TECHNISCHE UNIVERSITÄT MÜNCHEN

TUM School of Engineering and Design

Fixed-Wing Flight Control Law

Design using Incremental Control

Rasmus Steffensen

Vollständiger Abdruck der von der TUM School of Engineering and Design der Technischen
Universität München zur Erlangung eines

Doktors der Ingenieurwissenschaften (Dr.-Ing.)

genehmigten Dissertation.

Vorsitz: Prof. Dr.-Ing. Mirko Hornung

Prüfer der Dissertation: 1. Prof. Dr.-Ing. Florian Holzapfel

2. Prof. Dr.-Ing. Stephan Myschik

Die Dissertation wurde am 30.09.2024 bei der Technischen Universität München eingereicht
und durch die TUM School of Engineering and Design am 26.02.2025 angenommen.
This page is intentionally left blank.
Abstract

Flight control laws play a central role in the development of modern flying vehicles. In
order to safely and efficiently fulfil the mission objectives, a robust, safe and reliable
method of controlling the motion of the vehicle is necessary. Automatic control systems
are usually modularized, having mission handling, trajectory guidance and baseline control.
This thesis focuses on the baseline control of fixed-wing aircraft. For piloted aircraft, the
baseline controller is usually the interface between the pilot and the aircraft when the
pilot is manually flying the aircraft. In automatic modes, the baseline control serves as
the last control law of the cascaded control chain. Of all the control loops, the baseline
control is closest to the physics of the aircraft, and has to take into account the dynamics
influenced by aerodynamics, propulsion, inertia, airframe and external factors such as the
atmosphere and weather. Even though engineers are estimating very accurate models of
all these elements, in practice, the real aircraft dynamics will differ from the models. This
thesis investigates the emerging methods of sensor-based incremental control to increase
the robustness against parameter variations in the flight dynamic models, or allow the
use of lower quality models for the control law design. In order to successfully integrate
these methods into actual aircraft, several factors need to be properly taken into account.
Integration of control allocation, flight envelope protections, turn compensation, non-linear
compensations for kinematic couplings, gravity rate etc. has to be properly connected with
the incremental control method. Also effects such as structural mode couplings, sensor
noise and delays have to be properly accounted for. In this thesis, methods are developed
to accomplish this, such that incremental control can be applied in practice for fixed-wing
aircraft. Two main methods are developed, 1. incremental differential proportional-integral
(iDPI) control, and 2. incremental reference model based control. The first method is
demonstrated as a retrofit to a DA42 baseline control law and with implementation and
flight test using a small subscale remote controlled aircraft. Several improved properties
are demonstrated such as enhanced robust performance, avoiding hidden coupling terms
and better disturbance rejection. The second method is applied to the simulation of a
high performance turboprop demonstrator for piloted flight, showing excellent handling
qualities for both fine tracking and aggressive maneuvering.

i
This page is intentionally left blank.
Zusammenfassung

Flugregelungsgesetze spielen eine zentrale Rolle bei der Entwicklung moderner Fluggeräte.
Um die Missionsziele sicher und effizient zu erfüllen, ist eine robuste, sichere und zuverläs-
sige Methode zur Regelung der Fahrzeugbewegung erforderlich. Regelungssysteme sind in
der Regel modular aufgebaut und umfassen Missionsführung, Flugbahnführung und Basis-
regelung. Diese Arbeit konzentriert sich auf die Basisregelung von Starrflügelflugzeugen.
Bei bemannten Flugzeugen bildet der Basisregler in der Regel die Schnittstelle zwischen
Pilot und Flugzeug, wenn der Pilot das Flugzeug manuell steuert. Im automatischen Modus
fungiert die Basisregelung als innerstes Regelungsgesetz in der kaskadierten Regelkette. Von
allen Regelkreisen ist die Basisregelung am nächsten an der Physik des Flugzeugs und muss
die Dynamik berücksichtigen, die durch Aerodynamik, Antrieb, Trägheit, Flugzeugzelle
und externe Faktoren wie Atmosphäre und Wetter beeinflusst wird. Obwohl Ingenieure
sehr genaue Modelle all dieser Elemente erstellen, weicht die tatsächliche Flugzeugdynamik
in der Praxis von den Modellen ab. Diese Arbeit untersucht neue Methoden der sensorba-
sierten inkrementellen Regelung, um die Robustheit gegenüber Parameterschwankungen
in den Flugdynamikmodellen zu erhöhen oder die Verwendung von Modellen geringerer
Qualität für die Gestaltung des Regelungsgesetzes zu ermöglichen. Um diese Methoden
erfolgreich in echte Flugzeuge zu integrieren, müssen mehrere Faktoren angemessen be-
rücksichtigt werden. Die Integration von Control Allocation, Schutz vor Überschreitungen
der Flugbereichsgrenzen, Kurvenkompensation und nichtlinearen Kompensationen für
kinematische Kopplungen, Schwerkraft usw. muss angemessen mit der inkrementellen
Regelungsmethode verbunden werden. Auch Effekte wie strukturelle Moduskopplungen,
Sensorrauschen und Verzögerungen müssen angemessen berücksichtigt werden. In dieser
Arbeit werden Methoden entwickelt, um dies zu erreichen, sodass die inkrementelle Re-
gelung in der Praxis für Starrflügelflugzeuge angewendet werden kann. Es werden zwei
Hauptmethoden entwickelt: 1. inkrementelle Differential-Proportional-Integral-Regelung
(iDPI) und 2. inkrementelle referenzmodellbasierte Regelung. Die erste Methode wird als
Nachrüstung eines DA42-Basisregelungsgesetzes demonstriert und durch Implementierung
und Flugtests mit einem kleinen ferngesteuerten Flugzeug im Submaßstab validiert. Dabei
werden mehrere verbesserte Eigenschaften nachgewiesen, wie z. B. eine verbesserte Leis-
tung, die Vermeidung versteckter Kopplungsterme und eine bessere Störunterdrückung.

iii
Die zweite Methode wird auf einen Hochleistungs-Turboprop-Demonstrator für bemannten
Flug angewendet und zeigt hervorragende Handling Qualities sowohl für präzises Tracking
als auch für aggressives Manövrieren.

iv
Acknowledgment

First of all, I would like to thank my doctoral supervisor Prof. Dr.-Ing. Florian Holzapfel
for the time and support that went into this thesis. I would also like to thank him for
giving me the opportunity to contribute to so many exciting projects for the aerospace of
tomorrow.
The pleasant atmosphere at the chair would not have been possible without my
colleagues, who meanwhile have become great friends. I am grateful for the scientific
discussions and the social activities that have made the last few years so positive. I would
like to thank Agnes Steinert for the countless discussions and derivations, and for always
having a critical look to find all my mistakes. I would also like to thank Moritz Speckmaier
for the many interesting discussions, Alexander Zwenig for our work on the Albadrone,
Daniel Dollinger and Maximilian Wechner for the many late night hours preparing software
delivery to the turboprop project partner, and at last Max Söpper, Fredric Schwedler,
Rodolfo Hofman and the whole Grille team for the excitement of flight testing a large
multicopter vehicle.
I would like to thank Daniel Teubl & Sebastian Köberle from LLS who spent countless
hours with me on the TUM model airfield, teaching me remote piloting techniques so that
I could test my control software.
I would also like to thank Ewoud Smeur at TU-Delft for always challenging my thinking,
giving insightful ideas for solving the problems that I bring to him, it is always a pleasure
discussing with you.
I would also like to thank my friend Robert Miklos, who always gave feedback on my
control theory ideas with his huge knowledge of the field, and Adrien Østergaard for his
insights on flying and piloting.
Of course, I would also like to thank my students, who through their commitment
contributed very much to the success of this work. In particular Stefan Langen who worked
on handling qualities design of the control law, Sebastian Gölz who spent countless hours
in the simulator with pilots testing different control law variants, and Kilian Ginnell who
helped build and test the Albadrone demonstrator aircraft.
At last, I would like to thank my parents and grandparents, my sister and not the least
my girlfriend for always supporting me.

v
This page is intentionally left blank.
List of Publications

First author or shared lead author

• R. Steffensen, K. Ginnell, F. Holzapfel Practical System Identification and Incre-
mental Control Design for a Subscale Fixed-Wing Aircraft. MDPI Actuators, 13,
2024.

• R. Steffensen, A. Steinert, F. Holzapfel Incremental control as an enhanced and

robust implementation of gain scheduled controllers avoiding hidden coupling terms.
Aerospace Science and Technology 141, 2023.

• A. Steinert, R. Steffensen, F. Holzapfel Bias of anticipated non-parametric closed-loop

frequency response for controller gain re-tuning. Automatica, Volume 151, 2023.

• R. Steffensen, A. Steinert, E.J.J. Smeur Nonlinear Dynamic Inversion with Actuator

Dynamics: An Incremental Control Perspective. Journal of Guidance, Control, and
Dynamics 46 (4), 2023.

• R. Steffensen, A. Steinert, Z. Mbikayi, S. Raab, J. Angelov, F. Holzapfel Filter and

sensor delay synchronization in incremental flight control laws. Aerospace Systems,
2023.

• D. Braun, R. Steffensen, A. Steinert, F. Holzapfel Counter Optimization-Based

Testing of Flight Envelope Protections in a Fly-By-Wire Control Law Using Deep
Q-Learning. EuroGNC Conference, 2022.

• R. Steffensen, A. Steinert, F. Holzapfel Longitudinal Incremental Reference Model

for Fly-By-Wire Control Law using Incremental Non-Linear Dynamic Inversion.
AIAA SciTech Conference, 2022.

• A.C. Gabrys, R. Steffensen, R. de Angelis Cordeiro, J.R. Azinheira, A. Moutinho,

F. Holzapfel Integration of Phase Plane Flight Envelope Protections in Cascaded
Incremental Flight Control. IFAC-PapersOnLine 52 (12), 2019.

• R. Steffensen, A. Gabrys, F. Holzapfel Flight envelope protections using phase plane

limits and backstepping control. EuroGNC Conference, 2019.

vii
Co-author
• B. Hosseini, A. Steinert, R. Hofmann, X. Fang, R. Steffensen, F. Holzapfel, C
Göttlicher, Christoph Advancements in the Theory and Practice of Flight Vehicle
System Identification. Journal of Aircraft 60.5, 2023.

• M. Blanke, R. Stefensen Dynamic Wake Identification for Model-and Full-Scale

Container Ship. IFAC-PapersOnLine 55 (31), 2022.

• S. Scherer, M. Speckmaier, D. Gierszewski, C. Mishra, A. Steinert, R. Steffensen,

S. Wulf, F. Holzapfel Automatic Take-off and Landing of a Very Light All Electric
Optionally Piloted Aircraft. ICAS Conference, 2022.

• A Steinert, R Steffensen, D Gierszewski, M Speckmaier, F Holzapfel, R. Schmoldt,

F. Demmler, U. Schell, M. Ornigg, M. Koop Experimental Results of Flight Test
Based Gain Tuning. AIAA SciTech Conference, 2022.

• I. González Cabeza, B. Molesworth, M. Good, C. Caponecchia, R. Steffensen Investi-

gating the predictive validity of the COMPASS pilot selection test. The International
Journal of Aerospace Psychology 31, 2021

• A. Steinert, F. Holzapfel, R. Steffensen, C. Merkl Flight Test Based Gain Tuning

using non-parametric Frequency Domain Methods. AIAA SciTech Conference, 2021.

• P. Nagarajan, M. Maly, J. Jaisle, P. Gesting, R. Steffensen, M. Wechner, D. Gier-

szewski, C. Krammer, F. Holzapfel Taking Autonomy Out-of-the-Loop–Novel Method-
ology for the Development and Automated Operation of UAS in Integrated Airspace.
AIAA/IEEE 39th Digital Avionics Systems Conference (DASC), 2020.

viii
Contents

Abstract i

Kurzfassung iii

Acknowledgment v

List of Publications vii

Acronyms xi

1 Introduction 1
1.1 Motivation and problem statement . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 State of the art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Industry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.2 Academia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2 Methodology 13
2.1 Non-linear Dynamic Inversion . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Differential Proportional Integral Control . . . . . . . . . . . . . . . . . . . 23
2.3 Eigenstructure Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4 Control Allocation and Redistributed Pseudo Inverse . . . . . . . . . . . . 28
2.5 Control Law Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3 Article Introductions 33
3.1 Paper 1: Nonlinear Dynamic Inversion with Actuator Dynamics: An Incre-
mental Control Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Paper 2: Incremental control as an enhanced and robust implementation of
gain scheduled controllers avoiding hidden coupling terms . . . . . . . . . . 35
3.3 Paper 3: Practical System Identification and Incremental Control Design
for a Subscale Fixed-Wing Aircraft . . . . . . . . . . . . . . . . . . . . . . 37
3.4 Paper 4: Filter and sensor delay synchronization in incremental flight control
laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

ix
CONTENTS

3.5 Paper 5: Longitudinal Incremental Reference Model for Fly-By-Wire Control

Law using Incremental Non-Linear Dynamic Inversion . . . . . . . . . . . . 41

4 Conclusion 43
4.1 iDPI, an incremental control method for classical controllers . . . . . . . . 43
4.2 Sensor based incremental control with filter and delay synchronization . . . 45
4.3 Incremental reference model based control concept for INDI . . . . . . . . 45
4.4 Other Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

5 Outlook 47

A Article Reprints I
A.1 Paper 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . I
A.2 Paper 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . XI
A.3 Paper 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . XXX
A.4 Paper 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . LXIV
A.5 Paper 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . LXXXV

x
Acronyms

ADS Aeronautical design standard

ARP Aerospace recommended practice
BIT Built-in-test
CS Certification specifications
DPI Differential proportional-integral
FSD Institute of Flight System Dynamics
HCT Hidden coupling terms
iDPI Incremental differential proportional-integral
IMU Inertial measurement unit
INDI Incremental non-linear dynamic inversion
LHP Left half-plane
MEMS Micro-electromechanical system
MIL-DTL Military detail specification
MIL-STD Military standard
MIMO Multiple-input multiple-output
MTE Mission task element
NDI Non-linear dynamic inversion
PI Proportional-integral
PID Proportional-integral-derivative
RHP Right half-plane
RPI Redistributed pseudo inverse
SAE Society of automotive engineers
SAE-AS SAE Aerospace standard
SISO Single-input single-output
TUM Technische Universität München
UAV Unmanned aerial vehicle
VTOL Vertical take-off and landing

xi
This page is intentionally left blank.
Chapter 1

Introduction

1.1 Motivation and problem statement

Over the last 120 years, there has been an enormous advance in control of fixed-wing
aircraft. From the pioneers of aviation who debated whether to use the rudder or ailerons
to turn the aircraft to full authority digital fly-by-wire aircraft, where the commands of the
pilot are checked and modified before being transferred to the control surfaces. Autopilots
have been developed that can fully control everything from take-off, mission flying and
landing, allowing for highly automated and even autonomous missions.
All the way back in 1914, an early autopilot, the Sperry Aeroplane Stabilizer was
demonstrated using gyroscopes to perform controlled flight. Starting in the early 1960’s,
development of closed-loop feedback flight control systems began to support the expansion
of the flight envelope of new advanced aircraft. The control laws were necessary to
achieve better performance in terms of speed and altitude, physical design and improved
maneuvering characteristics. A famous example is the X-15 experimental aircraft, which
featured a Stability Augmentation System with adaptive gain scheduling. With the
advancement of these new control techniques, also new challenges arrived. The X-15
suffered a fatal crash due to loss of stability combined with a pilot induced oscillation. This
made the importance of stability, robustness and handling qualities of feedback control
laws evident.
At present, and in the coming years, a lot of new aircraft configurations will emerge
as new technologies will be integrated into future airplanes. Hydrogen powered aircraft,
battery powered aircraft, transition aircraft and so on will need to go through development,
flight testing and in the end certification. To support the rapid iterative development of
such aircraft, a high reduction in development time and cost can be achieved if the flight
control laws can be designed using lower fidelity models while ensuring robust performance,
stability and good handling qualities. A branch of new techniques, namely sensor-based
incremental control has the potential to achieve this by reducing the model dependency
and retaining the desired performance even with large differences between the model used
for control design and the actual aircraft. Control laws for the next generation of aircraft

1
1.2 State of the art

need to incorporate state-of-the-art flight envelope protections, have proper selection of

control variables that support the mission tasks, have dynamic characteristics that provide
excellent handling qualities and must be easy to interpret and implement. In the end, this
thesis contributes to advance the methods of incremental control in order to accomplish
this.

1.2 State of the art

This section introduces the state of the art in terms of flight control laws for fixed-wing
aircraft. The section is split between industry and academia, as there is still a noticeable
gap between established industry and academia. This gap is though slowly reducing with
the many startups in the electrical urban air mobility race toward sustainable aviation.
New challenges spark the necessity of new techniques, and this is observed by the creativity
in the deployment of modern and advanced control solutions.

1.2.1 Industry
In industry several different methods and control techniques have been applied for inner-
loop control. Even though the basic concepts of control are similar in most cases for
fixed-wing flight control, several different methods are applied across different companies.
Even though some differences stem from different requirements, a large part of the control
algorithm choice is due to tradition in either the company or the region. As most of the
control law details are intellectual property of the aircraft manufacturers, only a short
overview is given based on data released to the public through patents and overview papers.

Control Law Strategies:

Most control laws in civil airliners use some sort of proportional-integral (PI) control law
combined with different kinds of linear or non-linear compensation terms. Both Airbus
[15, 53], Boeing [45] and Embraer [11] uses PI based feedback laws. In the newest Airbus
fly-by-wire aircraft (A350/A321XLR) [16], an additional pitch rate derivative compensa-
tion term is used in a similar fashion to incremental non-linear dynamic inversion. This
compensation is used to increase the robustness towards uncertainties in the pitching
moment coefficient derivative with respect to angle of attack.

For modern high performance fighter aircraft, also variants of PI control laws are used.
In the Eurofighter, the so called differential PI is used [51] together with several non-linear
compensation terms. The F-16 [18], YF-23 [42] and F-18 [13] also uses a PI type control law.

2
 Chapter 1: Introduction

The F-35 was the first production aircraft to employ a non-linear dynamic inversion con-
trol law [12], which allowed to perform post-stall maneuvering without thrust vectoring [35].

In the emerging market of electrical vertical take off and landing, more experimentation
with modern control laws has also been practiced. Incremental non-linear dynamic inversion
has been employed in simulation for hover, transition and wing-born flight for the eRC
concept vehicle [3].

Control Variables and Envelope Protections:

In terms of control variables, there are overall two distinct strategies employed by manu-
facturers of civil transport aircraft. One approach aims to retain the speed stability of the
aircraft, such that the basic response resembles the aircraft’s natural tendency as if there
was no closed-loop control law. The other strategy is to decouple the speed and flight
path control, in the end producing a neutrally speed stable aircraft.
The first approach is integrated in Boeing’s fly-by-wire aircraft with the C ∗ U control
variable [47]. C ∗ is a linear combination of pitch rate q and the incremental load factor
at the pilot station ∆nz,pilot . The C ∗ U adds a trim velocity error term with C ∗ U = C ∗ +
KU (U − Utrim ) as given in [47], such that the aircraft remains speed stable, i.e. with no
commands, the aircraft returns to its trim speed Utrim . The trim speed is adjusted using
classical trim switches. The behavior resembles a classical aircraft, and a control column
with force shaping is used as control inceptor.
The second approach is applied in Airbus fly-by-wire aircraft. In the Airbus aircraft,
a C ∗ control variable is used. This approach decouples the flight path from the speed
response, and allows for better tracking during approaches and other precision flight path
tracking tasks. However, this results in a behavior that does not satisfy Part 25 certification
regulations for static longitudinal stability. In case the aircraft is maneuvered away from
its trim speed with no change in thrust, the speed will diverge away from the trim speed.
In order to pass certification, a special condition is issued that sets requirements for speed
protections to avoid the divergence resulting in overspeed or low speed conditions [52].
Flight envelope protections can be split between hard and soft protections. This is
mostly a philosophical choice, and determines whether the pilot can override the limits
or not. Both Airbus and Boeing aircraft feature extensive envelope protections. Airbus
has historically had hard protections for both angle of attack, load factor, speed, pitch
angle and bank angle as described in [52]. The protections are implemented by a dynamic
min function for the baseline load factor commands. Boeing fly-by-wire aircraft have on
the other hand historically employed a soft limiting philosophy, that allows the pilot to
override the protections, but gives visual and haptic cues to avoid excursions of the limits
[47].

3
1.2 State of the art

Certification Requirements:

The development of fly-by-wire systems for civil aircraft is a complicated process and
several strict guidelines and standards need to be followed if the control laws shall be
approved and certified for an end product by the aviation authorities. In the end, classical
airplanes in Europe need to be certified according to either CS-23 [20] for small aircraft or
CS-25 [19] for large aircraft.
With regards to certification of control laws, there are several certification standards,
acceptable means of compliance and guidance material to take into account. In general,
different aspects of the control law need to satisfy different set of requirements. For the
software development, flight control software is classified as safety critical and needs to
follow DO-178C [58] with DO-331 [57] providing details for model based development.
Guidelines for the development are given in ARP-4754B [50] and the corresponding safety
assessment in ARP-4761A [49].
For system design, installation and testing the SAE AS94900A [48] specifies several
requirements, in particular regarding performance and stability of the control law in
nominal conditions, uncertain conditions and in degraded operation. Similar requirements
are found in the military standard MIL-DTL-9490E [80]. Background material for these
requirements is given in [79].
For handling qualities of piloted fixed-wing aircraft, MIL-STD-8785C [17] and its
successor MIL-STD-1797A [56] and MIL-STD-1797B and C draft (not released to public)
gives quantitative and qualitative requirements. The majority of the quantitative require-
ments are given in terms of frequency and time domain criteria, which can be checked
to be within certain boundaries defining Level 1, 2 and 3 handling qualities. Qualitative
requirements are often rated according to Cooper-Harper by the pilot in either simulation
or flight test. In that regard, Mission Task Elements (MTEs) are usually defined to check
the quality of the control laws with respect to solving specific flying tasks. MTEs have
been used extensively for rotorcraft evaluation defined in ADS-33E [68], and are slowly
being adapted to the handling quality standard documents for fixed-wing aircraft [43].

1.2.2 Academia

The development and application of control techniques are much more rapid and versatile
in academia, as there are less constraints in terms of certification and processes. Many new
concepts are tested on small subscale research platforms, both in terms of multi-copters
but also transition vehicles and fixed-wing aircraft. Especially advances in microcontrollers
and MEMS sensor technology have made small flight control computers cheaply available.
This has accelerated the testing and deployment of small research vehicles with very few
resources. In the following, the current state of the art within fixed-wing incremental flight
control design in a research context is summarized.

4
 Chapter 1: Introduction

Sensor-based incremental control:

In order to increase robustness and reduce the required model knowledge for control law
design, research has gone in the direction of replacing model dependent terms in the control
law with measurements. This is commonly achieved by a local linearization, and then
discarding the terms related to the system dynamics. The remaining part is dependent
only on the control effectiveness, and the measured effector position as an incremental
term, hence the term incremental control. This method has turned out to work very well
in practice, especially in the domain of multicopters. Several control methods have been
shown to have an incremental counterpart, in particular non-linear dynamic inversion [63]
and backstepping control [1].

Incremental Non-Linear Dynamic Inversion:

Incremental non-linear dynamic inversion (INDI) is a popular control method for flying
vehicles in a research setting at the moment. It has been applied with great success to
multi-copters for superior performance with minimal modelling effort. In [2], INDI is used
to track the velocity of a hexacopter UAV, and the controller is shown to be robust against
uncertainties affecting both translational and angular accelerations. In [66], a cascaded
attitude and position INDI controller is proposed. The controller is shown to have superior
disturbance rejection capabilities compared to a classical proportional-integral-derivative
(PID) controller using only sensor data and knowledge about the system’s input effec-
tiveness. In [54] a unified control strategy for all phases of flight of a vertical take-off
and landing (VTOL) transition vehicle is proposed using incremental non-linear dynamic
inversion. In contrast to a cascaded loop, the proposed approach can simultaneously
allocate forces and moments, allowing for higher bandwidth control. In [55], an approach
that considers first order actuators in an INDI setting is introduced, denoted extended
INDI. This approach is especially useful if several actuators with different bandwidth exists,
or can be used if the bandwidth of the actuators should be artificially increased or reduced
by the control law. In [36] a hybrid approach to INDI is proposed. Here, a combination of
complementary filtering and actuator feedback is used together with a system model, in
order to mitigate measurement time delays. In [78] a global incremental control algorithm
based on INDI is applied to a tailsitter flying wing for trajectory following. The controller
is able to track agile trajectories in an indoor setting with very little model information.
The largest fixed-wing aircraft yet to employ an INDI strategy is the Delft PH-LAB [26].
Here, an attitude control law is flight tested, with focus on the problems arising from
slower actuators and larger time delays, which is in part dealt with using the concept of
pseudo control hedging.

5
1.2 State of the art

Reference Models as Desired Dynamics:

Explicit model following approaches are sometimes used to shape the desired closed-loop
response, i.e. from reference to response. For flight control of manned fixed-wing aircraft,
very specific requirements exist for the aircraft response in order for pilots to rate the
handling qualities as desired. This is a good use case for a model following approach,
since a reference model can be designed with the desired dynamics, while the feedback
path can be tuned to fulfil requirements for disturbance rejection, stability, robustness,
structural couplings, loads, noise etc. A model following approach is well suited with
the control methods non-linear dynamic inversion (NDI) and INDI. A reference model
shapes the desired dynamics, and an associated error controller tries to minimize the error
between the plant and the desired reference response. In [82, 83] a physical reference
model is used to shape the desired dynamics. In [7] an integrated reference model is
used for a transition VTOL aircraft. The reference model provides reference dynamics for
the entire flight envelope, from hover, transition to wing-born flight in a continues and
unified manner. The reference model is based on physical signals and employs envelope
protections that correspond to the physical capabilities of the vehicle using a phase plane
based approach. Pseudo-control hedging is used to account for actuator dynamics and
non-linear actuator effects. In [6, 8], feed-forward terms are introduced from the reference
dynamics to provide better tracking in nominal conditions, and in [9] this is extended with
online identification to adapt the feed-forward in uncertain conditions. In [67] a second
order model is used to shape the handling qualities for pitch rate control of a fixed-wing
aircraft, which in combination with an adaptive element provides level 1 handling qualities
in case of inaccurate control effectiveness estimates and shifts in center of gravity.

Gain scheduling and hidden coupling terms:

When implementing gain scheduled controllers, the effect of the gain scheduling should also
be considered. Either it needs to be shown that the effects of gain scheduling are negligible,
for example, if the scheduling variable changes slowly, or implementation strategies that
support gain scheduling should be used. Analysis of gain scheduled controllers is thoroughly
investigated in [59], and is based on linear parameter varying system analysis. In baseline
flight control, gain scheduling is often done with respect to static and dynamic pressure
[75], which are slowly varying. Using extensive simulation testing of the full system, it
can be shown that it is highly unlikely that scheduling effects will cause degradation in
the stability and performance of the control law. In case that the scheduling variables are
varying quickly, modifications to the implementation can improve the susceptibility to
the gain scheduling. In [33] a velocity algorithm is proposed, often denoted differential
proportional-integral (DPI) control. The algorithm avoids hidden coupling terms, which
are unwanted additional terms that can appear in the linearization of a gain scheduled
controller. A straightforward implementation of a classical PI controller will contain

6
 Chapter 1: Introduction

hidden coupling terms. With the algorithm in [33], PI controllers can be immediately
implemented using the velocity method, and it is shown that they perform much better
when applied on the nonlinear system with gain scheduling. The proposed algorithm cannot
be directly implemented on a real system, since it requires direct non-causal derivates
of the measurements. An approximation can be used to circumvent this limitation, by
using a filtered derivative, and an integral with a corresponding zero as shown in [28].
This implementation was also shown to avoid hidden coupling terms in [38]. In addition
to theoretical studies in academia, this type of implementation has been used on the
production Eurofighter aircraft [51].

1.3 Objectives

In light of the above state of the art, there are still gaps for incremental control laws to be
deployed in production fixed-wing aircraft. In this section, three elaborated objectives are
formulated, forming the basis of the research direction. The overall aim is to reduce the
gap between the current state of research and future flight control products.

O1: Apply incremental control techniques to decrease dependency on accurate

models and increase robustness against plant variations for fixed-wing flight
control. (Paper 1 [73], Paper 2 [71], Paper 3 [74])

The degree of robustness towards unforeseen events or unknown system characteristics

is essential for the quality of a control system. A key aspect is the robustness towards
parameter variations or variations in the flight dynamics. A control law is usually designed
to obtain some desired performance based on a dynamic model of the aircraft. In actual
flight, the perceived parameters may vary from the expected values due to a number of
reasons: Good models are often very expensive, so cheaper, lower grade models might be
used for the control law design; parameters may change during flight, for example in case of
icing on the airframe, or damage to the aircraft; the dynamics may change with variation
in center of gravity location or mass when load is released, or in case of inadvertent shift
of internal load. If the changes are unknown, the robustness of the control system should
consider both the resulting performance and stability. A promising method to drastically
increase the performance of the control law in case of parameter variations is the emerging
sensor-based incremental control techniques. It is an objective of this thesis to further
develop the methods of incremental control in the area of fixed-wing flight control.

7
1.4 Contributions

O2: Develop a reference model based incremental control law strategy for
application in piloted high performance fixed-wing aircraft. (Paper 5 [70])

Incremental control has mostly been applied to drones and research aircraft for basic flight
control functions. In order for application in piloted high performance aircraft intended
for a variety of missions, a complete structure of the control law needs to be formulated. It
is sought to have a blending of the control variable over the entire envelope supporting the
tasks and limitations in the different areas of the flight envelope. The control law needs
to allow for incorporation of flight envelope protections to guard against inadvertently
damaging the aircraft, while being integrated in a way that allows for full usage of the
available performance of the aircraft. The control law needs to allow for turn coordination
and compensation in order to assist the pilot in various tracking tasks such as precision
approaches and formation flying. The control law should automatically compensate for
non-linear coupling effects such as gravity rate, kinematic and inertial coupling. All these
functions shall contribute to a system that features carefree handling for the pilot over the
entire flight envelope.

O3: Develop methods for incremental control that allow processing of measure-
ments, including attenuation of noise, structural couplings and compensation
of known measurement time delays. (Paper 4 [72])

It is a goal of this thesis to integrate classical methods of attenuating structural mode

couplings using notch filters, attenuating noise using roll-off filters and proper compensation
of time delays within incremental control laws. Severe damage and structural failure have
happened in flight control systems with inappropriate consideration of aeroservoelastic
effects. Therefore, an important factor for fixed-wing flight control is to account for the
interaction between the structural dynamics of the aircraft and the control law. The
objective is to use classical methods of notch filtering to attenuate the coupling effects
within the incremental control law. Sensor noise and vibrations on measurements can
lead to jitter and vibrations on the control surfaces, which causes tear and wear on the
mechanical parts. Therefore noise and vibrations shall be accounted for and appropriately
attenuated in the control law using roll-off filters. Lastly, sensor time delays can lead
to instability of the closed-loop system, and hence known sensor time delays shall be
accounted and compensated for in order to increase the stability and robustness.

1.4 Contributions
The contributions of this thesis address several challenges related to implementing incre-
mental flight control laws for fixed-wing aircraft. In particular, this thesis provides the
following contributions:

8
 Chapter 1: Introduction

C1: Novel derivation of sensor-based incremental control law based on non-

linear dynamic inversion allowing for modification of closed-loop effector band-
width without sacrificing the desired flight dynamic characteristics. (Paper 1
[73])

To establish a basis for analysis and development of incremental control laws, a mathemat-
ical derivation is proposed for a sensor-based incremental non-linear dynamic inversion
based controller. The derivation takes into account the dynamic characteristics of the
effector in the inversion. This is achieved by applying the chain rule and using the resulting
Jacobian matrices together with the actuator increments and state derivatives, such that
an exact incremental control law formulation is obtained. It is shown that by proper
choice of error dynamics, splitting up in system dynamics and actuation dynamics, a
limit value problem can be formulated, resulting in the popular incremental non-linear
dynamic inversion. This derivation reveals how to scale the input effectiveness to adjust the
closed-loop effector bandwidth without (considerably) changing the desired flight dynamic
characteristics. This is useful in case it is required to avoid non-linearities above a certain
frequency, such as non-linear aerodynamic effects on the control surfaces, measurement
non-linearities, actuator rate saturations or surface drive-train non-linearities.

C2: Novel robust control law based on incremental control, that allows for
gain scheduling without hidden coupling terms. (Paper 2 [71])

A novel control method is proposed, which combines the robust performance of incremental
non-linear dynamic inversion, the simplicity of a PI controller and the property of a
DPI implementation that allows for gain scheduling without hidden coupling terms. The
method, denoted incremental DPI (iDPI), is analyzed in terms of flight control for the
longitudinal motion of a fixed-wing aircraft, and shows superior robust performance towards
parameter variations in the plant compared to PI and DPI. Additionally, the method does
not lead to hidden coupling terms, allows for bump-less switching between different gain
groups, has better disturbance rejection and is less prone to integrator wind-up.

C3: Novel synchronization structures in incremental control laws accounting

for sensor filters, notch filters and known measurement time delays. (Paper 4
[72])

A method to include notch filtering, roll-off filtering and compensate for known measure-
ment time delays in several types of incremental control laws is proposed. Any filtering
and delay for the measurements will cause out of phase conditions with the incremental
part of the control law. A method is proposed to appropriately synchronize the filtering of
the measurements with the incremental part, since incorrect or lack of synchronization
can lead to loss of stability. A complementary filter method is proposed to cope with

9
1.4 Contributions

dissimilar filtering in the different measurement channels. This is important when consid-
ering different structural filtering for the longitudinal, lateral and directional part of the
measurement channel.

C4: Application of novel iDPI structure to an industry-level control law.

(Paper 2 [71])

The iDPI method is applied to a high fidelity simulation of the TUM/FSD Diamond
DA42 technology demonstrator aircraft. The method is extended with the necessary
implementation modifications required to apply it to an industry-level PI baseline control
law developed at TUM/FSD. In particular, innovative solutions are presented for integration
with notch filtering due to structural couplings, roll-off filtering due to noise, nonlinear
compensations for trim rate, inertia coupling, turn compensation/coordination and control
allocation. The evaluation is performed on a level 6 flight dynamics model of the DA42.
Significant improvement in robust performance is demonstrated while satisfying the
required stability margins.

C5: Application of eigenstructure assignment to iDPI with implementation

and flight test results for a remote controlled subscale aircraft. (Paper 3 [74])

Eigenstructure assignment for gain design is used together with a novel flow angle derivative
implementation for the iDPI control structure, to validate the method with flight test on a
small subscale aircraft. The results demonstrate that the concept works on a real aircraft,
with all the considerations for a functional flight control system. Models for both actuator
and flight dynamics are estimated from open-loop flight test, where the estimation data
is collected with sine-sweep excitation on all axes. The models are used together with
eigenstructure assignment to design gains for the control law. In the end, closed-loop flight
test is performed that demonstrates that the performance of the controller is as predicted
by the estimated model.

C6: Novel physically motivated incremental reference model allowing incorpo-

ration of control variable blending, handling qualities design, flight envelope
protections, pilot relief functions and a non-linear base value term for use in
incremental control laws. (Paper 5 [70])

In order to apply incremental control for a piloted aircraft, a structure is proposed where
the desired reference dynamics are split between a non-linear base value and a linear
maneuvering part. Quantitative handling quality requirements are usually specified in
terms of desired characteristics of linear low order equivalent models. Therefore, a physically
inspired short-period model is used for reference, with the dynamic characteristics satisfying
the relevant requirements. The internal signals of the reference model are in terms of
incremental/decremental physical quantities, which allows for seamless blending of the

10
 Chapter 1: Introduction

tracked variable. Additionally, it allows for predictive envelope protection of several

competing variables within the reference model by application of phase plane based
protections. The derived structure allows for carefree handling by the pilot over the entire
design flight envelope of the aircraft. The nonlinear base value is introduced to commands
that are not directly related to the maneuvering by the pilot. This includes automatic
kinematic coupling compensation, gravity rate compensation, turn compensation and
pilot relief autopilot functions. In addition, the base value can be adjusted to seamlessly
introduce a natural speed stability at low speeds, by incorporating a virtual trim value,
based on a dynamic min-function. In this way, the pilot is always aware of the exact
control margin with respect to stall, including the effects of automatic compensations, such
as for turns. The different contributions of the base value can be faded out when required,
e.g. the turn compensation is faded out for rapid roll commands, to avoid un-commanded
large variations in the load factor.

1.5 Outline
This thesis is structured in five chapters. The first chapter contains the introduction. The
second chapter goes through the methodology applied in this research. Chapter three
contains a summary each of the five publications that form this thesis. Chapter four gives
the conclusion and Chapter five contains the outlook for the future. The papers for this
thesis are reprinted Appendix A, Paper 1 is in Section A.1, Paper 2 is in Section A.2,
Paper 3 is in Section A.3, Paper 4 is in Section A.4, Paper 5 is in Section A.5.

11
This page is intentionally left blank.
Chapter 2

Methodology

In this chapter the methodology for the research of this thesis is summarized. In particular,
this consists of an introduction to the techniques used for control law development. It
consists of an overview of the control law functions, introduction to the control algorithms
and an overview of the methods used for control law assessment.
A control law for any aircraft consists of many features, functions and sub-functions.
When discussing the inner-loop control for piloted aircraft, these functions can be divided
into four main areas: Control and Stability Augmentation, Envelope Protections, Failure
Handling / Reconfiguration and Status & Integrity Checks. A typical set of functions and
features are depicted in the diagram in Fig. 2.1. This thesis will be mainly focussed on

Control Law Functions

Control and Stability Augmentation Envelope Protections Status and Integrity Checks

- Shaping of dynamics / handling - Angle of attack protection. - Power on BIT’s.

qualities. - Load factor protection. - Continuous BIT’s.
- Control variable blending. - Pitch/bank angle protection. - Control Law Status Indications.
- Actuator position/rate saturation - Overspeed protection. - Function Status Indications.
handling. - Low energy protection. - Internal Input Signal Integrity
- Control allocation. - Upset recovery. Checking.
- Disturbance rejection. - Automatic ground collision - Internal Command Signal
- Turn coordination & avoidance. Safety Checking.
compensation. - Internal Signal Checking.
- Gravity compensation. - Signal fault detection and
- Non-linear compensations. Failure Handling / Reconfiguration isolation.
- Structural mode attenuation.
- Load alleviation - Sensor failure reconfiguration or
- Ride quality enhancements. degradation.
- Noise attenuation. - Actuator failure reconfiguration
- Pilot relief functions or degradation.
- Automatic pitch/flight - Flight control computer failure.
path hold. - Engine malfunctions handling.
- Automatic altitude hold. - Transient free control law reset,
- Automatic bank hold. switching and degradation.
- Automatic course hold.

Figure 2.1: List of typical inner-loop control law functions.

13
2.1 Non-linear Dynamic Inversion

the methods involved with Control and Stability Augmentation and Envelope Protections
integration. In the following sections, the underlying methods used for development are
discussed.

2.1 Non-linear Dynamic Inversion

The foundation of this thesis is inspired by the concept of nonlinear dynamic inversion, also
commonly denoted input-output feedback linearization. The concept has been thoroughly
introduced in several textbooks such as [34] and [65]. Input-output feedback linearization
is used to transform a certain type of non-linear system into a system with a linear input-
output relationship. This is different from classical linearization of a dynamic system,
which only locally approximates a non-linear system in a single stationary point. Feedback
linearization seeks to establish a linear relationship between the input and the controlled
variable for a large part of the non-linear system at once. In particular the method can be
used both for the regulation problem and for the tracking problem. In this thesis focus
will be on the tracking problem using new incremental variants of the method.

The SISO case

Definition 1 (Lie Derivative [65]). Let h : Rn → R be a smooth scalar function, and
f : Rn → Rn be a smooth vector field on Rn , then the Lie derivative of h with respect to f
is a scalar function defined1 by Lf h = ∇hf .

A more general definition can be formulated in differential geometry, however, for use
in terms of this thesis, the definition just serves as a convenient notation. In this setting,
the Lie derivative is just the straightforward derivative using the chain rule, in particular
the derivative along a vector field.
The input-output feedback linearization procedure is conducted by taking derivatives
of the output until the input appears explicitly. The system to be considered is given by
.
x = f (x) + g(x)u
(2.1)
y = h(x),

where x ∈ Rn is the system states, u ∈ R is the input, y ∈ R is the output, f ∈

C (r) (Rn ; Rn ) and h ∈ C (r) (Rn ; R), where C (r) (D, R) denotes the space of functions element-
wise differentiable r times with continuous derivatives, from the domain D with range R.
Note that for the procedure to work, neither input or output delays can be present, and in
this derivation the system should be input-affine. Modifications exist that can allow for
 
1 ∂h ∂h
The gradient notation is understood as: ∇h(x) ≡ ∂x 1
, ..., ∂xn and the product is understood as
∇hf = (∇h) (f ), i.e. a row vector multiplied by a column vector resulting in a scalar. The composition
of the Lie Derivative operator is understood as: Lf Lf h = ∇ (Lf h) (f ) = (∇ ((∇h) (f ))) (f )

14
 Chapter 2: Methodology

certain types of non-affine inputs, see for example [65].

Let the relative degree r be defined as the smallest integer value where

Lg Lf (r−1) h(x) 6= 0, (2.2)

then the r’th derivative of the output can be written as

y (r) = Lf r h(x) + Lg Lf (r−1) h(x)u. (2.3)

This choice of relative degree will match the usual definition of relative degree for a linear
system, namely the surplus of poles compared to zeros of the transfer function from input
to output [65].
The control law is obtained by requiring y (r) = ν, where ν is a pseudo control input,
and then solving Equation (2.3) for u, such that:
 −1
u = Lg Lf (r−1) h(x) (ν − Lf r h(x)) . (2.4)

If inserting this control law into the dynamics given by Equation (2.3), the following
closed-loop dynamics is obtained:
y (r) = ν (2.5)
This can be seen to be a chain of integrators between the pseudo control input ν and the
output y, and it is straightforward to design a linear feedback controller to shape these
dynamics.
It shall be noted that there exist cases where the relative degree is undefined, see [65,
p. 248].

Internal Dynamics and Zero Dynamics

Since the closed-loop dynamics shaped by Equation (2.5) only relates to the output, it can
happen that other states of the system grow unbounded, even if the output stays bounded.
Theoretically, this would still work, but for most practical systems, the unbounded states
relate to physical quantities and restrictions from mechanics or actuators would in the
end destroy the system. The states that are unobservable from the output are called the
internal dynamics. Therefore, it shall be checked for a control law whether the internal
dynamics are stable. The internal dynamics will depend on the chosen control law, and
are difficult to check for stability [65]. Instead, something called zero dynamics will be
used to assess the (local) stability of the internal dynamics. To introduce the concept of
zero dynamics, it is best understood by considering a linear system.

Example 2.1.1 (Internal Dynamics of a Linear System). Consider the short period
dynamics given in Equation (2.6) with the output given by Equation (2.7) and parameters
given in Table 2.1, where α denotes the angle of attack, q the pitch rate and η the
elevator deflection. The parameters Zα , Zq , Zη , Mα , Mq and Mη are the stability and

15
2.1 Non-linear Dynamic Inversion

Table 2.1: Example aircraft parameters

Zα = −1.38 Zq =0 Zη = −0.1
Mα = −4.588 Mq = −2.0 Mη = −4.0
V0 = 51.54m/s

control derivatives as derived in [46] and V0 is the trim velocity. The ∆ denotes deviations
around the trim values. For simplicity assume Zq = 0.

.      
∆ α Z α 1 + Zq ∆α Z η

.  =  +  ∆η (2.6)
∆q Mα Mq ∆q Mη
In the following, the load factor, nz , is controlled directly via the lift on the tailplane,
by moving the elevator. This is a very poor strategy, and the example will show that the
right half-plane (RHP) zero of the elevator to load factor transfer behavior will cause the
internal dynamics to be unstable.
The output equation is given by:
V V
∆nz = − Zα ∆α − Zη ∆η (2.7)
g g
By inspection, it can be verified that the zeros of the transfer function from ∆η to ∆nz are
the roots of the the following polynomial:
Zα
s2 − Mq s − Mα + Mη =0 (2.8)
Zη
By applying the Routh-Hurwitz criteria, the roots are all in the left half-plane (LHP) if
and only if
!
Zα
−Mα + Mη >0
Zη
since (−Mq ) is positive. However, with the provided numeric values, there is a zero in the
RHP (z = 6.18) as expected. For a conventional wing-horizontal tail aircraft, the direct lift
effect of the elevator on the load-factor is momentarily negative for a pitch up manevuer
with a positive steady state load-factor.
In order to apply feedback linearization, the derivatives of the output shall be taken until
the input appears explicitly. Since the input ∆η appears in the output, no derivatives are
required and the relative degree is r = 0. The trivial pseudo control input is given by
ν = ∆nz . !
−g V0
∆η = ν + Zα ∆α (2.9)
V0 Zη g
By applying this control law, the resulting internal dynamics have an order of two and
can be calculated by inserting the control input in Equation (2.9) into the short period
dynamics in Equation 2.6 as follows

.   0 1
   
∆ α ∆α − Vg0

.  =  M − Mη Zα M   +
Mη g
 ν. (2.10)
∆q α Zη q ∆q − V0 Zη

16
 Chapter 2: Methodology

Since ν in this example is directly the reference, the stability of the internal dynamics is
determined by the eigenvalues of the system matrix in Equation (2.10). The associated
characteristic polynomial is given by
Zα
s2 − Mq s − Mα + Mη =0 (2.11)
Zη

which is exactly the same as the one defining the zeros in Eq. 2.8. This result holds in
general for linear systems, i.e., if the zeros of the transfer behavior are in the LHP, the
internal dynamics are stable and vice versa.
Note: The above example shows that this strategy does not work to control the load-factor.
A common feasible strategy is to control the load-factor at a point forward of the center of
percussion, which will not exhibit a RHP zero. Another strategy is to disregard the direct
lift effect of the elevator and only consider the load-factor generated by changes in angle of
attack.

The notion of zero dynamics can be extended to non-linear systems by the following
definition:

Definition 2 (Zero Dynamics[65]). The zero dynamics are the internal dynamics of the
system when the output is kept at zero by the input.

It is often much easier to check the stability of the zero dynamics than the internal
dynamics for non-linear systems. However, even with global exponential stability of the
zero dynamics, only local stability of the internal dynamics can be guaranteed, as argued
in [65]. Hence, the stability of the zero dynamics is often checked locally by classical
methods of linearization and verifying that the eigenvalues are in the LHP. In order to
check for global stability of the internal dynamics, one has to revert to Lyapnov methods,
which in practice can be difficult to apply to real systems.

The MIMO case

The concept of input-output linearization can be extended to multiple-input multiple-
output (MIMO) systems as explained in [65]. In the following, we restrict the analysis to
systems with as many inputs as outputs. Consider the non-linear system given by:

.
x = f (x) + G(x)u
(2.12)
y = h(x),
where x ∈ Rn is the system state, u ∈ Rm is the input, y ∈ Rm is the output, f ∈
C (r) (Rn ; Rn ), G ∈ C (r) (Rn ; Rn × Rm ) and h ∈ C (r) (Rn ; Rm ). The procedure is similar to
the SISO case, and is performed elementwise for each output, indexed by i.
m
yi (ri ) = Lf ri hi (x) + Lgj Lf (ri −1) hi (x)uj (2.13)
X

j=1

17
2.1 Non-linear Dynamic Inversion

The relative degree ri is defined for each output (i) such that:

Lgj Lf (ri −1) hi (x) 6= 0 (2.14)

for some j.
The above can be written in matrix form:
Lf r1 h1 (x) Lg1 Lf (r1 −1) h1 (x) ... Lgm Lf (r1 −1) h1 (x)
      
y1 (r1 ) u1
...  =  +
      


  ...   ... ... ...   ... 
  

ym (rm ) Lf rm hm (x) Lg1 Lf (rm −1) hm (x) ... Lgm Lf (rm −1) hm (x) um
| {z }| {z }
E u
(2.15)
and in case that the matrix E is invertible, the corresponding control law is:

Lf r1 h1 (x)
     
u1 ν1
 ...  = E−1  ...  −  (2.16)
     
     ... 

um νm Lf hm (x)
rm

This type of control law is called a decoupling control law, since νi only affects yi , and the
matrix E is called the decoupling matrix. The vectorial relative degree is just r = (r1 , ..., rm )
and the total relative degree r is defined as:

r= (2.17)
X
ri
i

In case the total relative degree is equal to the number of states, there will be no internal
dynamics. In case the total relative degree is less than the number of states, the analysis
of the zero dynamics should be conducted.

Invariant Zeros and Transmission Zeros

In this section the role of zeros will be introduced for linear MIMO systems. The section
is based on [41, 64]. The following analysis is based on linear time invariant multiple
input-output systems given on the form:
.
x = Ax + Bu
(2.18)
y = Cx + Du,

where x ∈ Rn is the system states, u ∈ Rm is the inputs, y ∈ Rr is the outputs, A ∈ Rn×n

is the system matrix, B ∈ Rn×m is the input effectiveness matrix, C ∈ Rr×n is the output
matrix and D ∈ Rr×m is the feed-through matrix. Note that in this context, r denotes
the number of outputs and not the relative degree. For simplicity assume that all the
eigenvalue pairs are distinct. Let the eigenvector matrix be given by V and the diagonal
eigenvalue matrix be given by Λ such that Λ = V−1 AV. Perform the following state
transformation to get a decoupled diagonal system:

x = Vx̃ (2.19)

18
 Chapter 2: Methodology

such that: .
x̃ = Λx̃ + V −1
| {z B} u
B̃ (2.20)
y = CV
|{z} x̃ + D,
C̃

where B̃ is the transformed input matrix and C̃ is the transformed output matrix. The
transfer function matrix can be written as:
 
X c̃i b̃T
G(s) = C̃ (sI − Λ)−1 B̃ + D = i
+ D, (2.21)
i s − λi
 
where c̃i is the i’th column of C̃ and b̃T is the i’th row of B̃. Two simple controllabili-
i
ty/observability criteria can be established as given in [22] and in [27]:

Criteria 1 (Controllability [22]). The system in (2.18) is controllable if:

 
b̃T 6= 0T for all i. (2.22)
i

or as given by [27] h i
rank( A − λi I B ) = n for all i. (2.23)

Criteria 2 (Observability [22]). The system in (2.18) is observable if:

c̃i 6= 0 for all i. (2.24)

or as given by [27]  
A − λi I 
rank( )=n for all i. (2.25)
C

As opposed to the classical rank criteria, these two criteria associate controllability
and observability to each eigenvalue. The second part of the two criteria also work for
repeated eigenvalues.

Definition 3 (Transmission Pole [41]). If an eigenvalue is both controllable and observable

according to criteria 1 and 2, then this eigenvalue value is called a transmission pole of
the system, as it appears in at least one element of G(s) as given in Eq. (2.21).

Definition 4 (System matrix formulation [64]). The laplace transform of the system in
Eq. (2.18) can be written in matrix form using the Rosenbrock system matrix 2 :
 
A − sI B 
P(s) =  (2.26)
C D
2
The first row is often written with opposite signs

19
2.1 Non-linear Dynamic Inversion

Table 2.2: Example aircraft parameters

Nr -0.52 Yr -1.98 Lr 0.29

Nβ 2.30 Yβ -0.20 Lβ -4.67
Np -1.12 Yp 0.02 Lp -6.67
Nξ 0.01 Yξ 0.00 Lξ 0.25
Nζ -0.09 Yζ 0.00 Lζ -0.01
V0 40.00 θ0 0.09 α0 0.09

such that     
−x(0)   A − sI B   X(s) 
 = (2.27)
Y(s) C D U(s)

Definition 5 (Invariant Zero [41]). If the rank of the Rosenbrock system matrix defined in
Eq. (2.26) looses full rank at s = η, then η is denoted an invariant zero of the system. I.e.
the invariant zeros are the solutions of det(P(η)) = 0.

The term invariant zero is justified in the sense that the zeros η is not affected by any
state feedback. Consider the state feedback u = −Kx + Fr, where K is the gain matrix,
r the reference and F an feed-forward gain matrix. The closed loop system can be defined
using Eq. (2.27):
       
−x(0)   A − sI B   X(s) A − sI − BK BF   X(s) 
 = =
Y(s) C D −KX(s) + FR(s) C − DK DF R(s)
(2.28)
The Rosenbrock-matrix from Eq. (2.26) for the closed loop system is given by:
    
A − BK − ηI BF   A − ηI B   I 0 
Pcl (η) =  = (2.29)
C − DK DF C D −K F

which reveals that the invariant zeros of the plant are also invariant zeros of the intercon-
nected state feedback system, if the right most matrix in Eq. (2.29) is full rank.

Definition 6 (Transmission Zero [41]). A transmission zero is an invariant zero that

either a) does not coincide with any eigenvalue or b) coincide with an eigenvalue which is
both controllable and observable.

Note, if the system is a minimal realization, then all the invariant zeros are transmission
zeros. This is justified because a realization is minimal if and only if all the eigenvalues are
both observable and controllable. Using a minimal realization of the system is therefore
practical for many control designs.

20
 Chapter 2: Methodology

Example 2.1.2 (Multivariate zeros and dynamic inversion). Consider the linearized lateral
dynamics given by:
.
0
      
r Nr Nβ Np r Nξ Nζ  
.
Yr − cos α0 Yβ Yp + sin α0 g
cos α0
      
 β    β  
 ξ 
Yξ Yζ 
= V0
+ , (2.30)
     
.
0
   
 p   Lr Lβ Lp  p  
 ζ
Lξ Lζ 
 .     
φ tan θ0 0 1 0 φ 0 0

with the parameters defined in Table 2.2, where r is the yaw rate, β is the side-slip an-
gle, p is the roll rate, φ is the bank angle, ξ is the aileron deflection, ζ is the rudder
deflection, θ0 is the trim pitch angle and α0 is the trim angle of attack. The parame-
ters Nr , Nβ , Np , Nξ , Nζ , Yr , Yβ , Yp , Yξ , Yζ , Lr , Lβ , Lp , Lξ and Nζ are the stability and control
derivatives as derived in [46]. Let the desired output to be controlled be the yaw and roll
rate, hence:    
r 1 0 0 0
y= = x (2.31)
p 0 0 1 0
The transfer function for the open-loop system is:
    
r(s)  1 z (s) z12 (s)   ξ(s) 
 11
= (2.32)
(s + 6.90)(s + 0.03)(s2 + 0.46s + 6.52) z21 (s) z22 (s)

p(s) ζ(s)

and the SISO zeros are given by:

z11 (s) = 0.013(s − 13.65)(s − 1.726)(s + 1.576)

z12 (s) = −0.09(s + 6.65)(s2 + 0.04s + 0.58)
(2.33)
z21 (s) = 0.25s(s2 + 0.72s + 5.16)
z22 (s) = −0.01s(s2 + 4.77s + 94.29)

and the transmission zeros are computed numerically to:

z(s) = s(s + 0.11). (2.34)

Note that the transmission zero at s = −0.11 is not a zero in any of the SISO transfer
functions, highlighting the difference between SISO zeros and transmission zeros.

Now consider the dynamic inversion control law given by:

−1 
  
.   
ξ   Nξ Nζ  r N r + Nβ β + Np p 
 d  −  r
 = . . (2.35)
ζ Lξ Lζ pd Lr r + Lβ β + Lp p
. .
where rd and pd are the desired yaw and roll accelerations. The closed-loop system
eigenvalues are calculated to be:
λ1 = −0.11
λ2 = 0
(2.36)
λ3 = 0
λ4 = 0

21
2.1 Non-linear Dynamic Inversion

with pole λ1 and λ2 at the two transmission zeros, and two additional poles λ3 and λ4 in
the origin from the NDI control law, one for each of the two outputs of relative degree one.
This is similar to what is observed for the SISO dynamic inversion control law in Example
2.1.1 and highlights the importance of transmission zeros for MIMO design.

Incremental Nonlinear Dynamic Inversion

The control law in Eq. (2.16) and (2.4) both depend heavily on the non-linear model.
In order to reduce the modelling effort, [4] proposed a variation of the control law by
substituting some model parts with measurements. In this procedure, the resulting control
law commands increments to the current actuator position rather than the absolute value,
hence the name incremental control. The proposed control law was motivated by a Taylor
expansion of Eq. (2.3) and then neglecting the part related to the system dynamics and
higher order terms. Rewriting Eq. (2.3) when considering a SISO system, i.e. y and u are
scalar, gives:
y (r) = a(x) + b(x)u (2.37)
where a(x) = Lf r h(x) and b(x) = Lg Lf (r−1) h(x). Now perform a Taylor expansion around
(x0 , u0 ) such that:
∂a(x) + b(x)u
y (r) (x, u) = y (r) + (x − x0 ) + b(x0 ) (u − u0 ) + O(x2 , u2 ).
(x0 ,u0 ) ∂x (x0 ,u0 )
(2.38)
Let x0 = x(t − Ts ) and u0 = u(t − Ts ), where t is the time and Ts is the sample time.
Let t − Ts be the time where we obtain a measurement of the system, and the time we
calculate and apply the new control command such that the input at time t is equal to
u(t). The time t is one step later, where we want the system to have obtained some new
desired value, namely y (r) (x(t), u(t)) = ν. If the sample time Ts is small, and we assume
that the actuators move much faster than the states change, i.e. |x − x0 | << u − u0 , then
by neglecting the second term on the right side of Eq. (2.38) and the higher order terms,
we obtain the following control law:
1
 
u= ν − y (r) + u0 (2.39)
b(x0 ) (x0 ,u0 )

where y (r) is the r’th derivative of the output at time t − Ts , and the idea is to use
(x0 ,u0 )
a measurement to obtain this value. It is seen that the control law in Eq. (2.39) only
depends on the input effectiveness b(x). This heavily reduces the modelling effort to apply
this type of control law. Since the control law does not depend on large parts of the model
parameters, it has also been shown to be much more robust against parameter variations
in the flight dynamics [63]. This thesis develops on the concept of incremental non-linear
dynamic inversion in paper 1 [73] with a novel derivation of the concept, that reveals
a method to compensate for state dependent terms and define the closed-loop actuator
bandwidth in the control design.

22
 Chapter 2: Methodology

2.2 Differential Proportional Integral Control

Another fundamental control concept in this thesis is the differential proportional-integral

(DPI) control concept originally introduced in [33]. This should not be confused with
the proportional- integral-derivative (PID) control concept from classical control theory.
DPI was introduced as a special implementation of the PI controller, well suited for gain
scheduling applications. Gain scheduling is the process of adapting the control gains with
either external factors, or as a function of some of the system variables. In flight control,
controller gains are usually adapted with variation in static pressure (altitude), dynamic
pressure (airspeed), mach number, mass, center of gravity, etc. Additionally, the gains
are usually adapted for each configuration of flaps, slats and landing gear. When the
gains are adapted as a function of system variables, this introduces an additional feedback.
A classical design technique consists of trimming and linearizing the plant at a set of
envelope trim points. For each trim point, the controller gains are calculated to satisfy
performance, stability and robustness requirements. If the controller gains are naively
interpolated as a function of the scheduling variables, this can lead to oscillations and loss
of stability when applied on the non-linear system. This can be analyzed by trimming and
linearising the entire closed-loop control system including the gain scheduling, and then
assessing the stability. For a classical PI controller, it is seen that additional terms will
appear in the linearization, such that even at the design trim points of the controller, the
calculated requirements might not be fulfilled. These additional terms that appear due to
the gain scheduling are denoted hidden coupling terms (HCT). The DPI controller by [33]
is designed in such a way that the linearization of the gain scheduled controller and the
linear PI representation are identical and hence avoids any hidden coupling terms.
Different general variants exist, but the most simple version is used for a classical PI,
which is also the variant used in the Eurofighter [51]. Hence for the following, this will be
the variant in focus. The DPI algorithm works by differentiating all the measurements for
the control law, and then integrating the control signal in the end. For a linear system,
this does not change anything with regard to the control law, as the differentiation and
integration cancel out. However, for the gain scheduled implementation, which is not a
linear controller due to the interpolation, this will make a difference. In the following, it
will be demonstrated by an example how the DPI implementation avoids hidden coupling
terms. A proof for the general case is given in [33].

Example 2.2.1 (DPI and Hidden Coupling Terms). Consider the point mass motion of
an aircraft along a straight horizontal trajectory. For simplicity, assume that the airspeed
is so fast that the drag is dominated by parasite drag such that:

.
mV = T − kV 2
(2.40)
y=V

23
2.2 Differential Proportional Integral Control

where V is the kinematic speed, T is the thrust, k = 1/2ρSCd,0 and m is the mass. Again,
for simplicity, assume that k = m = 1. Then Eq. (2.40) reduces to:
.
V =T −V2
(2.41)
y = V.

The trim thrust is simply:

T0 = V0 2 , (2.42)
where V0 is the trim velocity. The linearized system is given by:
.
∆V = −2V0 ∆V + ∆T
(2.43)
∆y = ∆V,

where the ∆ denotes deviations from the trim point. In order to control the speed V , a
(very simple) PI controller is designed, which is to be gain scheduled with velocity V :
.
eI = ∆Vref − ∆V

|{z} =
∆T eI − (β − 2V0 ) ∆V
|{z} (2.44)
|{z} | {z }
control integral feedback feedback
feedback gain variable

where eI is the integrator state, Vref is the speed reference, ∆Vref = Vref − Vref,0 with Vref,0
as the trim reference speed, T is the thrust command, ∆T = T − T0 and β > 0 is a tuning
parameter. Note that the feedback gain (β − 2V0 ) is constant for the linear design model,
but will be varying when gain scheduled, i.e. V0 will be changing. In many controllers, this
would be implemented by having the gain vary using a lookup table, however the gain is
shown algebraically in this example to demonstrate analytically the consequence of gain
scheduling on stability. Inserting the controller in Eq. (2.44) into the linear system in Eq.
(2.43), reveals the following linear closed-loop dynamics:
.. .
∆V + β∆V + ∆V = ∆Vref (2.45)

with characteristic polynomial p(s) given by:

p(s) = s2 + sβ + 1. (2.46)

By the Routh-Hurwitz criterion it is confirmed that the system is stable as β > 0.

Consider now the equivalent gain scheduled controller:
.
eI = Vref − V
(2.47)
T = eI − (β − 2V ) V.

The linearized closed-loop system for the gain scheduled controller is obtained by substituting
the control law in Eq. (2.47) into the system given by Eq. (2.41). When linearizing, the
following linear dynamics is obtained:
.. .
∆V + (β − 2V0 ) ∆V + ∆V = ∆Vref (2.48)

24
 Chapter 2: Methodology

.
It can be identified that an additional hidden coupling term −2V0 ∆V appeared in the
linearization. The characteristic polynomial is given by:

p(s) = s2 + s (β − 2V0 ) + 1. (2.49)

This system is unstable for β < 2V0 . If the controller was designed with β < 2V0 for the
trim point V0 , according to the design analysis in Eq. (2.46) the system would be stable,
however when implemented as a gain scheduled controller in the exact same trim point,
the resulting closed-loop system would be unstable.
Consider now the scheduled DPI implementation, where the proportional feedback is moved
behind the integrator:
. .
eI = Vref − V − (β − 2V ) V
(2.50)
T = eI .

Inserting T from Eq. (2.50) into the system equation (2.41), and differentiating once with
respect to time gives the following closed-loop dynamics:
.. . . . .
V = eI − 2V V = Vref − V − (β − 2V ) V − 2V V (2.51)

which reduces to:

.. .
V + β V + V = Vref (2.52)

It is seen that the characteristic polynomial of that system is equal to Eq. (2.46) and
that the linear system is equivalent to Eq. (2.45) and this implementation hence does not
produce any hidden coupling terms.

The above example demonstrated that the implementation of gain scheduled controllers
matters, and that a naive implementation can lead to instability even though the linear
design is stable. It additionally shows that implementing a classical PI controller using
the DPI algorithm recovers the design linear behavior in every linearized envelope point.
This thesis expands the concept of DPI to incremental control laws in paper 2 [71].

Filtered DPI
For practical implementations, the differentiation used in DPI is not possible. Another
version of the DPI exists, in this thesis denoted the filtered DPI. This method was used
in [28] in relation to adaptive flight control. Instead of a direct derivative, a filtered
derivative is used. In order to synchronize the filtering, a corresponding zero is added to
the integration, such that in the linear case, no change is made to the PI controller. It is
shown in [38] that such an implementation also does not introduce hidden coupling terms
when applied with gain scheduling. A similar implementation is developed in this thesis in
paper 2 [71] for incremental control laws.

25
2.3 Eigenstructure Assignment

2.3 Eigenstructure Assignment

Eigenstructure assignment is a method to design control law gains for a linear output
feedback control law. The method has been investigated in e.g. [21, 40, 60]. The following
derivation is based on [30]. Given a linear system described by:
.
x = Ax + Bu
(2.53)
y = Cx

where A ∈ Rn×n is the system matrix, B ∈ Rn×m is the input effectiveness, C ∈ Rr×n
is the output matrix, x ∈ Rn×1 is the state vector, u ∈ Rm×1 is the input vector, and
y ∈ Rr×1 is the output vector.
Note that the flight dynamics often have a direct feed-through to one of the measured
variables (e.g. the normal or lateral load-factor nz , ny ), but when the sensor and actuator
dynamics are included in the design, this does not appear as a D matrix entry, and the
system in (2.53) is appropriate for design.
Consider the output feedback given by:

u = −Ky + r (2.54)

where K ∈ Rm×r is the control feedback gain matrix, and r ∈ Rm×1 is the reference input
vector. Substituting Eq. (2.54) into Eq. (2.53), gives the modified closed loop dynamics:
.
x = (A − BKC) x + Br (2.55)

The goal is to ensure that the new system matrix has a specific set of eigenvalues and
eigenvectors given by a desired system matrix, Ad , desired eigenvalues, λd,k and de-
sired eigenvectors vd,k , where k ∈ {1, ..., r}. Based on the definition of eigenvalues and
eigenvectors, the following can be stated:

λd,k vd,k = Ad,k vd,k = (A − BKC) vd,k , (2.56)

and rearranging the terms gives,

[λd,k I − A + BKC] vd,k = 0. (2.57)

Now write this as a scalar product:

 
vd,k 
[λd,k I − A B]  = 0. (2.58)
KCvd,k

Define the input directions z ∈ Rm×1 as:

zd,k = KCvd,k , (2.59)

such that:  
vd,k 
[λd,k I − A B]  = 0. (2.60)
zd,k

26
 Chapter 2: Methodology

It is important that the desired eigenvalues are distinct from the open loop eigenvalues
such that (λd,k I − A) has full rank. Let N̄λd,k be a basis for the m-dimensional nullspace of
[λd,k I − A B], such that any solution to Eq. (2.60) can be written as a linear combination
with coefficients lλd,k :
   
v
 d,k  = N̄λ lλ = 
Nλd,k 
d,k d,k
lλd,k , (2.61)
zd,k N̂λd,k

where N̄λd,k is split into an upper and lower part such that vd,k = Nλd,k lλd,k and zd,k =
N̂λd,k lλd,k . The dimension of lλd,k is equal to the dimension of the nullspace, which is equal
to the number of controls in u, denoted m. Hence the number of values that can be
specified in each eigenvector can be identified to be m.
Denote s ≤ m the number of eigenvector entries that are sought specified. Choosing less
number of entries can be done to leave room for increasing robustness.
Define now a re-ordered eigenvector, where the first s entries are the ones to be specified:
       
S
vd,k PSd,k
 s×1 =  s×n  vd,k , (2.62)
  
 U  U
vd,k Pd,k
(n−s)×1 s×n
   
where PSd,k and PUd,k are permutation matrices that arrange the entries to be
s×n  s×n  
specified in vd,k
S
, and entries not to be specified in vd,k
U
. Now the specified
s×1 (n−s)×1
part of the eigenvector can be written by combining the top part of Eq. (2.61) and top
part of Eq. (2.62):
     
S
vd,k = PSd,k vd,k = PSd,k Nλd,k lλd,k . (2.63)
s×1 s×n s×n
 
In this thesis s = m, then PSd,k Nλd,k is square (m × m) and the solution can be found
s×n
by the inverse:  −1 
 
lλd,k = PSd,k Nλd,k S
vd,k . (2.64)
s×n s×1

In case s < m then there are several solutions, and other objectives in terms of robustness
can be optimized by choosing the best solution of lλd,k . One could also choose s > m, and
then find a weighted optimal solution by optimization, introducing a weighting matrix
Qλd,k .
The controller gains K can now be found using Equation (2.59) and inserting Equation
(2.61) for each eigenvalue/vector pair:
h i h i
zd,1 zd,2 ... zd,r = KC vd,1 vd,2 ... vd,r , (2.65)

which can be solved for K by:

h i h i−1
K= zd,1 zd,2 ... zd,r C vd,1 vd,2 ... vd,r . (2.66)

General guidelines given by [29] for the choice of eigenvalues and eigenvectors are:

27
2.4 Control Allocation and Redistributed Pseudo Inverse

• Choose only elements necessary for decoupling.

• Assign eigenvalue regions, allows to vary the nullspace.

• Take into account all subsystems (actuators, estimators,...).

• Minimize the modification of the dynamics.

• Remaining degrees of freedom can be used to increase robustness.

In this thesis, eigenstructure assignment is combined with incremental control in paper 3

[73] to control the lateral motion of a fixed-wing aircraft.

2.4 Control Allocation and Redistributed Pseudo In-

verse
In over actuated systems, there are more actuators than channels to control. This is very
often the case in fly-by-wire aircraft, both due to the requirement of redundancy, but also
because of the possibility to satisfy several secondary control objectives. Because of the
surplus of actuators, there are a multitude of ways to deflect the control surfaces in order
to obtain the same aircraft response. Control allocation is the task of choosing which of
the many possibilities that will be commanded to the control effectors.
Consider now the multi-variable case of the control law given by Eq. (2.39), in short
notation it can be written as:
∆u = B† ∆ν, (2.67)
where (·)† is an arbitrary right inverse operator. The control allocation consists of finding
a suitable B† that satisfies the desired control allocation objectives. A common approach
is to use the Moore-Penrose pseudo-inverse [5]. The pseudo-inverse is the solution to a
minimization problem that minimizes the incremental control effort. The optimization
problem is given as follows:
min 12 ∆uT ∆u
∆u
s.t. (2.68)
ν = Bu.
In case the solution exists, i.e. the dimension of ∆u is larger than or equal to the dimension
∆ν and B has full row rank, the solution to the problem is denoted ∆usol = B+ ∆ν where
B+ is given by:  −1
B+ = BT BBT . (2.69)
One issue with this allocation strategy is that ∆u is allowed to take on any value in Rm ,
with m being the dimension of u. In practice, the actuators are restricted in terms of
position and rate limitations. Many different strategies exist to accommodate this. A
strategy denoted redistributed pseudo inverse, introduced in [81], has been chosen due to

28
 Chapter 2: Methodology

its simplicity and deterministic behavior. The algorithm is somewhat heuristic and does
not guarantee to converge to a solution even though one exists. However, it has proven to
work well for practical applications. Control allocation using pseudo-inverse methods is an
active research area and several ideas and extensions have been proposed. For example
the redistributed scaled pseudo inverse, [84], tries to preserve the direction of the solution
by introducing a scaling.

The basic redistributed pseudo inverse (RPI) algorithm is depicted in the flow diagram
in Fig. 2.2. ∆ν t is a temporary pseudo control variable, ∆ut is a temporary control
input and W usat is a diagonal matrix with ones on the diagonal for which actuators are
unsaturated, and zeros for those that are saturated. ∆u and ∆u is the lower and upper
incremental limits due to position, rate saturations and possibly acceleration saturations.
The final ∆u solution is composed of all the saturated controls plus the final allocated
control ∆ut , if any. This algorithm was used for flight control in this thesis, within the
paper [70], due to the simple and almost decoupled configuration of the effectors of the
aircraft under investigation.
In case the actuators are first order, the limit on the incremental control command
due to rate saturations can be calculated directly from the first order differential equation
of the actuator such that: . .
umax umin
∆ūc = , ∆uc = , (2.70)
ωa ωa
where ωa is the eigenfrequency of the actuator and (∆ūc , ∆uc ) are the upper and lower
limits on the actuator command increment. In case of second order actuators, extensive
analysis of approximate limitations is given in [3]. In this thesis, a simple approach is taken,
where the incremental limit is calculated by approximating the second order actuator as a
first order actuator with the modified actuator eigenfrequency given as a function of the
second order system eigenfrequency and relative damping ωa,2o and ζa,2o , where subscript
(2o) refers to the second order system values, and (1o) refers to the approximate first order
equivalent value:
ωa,2o
ωa,1o ≈ , (2.71)
2ζa,2o
such that: . .
umax 2ζa,2o umin ζa,2o
∆ūc = , ∆uc = . (2.72)
ωa,2o ωa,2o

One issue that arises with the incremental formulation of the control allocation problem,
is related to the path dependency. Because only increments are allocated, the pseudo
inverse minimizes the increments, but not the total value of the control signal. This can lead
to the algorithm converging to inappropriate trim solutions. In order to accommodate this,
secondary objectives can be incorporated by using the nullspace of the input effectiveness
matrix. I.e., additional increments that lies in the nullspace of the control effectiveness
matrix can be added without changing the incremental pseudo-control direction. In this

29
2.5 Control Law Assessment

Start

Set 𝚫𝝂𝑡 = 𝚫𝝂 and 𝑾𝑢𝑠𝑎𝑡 = 𝐈𝑘×𝑘

Calculate temporary
pseudo-inverse solution
Δ𝒖𝑡 = (𝑩𝑾𝑢𝑠𝑎𝑡 )+ 𝚫𝝂𝑡

Cut saturated controls

Δ𝒖𝑠𝑎𝑡 = max(min Δ𝒖𝑡 , 𝚫ഥ
𝒖 , 𝚫𝒖)

Check if any controls

FALSE
were saturated:
?
𝚫𝒖𝑡 ≠ Δ𝒖𝑠𝑎𝑡
TRUE

Update identity saturation

matrix with zeros at
saturated controls.
𝑾𝑢𝑠𝑎𝑡

Update 𝚫𝝂𝑡
𝚫𝝂𝑡,𝑜𝑙𝑑 = 𝚫𝝂𝑡
𝚫𝝂𝑡 = 𝚫𝝂𝑡,𝑜𝑙𝑑 − 𝑩(~𝑾𝑢𝑠𝑎𝑡 )Δ𝒖𝑠𝑎𝑡

FALSE
If 𝑾𝑢𝑠𝑎𝑡 = 𝟎
TRUE
In-exact solution-> Exact solution->

Finish
Figure 2.2: Flow diagram of the RPI algorithm.

way, secondary objectives can be incorporated, for example the movement to desired trim
allocation. Other objectives such as system identification [39] by maneuver injection in the
nullspace, maneuver load alleviation, specific force shaping and so on can also be injected
in the nullspace. A specific algorithm to do this is the nullspace transition, as explained
in [84, 85]. The redistributed pseudo inverse with null space transition was used in this
thesis for control allocation in paper 5 [70].

2.5 Control Law Assessment

Prior assessment of a given control system is immensely important before flight testing.
Many branches of engineering are involved for the testing of the overall control system.
Testing of mechanical parts, surfaces, actuators, linkages and their interaction. Also testing

30
 Chapter 2: Methodology

of flight control computers, sensors, network and data communications, power sources
and so on needs to all be appropriately verified before flight. The controller software
on the flight control computer needs to be tested and verified for correctness, execution
performance, latency, maximum delay and so on. In this thesis only the control law is
discussed, meaning the specific equations/blocks that implement the desired functionality
of the controller. Prior to flight, the control law shall be verified to deliver the expected
performance, stability and robustness, failure handling, resilience and recovery. The
following lists typical evaluations of the control law for 3 different branches of testing:

• Performance:

− Tracking performance
− Disturbance rejection
− Handling qualities
− Initialization and resetting

• Stability & Robustness:

− SISO stability margins

− Singular value stability margins
− Monte-Carlo analysis
− Noise amplification evaluation
− Delay and execution time analysis
− Optimization based analysis (e.g. counter optimization)
− Classic µ-analysis
− Probabilistic µ-analysis

• Failure Handling:

− Sensor malfunctions
− Actuator malfunctions
− Engine malfunctions
− Computation resource malfunctions

Although not all methods mentioned above are in the scope of this thesis, thorough
assessment of the performance, stability and robustness have been applied during the
development of the control laws in this thesis. Not all methods were applied directly by
the author, but in connection with colleagues working in the same project. In particular,
novel methods of counter optimization using re-enforcement learning were applied in [14],
where worst case pilot inputs combined with worst case disturbances were investigated in

31
2.5 Control Law Assessment

sense of exceeding flight envelope limits. This analysis was used to gain confidence in the
envelope protections, together with giving guidance in choosing protection parameters in
order to maximize the safety and robustness of the response.
In addition, advanced methods of probabilistic µ-analysis as described in [61, 62] was
investigated by a colleague, in order to increase the confidence in the robustness of the
control law considering probabilistic uncertainties. Design of piloted handling qualities
was investigated in a student thesis [37], and was evaluated with pilot in the loop tests in
another student thesis [24].

32
Chapter 3

Article Introductions

3.1 Paper 1: Nonlinear Dynamic Inversion with Ac-

tuator Dynamics: An Incremental Control Per-
spective
The paper, [73], is reprinted in full in Appendix A, Section A.1.

In this paper the mathematical foundation for the analysis and development of the
incremental control laws is established. The basis for the derivation is non-linear dynamic
inversion, also referred to as input-output feedback linearisation. The concept is often
credited to [32], where the method is used to linearize the system from inputs to outputs
to form a chain of integrators.

In order to arrive at an incremental control law, the concept of nonlinear dynamic

inversion is used on an arbitrary nonlinear system appended with linear first order actu-
ators. Non-linear dynamic inversion is then applied all the way to the actuator output.
In classical NDI, the actuator dynamics are neglected and assumed much faster than the
system dynamics. In this paper the chain rule is used instead to arrive at a formulation
that uses the state derivatives, the increment in actuator command and the Jacobian
matrices of the derived input - differentiated output relationship. This can now be solved
for the required incremental actuator command. The resulting control law is an exact
inversion of the system.

This control law requires knowledge of the system matrix and the state derivatives. In
order to arrive at the popular incremental nonlinear dynamic inversion as introduced in
[4], an approximation can be made in the derived control law, developed by the candidate.
The error dynamics are split between actuator dynamics and system dynamics. These
dynamics are substituted into the exact non-linear control law. The contributions due to
the model dynamics and state derivatives are observed to be inversely proportional to the

33
3.1 Paper 1: Nonlinear Dynamic Inversion with Actuator Dynamics: An Incremental
Control Perspective

actuator bandwidth. If the actuator bandwidth becomes sufficiently large compared to

the dynamics of the system, the system dynamics vanish and the remaining control law is
exactly the incremental non-linear dynamic inversion law.

In case the actuator dynamics are slow compared to the system dynamics, the neglected
terms can be re-introduced in the control law. It is shown in the paper, that, unsurprisingly,
keeping the terms will result in the closed-loop system behaving as the design dynamics.

Based on this analysis, it is clear that the error dynamics in incremental non-linear
dynamic inversion will depend on the size of the neglected terms, which will vary over the
flight envelope. Hence another solution to obtain the desired error dynamics is to introduce
gain scheduling in the error controller. In Paper 2 of this thesis, a simple method is
proposed to obtain the desired dynamics by gain scheduling, while keeping the robustness
of incremental non-linear dynamic inversion.

Author contributions

Rasmus Steffensen did the concept development, mathematical derivations, simulation

results and manuscript preparation. Agnes Steinert did concept discussion, manuscript
review and editing. Ewoud Smeur did concept discussion, manuscript editing and review.

34
3.2 Paper 2: Incremental control as an enhanced
and robust implementation of gain scheduled con-
trollers avoiding hidden coupling terms

The paper, [71], is reprinted in full in Appendix A, Section A.2.

In this paper, a control law is derived by the candidate, that allows for performing
gain scheduling in incremental control laws in a simple and safe manner. The concept is
based on the ideas of DPI control laws [33], which implement gain scheduled PI control
laws such that hidden coupling terms are avoided. The main benefit of the proposed incre-
mental DPI control law, iDPI, is inherent robustness characteristics similar to those of INDI.

The paper starts by introducing an ideal algorithm, developed by the candidate, where
direct derivatives are used for the derivative dependent parts of the control law. This can-
not be implemented in an actual control system, but serves as a basis for the derivation and
understanding of the control strategy. It is shown that this control law exhibits significantly
increased robustness with respect to variation in the main aerodynamic parameters when
considering closed-loop pole locations, time domain tracking and disturbance rejection.

Thereafter, a filtered algorithm is introduced, also developed by the candidate, where

the derivatives are estimated by a low-pass filter with a zero in the origin. The filtered
differentiation has to be synchronized in the incremental part of the control law, as ex-
plained in Paper 3 of this thesis [72]. It is shown that by varying the filter time constant,
the control law moves between the ideal iDPI algorithm and a basic PI. In the limit of
infinite filter time constant (slow), the control law turns into a pure PI, where the other
limit of infinitely small (fast) filter time constant, the control law turns into the ideal iDPI
algorithm. Since the filtered derivative amplifies noise and structural mode couplings, a
compromise has to be chosen for the filter time constant, between achieving the robustness
of the ideal algorithm, and the amount of amplification of noise and structural modes. In
a flight test campaign, the filter time constant can be slowly lowered, when confidence in
the system has been gained, to increase the robust performance of the system in a safe
manner.

Hereafter, a derivation made by the candidate, demonstrates that the gain scheduled
implementation of the filtered iDPI algorithm does not lead to hidden coupling terms.
This is confirmed with simulation on a non-linear model with online pole-placement. It
is shown that a PI designed with the exact same linear closed-loop response as the iDPI
oscillates due to the scheduling of the gains, where the proposed iDPI has no oscillations.
3.2 Paper 2: Incremental control as an enhanced and robust implementation of gain
scheduled controllers avoiding hidden coupling terms

In the end, the proposed approach is applied to an industry level PI controller developed
by FSD. The resulting enhanced control law is verified on a high fidelity simulation of the
TUM/FSD Diamond DA42 Research Platform. The complete simulation demonstrates the
expected increase in robust performance while satisfying the required stability margins.

Author contributions

Rasmus Steffensen did the concept development, theoretical derivations, simulation results
and manuscript preparation. Agnes Steinert assisted in integration with DA42 PI control
law and with review of the manuscript. Florian Holzapfel did manuscript review.

36
 Chapter 3: Article Introductions

3.3 Paper 3: Practical System Identification and In-

cremental Control Design for a Subscale Fixed-
Wing Aircraft

The paper, [74], is reprinted in full in Appendix A, Section A.3.

In this paper, flight test with a subscale fixed-wing aircraft is used to verify the appli-
cability of the proposed iDPI control structure in [71] to a real world application.

A subscale fixed-wing aircraft is equipped with a Pixhawk 4 flight control computer

containing an inertial measurement unit (IMU) with three axis accelerometers and gyro-
scopes. Also a GPS, airspeed sensor, surface and actuator deflection sensors were installed.
Using MATLAB/Simulink and the UAV Toolbox Support Package for PX4 Autopilots,
the iDPI controller and supporting flight code, system identification, signal processing,
mode logics, signal logging, backup control laws and system integration was implemented
by the candidate, and using the Embedded Coder toolbox, the models were build to C++
code, and then compiled and flashed onto the microcontroller in the Pixhawk 4.

First, the aircraft was used to conduct open-loop flight test with sine-sweeps on each
axis. For the data collection for the paper, the aircraft was flown by the candidate with
several flights, to iteratively improve flight test techniques and flight software. The software
for the frequency domain system identification was written by the candidate, and the
analysis, plotting and presentation were performed by the candidate.

The identified models were used by the candidate to design controller gains for the
lateral-directional motion using eigenstructure assignment. The control law was assessed
for stability and robustness using classical methods. The developed Simulink models were
automatically linearized, and the different loop-breaks were checked for gain and phase
margin together with using Nichols and Nyquist charts to check the loop transfer functions
shape.

The controller was flight tested, and the response was compared against the design
response. It was shown that the expected and observed response matched well, and thereby
verified the control method in flight.

37
3.3 Paper 3: Practical System Identification and Incremental Control Design for a
Subscale Fixed-Wing Aircraft

Author contributions

Rasmus Steffensen did the concept development, theoretical derivations, flight tests and
manuscript preparation. Kilian Ginnell assisted with integration of sensors and actuators
in the aircraft, in addition to supporting flight tests. Florian Holzapfel contributed with
manuscript review.

38
 Chapter 3: Article Introductions

3.4 Paper 4: Filter and sensor delay synchronization

in incremental flight control laws
The paper, [72], is reprinted in full in Appendix A, Section A.4.

In this paper, it is investigated how to take into account delays and filtering on the
measurements in incremental control laws. The data from almost all sensors is subject
to measurement delay, which stems from internal sensor delays, transmission delays or
processing delays. Additionally, any filtering of the sensor data will also cause phase delay.
In the sensor, there will often be anti-aliasing filters or low-pass filters for attenuating noise.
In the control law, roll-off/low-pass filters and notch filters are often also applied. Incremen-
tal control laws are especially sensitive to this kind of filtering, if not properly synchronized.

The paper includes an introduction to the concept of complementary filters as the later
derivations are based on incorporating complementary filters as synchronization structures.
The initialization of the complementary filter is derived, which considers the transient
behavior when resetting the filter. A method is derived to initialize the states of the
complementary filter such that no transients are produced by the reset.

Then the different synchronization techniques are compared. The starting point is
the simple INDI control law where the derivative is exactly known without any delay or
filtering, denoted "vanilla" INDI. This is the baseline to which the different synchronizations
are compared to. If the derivative in the "vanilla" INDI is obtained by a filter without
synchronization, the tracking becomes very poor and the closed-loop system can even
become unstable and is demonstrated by an example.
It is shown if one can construct a perfect model based estimate of the derivative, the
proposed complementary filter approach will perfectly recover the "vanilla" INDI with
respect to the influence of the system dynamics (CAX). In terms of disturbance rejection,
the sensor and filter bandwidth of the complementary filter determine the effect on the
closed-loop dynamics. The higher the bandwidth of the sensor and the complementary
filter, the better disturbance rejection is achieved. Another synchronization technique is
the one introduced in the original INDI concept in [4], where the filter for estimating the
derivative is applied on the feedback of u. It is shown that the influence of the system
dynamics on the regulated output is increased due to the sensor and filter dynamics.
Additionally, it is shown that in case of different filtering in the different measurement
channels, this technique will not recover the basic closed-loop structure. Different filters
are often needed when considering attenuation of structural mode couplings in the different
measurement channels. It is shown that moving the filter of the u feedback to a comple-
mentary filter, the original synchronized closed-loop structure is recovered. Similarly, it
is shown for the Hybrid INDI concept proposed in [36], that the influence of the state

39
3.4 Paper 4: Filter and sensor delay synchronization in incremental flight control laws

dynamics on the output derivative can be made independent of the complementary filter
by a modified choice of filter structure. However, if different filtering in the separate
measurement channels exists, the basic closed-loop structure is not recovered. It is shown
that moving the filter on the u feedback to the complementary filter, recovers the basic
closed-loop structure. It is shown that all the synchronization concepts only modify the
effect of the system dynamics CAX and not the disturbance term CBd D.

A simulation example is given comparing the pros and cons of the different synchro-
nization techniques. The example highlights the benefit of using a complementary filter
when different notch filters are used for attenuating structural modes in the different
measurement channels.

Author contributions

Rasmus Steffensen did the concept development, mathematical derivations, simulation

corrections and manuscript preparation. Agnes Steinert did manuscript review and editing.
Zoe Mbikay prepared the initial simulation results and figures. Jorg Angelov, Stefan Raab
and Florian Holzapfel did concept discussion and manuscript review.

40
 Chapter 3: Article Introductions

3.5 Paper 5: Longitudinal Incremental Reference

Model for Fly-By-Wire Control Law using In-
cremental Non-Linear Dynamic Inversion

The paper, [70], is reprinted in full in Appendix A, Section A.5.

This paper proposes several techniques to combine incremental control laws with
structures required to provide a full set of functional features necessary for a piloted fixed-
wing aircraft. The proposed concept incorporates appropriate control variable selection
and blending, straightforward design of excellent handling qualities, considerations for
actuator rate and position constraints, flight envelope protections, pilot relief functions,
seamless transition to a low speed, speed stable configuration supporting control margin
awareness and landing flare capabilities, compensation of non-linear coupling terms for
agile maneuvering and modern control allocation for efficient use of control effectors.

In particular, a split between an incremental part and a base value is proposed. The
incremental part can be considered as the maneuvering part that is commanded directly
by the pilot in the longitudinal plane where the base value contains compensations in the
longitudinal plane due couplings with maneuvering in the lateral and directional plane.

The paper starts by introducing the base value calculation based on the rigid body
non-linear equations of motion in the aerodynamic frame, which are derived in the ap-
pendix of the paper. The key idea of the base value is to avoid large excursions in angle
of attack by compensating for the maneuvering in the lateral and directional plane. The
base value contains contributions related to kinematic coupling, gravity rate and turn
compensation. The base value can be limited to appropriately fade out turn compensation
during rapid rolling. Additionally, the base value can be limited at low speeds to intro-
duce a natural speed stability and by this an intuitive control margin awareness for the pilot.

Hereafter the incremental reference model is introduced. The reference model contains
several subsystems required to shape the desired response. First the incremental short-
period plant model is introduced, which is formulated as a modified linear short-period
system. This allows the dynamics to be shaped to fulfill handling qualities requirements as
given in the respective standards and guidelines in terms of a low order equivalent system.
The reference state of the angle of attack is based on a decrement from the maximum
angle of attack, and is hence straight forward to limit in terms of stall and maximum load
factor. An equivalent pitch acceleration dynamics block is implemented to account for the
actuator dynamics. Non-linear effects such as rate and position limits are propagated to
this block from the control allocation, in order to account for the deficiency resulting from

41
3.5 Paper 5: Longitudinal Incremental Reference Model for Fly-By-Wire Control Law
using Incremental Non-Linear Dynamic Inversion

any of those limits being active. A limiting scheme is applied to avoid counter intuitive
propagations as perceived by the pilot to the desired dynamics. Two main factors were
identified, 1. the reference model command shall not change sign due to the deficiency,
and 2. the reference model command shall not amplify due to the deficiency. A double
dynamic saturation for the deficiency is proposed to achieve those two goals.

Hereafter the control variable selection is introduced. The control variable is determined
by handling qualities and by desired envelope limits. The control variable is changed
depending on the velocity, stick deflection and a selection function using the consolidated
angle of attack or body frame load-factor. The control variable is smoothly blended
between the regions and hence provides a seamless transition for the pilot. The relevant
control variable is therefore always tracked and provides a care-free maneuvering concept
for the pilot. The dynamics are shaped by a reference control law using a pole placement
strategy, together with a shaping filter to account for attitude drop back and required
initial/maximum pitch acceleration.

The next section introduces a concept for error controller blending and gain design, and
at last the control allocation structure is presented together with the method to propagate
the pseudo control deficiency from the control allocation to the reference model.

The resulting control law contains all elements for piloted manual flight that would be
expected for a state of the art, agile, high performance aircraft.

Author contributions

Rasmus Steffensen did the concept development, mathematical derivations and manuscript
preparation. Agnes Steinert did concept discussion, review of mathematical derivations,
manuscript review and editing. Florian Holzapfel did concept discussion and manuscript
review.

42
Chapter 4

Conclusion

In this thesis, concepts for integrating incremental control for fixed-wing aircraft were
developed. The concepts were applied to three different aircraft types, a subscale remote
controlled aircraft, a general aviation aircraft and an agile, high performance turboprop
aircraft.

There were in total three objectives for this thesis: 1. advance the use of incremental
control for fixed-wing aircraft for increased robust performance, 2. develop methods for
using incremental control for piloted high performance fixed-wing aircraft and 3. integrate
proper handling of notch and roll-off filters and compensation of known measurement time
delays.

These three goals were achieved by developing several new techniques for incremental
control and validating the methods by applying them in high fidelity simulation of
two different aircraft with very different characteristics, together with flight test of the
incremental control algorithm on a small subscale remote-controlled aircraft. Incremental
control was implemented for the Diamond DA42 level 6 simulation, as a modification to the
TUM/FSD baseline PI controller. Secondly, incremental control and supporting functions
were implemented for a turboprop demonstrator aircraft, for highly agile maneuvering
including aerobatics and basic fighter maneuvers. At last, iDPI was implemented and
flight tested on the Albadrone subscale demonstrator, a remote controlled fixed-wing V-tail
configuration aircraft.

4.1 iDPI, an incremental control method for classical

controllers
A novel incremental control method named iDPI was developed, a concept which combines
the simplicity of a PI controller with the robustness of incremental nonlinear dynamic
inversion and avoids hidden coupling terms in a similar manner to DPI. The control

43
4.1 iDPI, an incremental control method for classical controllers

method was presented in paper 2 [71] resulting in contribution C1. The control method
includes additional benefits, namely bumpless switching between different gain groups,
easy to avoid integrator wind-up, enhanced robust performance with respect to parameter
variations, enhanced disturbance rejection and increased time delay robustness at the
input. The benefits are obtained with the drawback of the need to measure the effector
position unless estimated by a model, it is more sensitive to sensor noise and unknown
measurement time delays and may result in lower stability margins. A tuning parameter
is contained in the control law, that determines the degree of noise amplification versus
increased robustness. In one corner case, the basic PI controller is recovered, with no
additional robustness benefits and low noise amplification, and in the other corner case,
an ideal incremental control law is obtained with very good robustness qualities and high
noise amplification. The simple tuning parameter determines the relation between noise
amplification and robustness of the control law.
The method can be used for either new control laws, or as a retro fit for PI based
control laws. The control law is very simple, and the structure is similar to the DPI control
concept. Implementation of gain design and assessment routines are straight forward, and
existing tools for PI control laws can be used. The method is intended to be used with
gain scheduling for aircraft with a very large flight envelope, or highly agile aircraft that
operate swiftly between envelope points. As the method does not produce hidden coupling
terms, it is suitable for this type of application.
The concept can also be combined with incremental non-linear dynamic inversion, such
that subsystems such as reference models and control allocation can be directly applied.
This reduces the development time and allows for transfer of knowledge from existing
implementations of incremental non-linear dynamic inversion control laws.

The iDPI concept was tested in simulation as a retrofit for the TUM/FSD DA42
baseline longitudinal PI control law, as presented in paper 2 [71] resulting in contribution
C4. The modification demonstrated a significant increase in robust performance with
respect to variation in the stability and control derivatives and input time delay. The
modification was simple both with respect to implementation, but also with respect to
updating the gain design routines. Hence, it was demonstrated that the concept is viable
for future retrofits in other PI based control laws.

The iDPI concept was also applied to the lateral-directional motion of a small fixed-wing
subscale aircraft as presented in paper 3 [74] resulting in contribution C5. This validated
the use of iDPI for multiple-input-multiple-output systems with real flight test results.
Kinematic relations for the aerodynamic side-slip angle was used to avoid taking the
derivate of the lateral load-factor, which decreased the influence of noise on the load-factor
feedback. The concept showed good performance, and was able to reproduce the predicted
performance from the simulation model.

44
 Chapter 4: Conclusion

4.2 Sensor based incremental control with filter and

delay synchronization

The development of a sensor based incremental control strategy taking into account effector
bandwidth and desired flight dynamic characteristics was presented in paper 1 [73] resulting
in contribution C1. Several techniques to compensate for measurement time delays and
filter dynamics, such as notch, roll-off and anti-aliasing filters were presented in paper 4
[72] resulting in contribution C3. It was shown that the different compensation schemes
affect how the system dynamics affect the controlled output. In particular, it was shown
that having a model of the output derivative, the ideal ("vanilla") INDI control law can
be recovered with a complementary filter in nominal conditions. It was also shown that
choosing a complementary filter approach for estimating the output derivatives allows for
incorporation of different filtering in the different measurement channels. This is useful in
particular if structural modes affect the different measurement channels in a dissimilar
manner, such that different notch filters are required.

4.3 Incremental reference model based control con-

cept for INDI

A novel incremental reference model strategy was developed and presented in paper 5 [70]
resulting in contribution C6. The desired dynamics are split up between a nonlinear base
value and an incremental part related to the pilot maneuvering. The base value contains
turn compensation, gravity rate compensation, kinematic coupling compensation and pilot
relief autopilot functions. The incremental part contains the dynamics directly related
to the pilot maneuvering as commanded by the control inceptor. The incremental part
is implemented as a low order equivalent system and a reference control law shapes the
dynamics to satisfy all relevant handling qualities. This combination has shown to produce
very good results in terms of shaping the handling qualities, providing good compensation
of non-linear effects, and with assisting pilots with appropriate turn compensation during
maneuvering. The method of an incremental reference model with associated base value
was implemented for a fixed base, two pilot tandem simulation of a turboprop demonstrator
aircraft capable of advanced aerobatics and fighter maneuvering. Subsequent pilot in the
loop studies showed good performance of the control law with satisfactory pilot ratings.

To that end, the objectives set out in Section 1.3 were successfully met, and the
results and contributions of this thesis allow for further development and integration of
incremental control laws into future control of fixed-wing aircraft.

45
4.4 Other Research

4.4 Other Research

Beyond the contributions mentioned in this thesis, the author also contributed with
publications that were not included in the thesis. The additional research was published
in the area of flight envelope protections [23, 69], counter optimization based control law
testing [14], flight test based gain tuning [75–77], pilot selection [25], concept of operations
development for an optionally piloted vehicle [44] and system identification [10, 31].

46
Chapter 5

Outlook

This thesis contributed with solutions to many open issues related to integrating incre-
mental control laws into a fully operational setting. During the work, several new insights,
improvements and novel control law concepts were developed. This chapter lays out a
future direction to further advance the methods of incremental control for integration into
industry level products.

One of the incremental control techniques that were developed and studied in this
thesis was the iDPI. In order to further develop the iDPI concept, several concepts can be
investigated for incorporation. First, the inclusion of advanced control allocation could be
investigated. Because of the similarities between iDPI and INDI, the control allocation
methods from INDI may be adjusted to be used within iDPI. This could be used to
more effectively distribute the control action in over-actuated configurations. Additionally,
secondary control objectives could be fulfilled by using null-space methods, which has
already been demonstrated in INDI control laws.

The iDPI control structure could be combined with a reference model in order to do ex-
plicit model following control. Using a physical reference model, which was also investigated
in this thesis in combination with INDI, could be combined with the iDPI controller struc-
ture. The differential-integral formulation of the control law could potentially eliminate
some model parameter dependent errors arising from the reference model as outlined in [71].

It could additionally be studied how to integrate the surface deflection measurements

in the incremental loop for a flight test implementation. The flight test presented in this
thesis used an actuator model to generate the surface deflections for the incremental loop.
In order to use the direct measurements, it should be investigated whether an observer
can be used to generate well behaving signals for use in the incremental-loop.
To verify the robustness properties of the control law, further flight tests could be con-

47
ducted with perturbations to the design model used for gain design. The flight test could
reveal whether the predicted robust performance is also achieved on the real aircraft.

The concepts for piloted control of high-performance fixed-wing aircraft developed

within this thesis could also be applied to the iDPI control technique. It should be investi-
gated which further adjustments to the control law structure are necessary to successfully
combine these concepts. Due to the similarity between iDPI and INDI, it seems reasonable
that this can be achieved in a straight-forward way.

In summary, there is an exciting future ahead for advancing the techniques of incre-
mental flight control for practical applications.

48
Bibliography

[1] Paul Acquatella, E van Kampen, Qi Ping Chu, et al. Incremental backstepping for
robust nonlinear flight control. Proceedings of the EuroGNC, 2013.

[2] Venkata Sravan Akkinapalli and Florian Holzapfel. Incremental dynamic inversion
based velocity tracking controller for a multicopter system. In 2018 AIAA Guidance,
Navigation, and Control Conference, page 1345, 2018.

[3] Jorg Nicolai Angelov. Model-Based Systems Engineering of Flight Control for VTOL
Transition Aircraft. PhD thesis, Technische Universität München, 2023.

[4] Barton Bacon and Aaron Ostroff. Reconfigurable flight control using nonlinear
dynamic inversion with a special accelerometer implementation. In AIAA Guidance,
Navigation, and Control Conference and Exhibit, page 4565, 2000.

[5] Adi Ben-Israel and Thomas NE Greville. Generalized inverses: theory and applications,
volume 15. Springer Science & Business Media, 2003.

[6] Pranav Bhardwaj. Nonlinear Flight Control Strategies for Urban Air Mobility. PhD
thesis, Technische Universität München, 2024.

[7] Pranav Bhardwaj, Stefan A Raab, Jiannan Zhang, and Florian Holzapfel. Integrated
reference model for a tilt-rotor vertical take-off and landing transition uav. In 2018
Applied Aerodynamics Conference, page 3479, 2018.

[8] Pranav Bhardwaj, Stefan A Raab, and Florian Holzapfel. Higher order reference
model for continuous dynamic inversion control. In AIAA Scitech 2021 Forum, page
1130, 2021.

[9] Pranav Bhardwaj, Philipp Müller, and Florian Holzapfel. Design plant adaptation in
jerk-level reference model. In AIAA Scitech 2022 Forum, page 1379, 2022.

[10] Mogens Blanke and Rasmus Stefensen. Dynamic wake identification for model-and
full-scale container ship. IFAC-PapersOnLine, 55(31):463–469, 2022.

49
BIBLIOGRAPHY

[11] Fabrício Reis Caldeira, Marcos Vinicius Campos, Reneu Luiz Andrioli Junior, Wagner
de Oliveira Carvalho, Dagfinn Gangsaas, Eduardo Camelier, Daniel Siqueira, and
Lucas Rubiano. Flight control system mode and method providing aircraft speed
control through the usage of momentary on-off control. US Patent US20130138274A1,
05 2013. Assignee: Embraer SA.

[12] Dan Canin. F-35 high angle of attack flight control development and flight test results.
In AIAA aviation 2019 forum, page 3227, 2019.

[13] Abhijit Chakraborty, Peter Seiler, and Gary J Balas. Susceptibility of f/a-18 flight
controllers to the falling-leaf mode: Nonlinear analysis. Journal of guidance, control,
and dynamics, 34(1):73–85, 2011.

[14] Agnes Steinert David Braun, Rasmus Steffensen and Florian Holzapfel. Counter
optimization-based testing of flight envelope protections in a fly-by-wire control law
using deep q-learning. CEAS EuroGNC, 2022.

[15] Stephane Delannoy. Airbus generic control laws. Preceedings, EuroGNC, 2019.

[16] Stéphane Delannoy and Simon Oudin. Longitudinal control law for modern long
range civil aircrafts. Proceedings of the EuroGNC, 2013.

[17] Department of Defense. Mil-std-8785c: Military standard - flying qualities of piloted

airplanes. Department of Defense, United States of America, March 1993.

[18] CS Droste. Lessons learned in the development of the f-16 flight control system.
Technical Report 1, General Dynamics Fort Worth TX Fort Worth Div, 1983.

[19] EASA. Cs25, certification specifications for large aeroplanes, 2009.

[20] EASA. Cs23, certification specifications for normal utility, aerobatic, and commuter
category aeroplanes, 2012.

[21] Lester F Faleiro. The application of eigenstructure assignment to the design of flight
control systems. PhD thesis, Loughborough University, 1998.

[22] Otto Föllinger. Regelungstechnik. 8., überarbeitete auflage. Hüthig-Verlag, Heidelberg,

1994.

[23] Agnes Christine Gabrys, Rasmus Steffensen, Rafael de Angelis Cordeiro, José Raul
Azinheira, Alexandra Moutinho, and Florian Holzapfel. Integration of phase plane
flight envelope protections in cascaded incremental flight control. IFAC-PapersOnLine,
52(12):429–435, 2019.

[24] Sebastian Gölz. Qualitative Evaluation of Fixed Wing Flight Control Law. Semester
thesis, TUM-FSD, 2022.

50
 BIBLIOGRAPHY

[25] Iñaki González Cabeza, Brett Molesworth, Malcolm Good, Carlo Caponecchia, and
Rasmus Steffensen. Investigating the predictive validity of the compass pilot selection
test. The International Journal of Aerospace Psychology, 31(3):252–268, 2021.

[26] Fabian Grondman, Gertjan Looye, Richard O Kuchar, Q Ping Chu, and Erik-Jan
Van Kampen. Design and flight testing of incremental nonlinear dynamic inversion-
based control laws for a passenger aircraft. In 2018 AIAA Guidance, Navigation, and
Control Conference, page 0385, 2018.

[27] Malo LJ Hautus. Controllability and observability conditions of linear autonomous

systems. Ned. Akad. Wetenschappen, 72:443–448, 1969.

[28] Fabian Hellmundt. Adaptive Flight Control for Fault Tolerance. PhD thesis, Technis-
che Universität München, 2017.

[29] Florian Holzapfel. Advanced flight control lecture slides. TUM-FSD, 2023.

[30] Florian Holzapfel, Orlando da Costa, Matthias Heller, and Gottfried Sachs. Devel-
opment of a lateral-directional flight control system for a new transport aircraft. In
AIAA Guidance, Navigation, and Control Conference and Exhibit, page 6222, 2006.

[31] Barzin Hosseini, Agnes Steinert, Rodolfo Hofmann, Xiang Fang, Rasmus Steffensen,
Florian Holzapfel, and Christoph Göttlicher. Advancements in the theory and practice
of flight vehicle system identification. Journal of Aircraft, 60(5):1419–1436, 2023.

[32] Alberto Isidori. Nonlinear control systems: an introduction. Springer, 1985.

[33] Isaac Kaminer, Antonio M Pascoal, Pramod P Khargonekar, and Edward E Coleman.
A velocity algorithm for the implementation of gain-scheduled controllers. Automatica,
31(8):1185–1191, 1995.

[34] Hassan K Khalil. Nonlinear systems third edition. Patience Hall, 115, 2002.

[35] Chongsup Kim, Changho Ji, Giok Koh, and Nagsun Choi. Review on flight control law
technologies of fighter jets for flying qualities. International Journal of Aeronautical
and Space Sciences, 24(1):209–236, 2023.

[36] Yagiz Kumtepe, Tijmen Pollack, and Erik-Jan Van Kampen. Flight control law design
using hybrid incremental nonlinear dynamic inversion. In AIAA SciTech 2022 Forum,
page 1597, 2022.

[37] Stefan Langen. Handling quality design of fixed wing normal law. semester thesis,
tum-fsd, 2021.

[38] Hugo Lhachemi, David Saussié, and Guchuan Zhu. An enhanced velocity-based
algorithm for safe implementations of gain-scheduled controllers. International Journal
of Control, 90(9):1973–1989, 2017.

51
BIBLIOGRAPHY

[39] Hangxu Li, Stephan Myschik, and Florian Holzapfel. Null-space-excitation-based

adaptive control for an overactuated hexacopter model. Journal of Guidance, Control,
and Dynamics, 46(3):483–498, 2023.

[40] Darren M Littleboy. Numerical techniques for eigenstructure assignment by output

feedback in aircraft applications. PhD thesis, Citeseer, 1994.

[41] Boris Lohmann. Nullstellen und Kompensation im Zustandsraum und ihre Bedeutung
für den Entwurf. Institut für Automatisierungstechnik, Universität Bremen, 1997.

[42] Paul Metz. Flight test of the yf-23a advanced tactical fighter. In Aerospace design
conference, page 1039, 1992.

[43] David G Mitchell, Roger H Hoh, Bimal L Aponso, and David H Klyde. Proposed
incorporation of mission-oriented flying qualities into mil-std-1797a. Technical report,
Systems Technology Inc Hawthorne CA, 1994.

[44] Pranav Nagarajan, Markus Maly, Jerg Jaisle, Paul Gesting, Rasmus Steffensen, Max-
imilian Wechner, Daniel Gierszewski, Christoph Krammer, and Florian Holzapfel.
Taking autonomy out-of-the-loop–novel methodology for the development and au-
tomated operation of uas in integrated airspace. In 2020 AIAA/IEEE 39th Digital
Avionics Systems Conference (DASC), pages 1–10. IEEE, 2020.

[45] Kioumars Najmabadi, Monte R. Evans, Edward E. Coleman, Robert J. Bleeg,

Richard S. Breuhaus, Dorr Marshall Anderson, and Timothy A. Nelson. Aircraft
pitch-axis stability and command augmentation system. US Patent US6325333B1, 12
2001. Assignee: Boeing Co.

[46] Robert C Nelson et al. Flight stability and automatic control. WCB/McGraw Hill
New York, 1998.

[47] Dominik Niedermeie and Anthony A Lambregts. Fly-by-wire augmented manual

control-basic design considerations. In International Congress of the Aeronautical
Sciences, volume 100, page 7, 2012.

[48] Society of Automotive Engineers Aerospace. As94900a, aerospace-flight control

systems-design, installation and test of piloted military aircraft, general specification
for: Aerospace standard. SAE International, Jul, 2018.

[49] Society of Automotive Engineers Aerospace. SAE ARP4761 Rev. A, guidelines and
methods for conducting the safety assessment process on civil airborne system and
equipment. 2023.

[50] Society of Automotive Engineers Aerospace. SAE ARP 4754 Rev. B.: Guidelines for
Development of Civil Aircraft and Systems. SAE, 2023.

52
 BIBLIOGRAPHY

[51] Robert Osterhuber, Martin Hanel, and Rainer Hammon. Realization of the eurofighter
2000 primary lateral/directional flight control laws with differential pi-algorithm. In
AIAA guidance, navigation, and control conference and exhibit, page 4751, 2004.

[52] Simon Oudin. Low speed protections for a commercial airliner: a practical approach.
In AIAA Guidance, Navigation, and Control Conference, page 1023, 2017.

[53] Simon Oudin and Clement Frot. Optimizing take-off performances using a protected
rotation law. In AIAA SCITECH 2022 Forum, page 0619, 2022.

[54] Stefan A Raab, Jiannan Zhang, Pranav Bhardwaj, and Florian Holzapfel. Proposal
of a unified control strategy for vertical take-off and landing transition aircraft
configurations. In 2018 Applied Aerodynamics Conference, page 3478, 2018.

[55] Stefan A Raab, Jiannan Zhang, Pranav Bhardwaj, and Florian Holzapfel. Considera-
tion of control effector dynamics and saturations in an extended indi approach. In
AIAA Aviation 2019 Forum, page 3267, 2019.

[56] HH Roger, DG Mitchell, IL Ashksnas, et al. Mil–std–1797a flying qualities of piloted

aircraft, 1990.

[57] RTCA, Inc. Do-331: Model-based development and verification supplement to do-178c
and do-278a, 2011.

[58] RTCA, Inc. Do-178c: Software considerations in airborne systems and equipment
certification. RTCA, Inc., January 2012. URL [Link]
do-178c. Available at: [Link]

[59] Wilson J Rugh and Jeff S Shamma. Research on gain scheduling. Automatica, 36(10):
1401–1425, 2000.

[60] Atsushi Satoh and Kenji Sugimoto. Partial eigenstructure assignment approach for
robust flight control. Journal of guidance, control, and dynamics, 27(1):145–150, 2004.

[61] David Seiferth, Rubens Afonso, Florian Holzapfel, and Matthias Heller. Reduced
conservatism proof of the balanced mimo gain and phase margins. In 2020 59th IEEE
Conference on Decision and Control (CDC), pages 905–911. IEEE, 2020.

[62] David Seiferth, Johannes Diepolder, Rubens Afonso, and Florian Holzapfel. Proba-
bilistic µ-analysis using mapped uncertainties. In 2021 American Control Conference
(ACC), pages 3670–3676. IEEE, 2021.

[63] Soeren Sieberling, Qiping Chu, and Bob Mulder. Robust flight control using incre-
mental nonlinear dynamic inversion and angular acceleration prediction. Journal of
guidance, control, and dynamics, 33(6):1732–1742, 2010.

53
BIBLIOGRAPHY

[64] Sigurd Skogestad and Ian Postlethwaite. Multivariable feedback control: analysis and
design. john Wiley & sons, 2005.

[65] Jean-Jacques E Slotine, Weiping Li, et al. Applied nonlinear control. Prentice hall
Englewood Cliffs, NJ, 1991.

[66] Ewoud JJ Smeur, Guido CHE de Croon, and Qiping Chu. Cascaded incremental
nonlinear dynamic inversion for mav disturbance rejection. Control Engineering
Practice, 73:79–90, 2018.

[67] Beau Smit, Tijmen Pollack, and Erik-Jan Van Kampen. Adaptive incremental
nonlinear dynamic inversion flight control for consistent handling qualities. In AIAA
SCITECH 2022 Forum, page 1394, 2022.

[68] Aeronautical Design Standard. Performance specification, handling qualities require-

ments for military rotorcraft. US Army Aviation and Missile Command, Redstone
Arsenal, AL, ADS-33E-PRF, 2000.

[69] Rasmus Steffensen, Agnes Gabrys, and Florian Holzapfel. Flight envelope protections
using phase plane limits and backstepping control. In 5th CEAS Conf. Guid. Navig.
Control, 2019.

[70] Rasmus Steffensen, Agnes Steinert, and Florian Holzapfel. Longitudinal incremental
reference model for fly-by-wire control law using incremental non-linear dynamic
inversion. In AIAA Scitech 2022 Forum, page 1230, 2022.

[71] Rasmus Steffensen, Agnes Steinert, and Florian Holzapfel. Incremental control as an
enhanced and robust implementation of gain scheduled controllers avoiding hidden
coupling terms. Aerospace Science and Technology, 141:108500, 2023.

[72] Rasmus Steffensen, Agnes Steinert, Zoe Mbikayi, Stefan Raab, Jorg Angelov, and
Florian Holzapfel. Filter and sensor delay synchronization in incremental flight control
laws. Aerospace Systems, pages 1–20, 2023.

[73] Rasmus Steffensen, Agnes Steinert, and Ewoud JJ Smeur. Nonlinear dynamic inversion
with actuator dynamics: An incremental control perspective. Journal of Guidance,
Control, and Dynamics, 46(4):709–717, 2023.

[74] Rasmus Steffensen, Kilian Ginnell, and Florian Holzapfel. Practical system identifica-
tion and incremental control design for a subscale fixed-wing aircraft. In Actuators,
volume 13, page 130. Multidisciplinary Digital Publishing Institute, 2024.

[75] Agnes Steinert, Florian Holzapfel, Rasmus Steffensen, and Christian Merkl. Flight
test based gain tuning using non-parametric frequency domain methods. In AIAA
Scitech 2021 Forum, page 1424, 2021.

54
 BIBLIOGRAPHY

[76] Agnes Steinert, Rasmus Steffensen, Daniel Gierszewski, Moritz Speckmaier, Florian
Holzapfel, Robert Schmoldt, Frank Demmler, Ulrich Schell, Markus Ornigg, and
Marius Koop. Experimental results of flight test based gain tuning. In AIAA SciTech
2022 Forum, page 2296, 2022.

[77] Agnes Steinert, Rasmus Steffensen, and Florian Holzapfel. Bias of anticipated non-
parametric closed-loop frequency response for controller gain re-tuning. Automatica,
151:110881, 2023.

[78] Ezra Tal and Sertac Karaman. Global incremental flight control for agile maneuvering
of a tailsitter flying wing. Journal of Guidance, Control, and Dynamics, 45(12):
2332–2349, 2022.

[79] James L Townsend and Eugene T Raymond. Background information and user’s
guide for mil-f-9490. Technical report, Boeing Wichita CO KS, 1975.

[80] United States Department of Defense. “mil-dtl-9490e, flight control systems - design,
installation and test of piloted aircraft, general specification for. Military Specification,
2008. MIL-F-9490.

[81] John Virnig and David Bodden. Multivariable control allocation and control law
conditioning when control effectors limit. In Guidance, navigation, and control
conference, page 3609, 1994.

[82] Fubiao Zhang and Florian Holzapfel. Flight control using physical dynamic inversion.
In AIAA Guidance, Navigation, and Control Conference, page 1758, 2015.

[83] Fubiao Zhang, Florian Holzapfel, and Matthias Heller. Nonlinear non-cascaded
reference model architecture for flight control design. In Advances in Aerospace
Guidance, Navigation and Control: Selected Papers of the Second CEAS Specialist
Conference on Guidance, Navigation and Control, pages 407–425. Springer, 2013.

[84] Jiannan Zhang, Pranav Bhardwaj, Stefan A Raab, Saurabh Saboo, and Florian
Holzapfel. Control allocation framework for a tilt-rotor vertical take-off and landing
transition aircraft configuration. In 2018 Applied Aerodynamics Conference, page
3480, 2018.

[85] Jiannan Zhang, Jian Wang, Fubiao Zhang, and Florian Holzapfel. Modeling and
incremental nonlinear dynamic inversion control for a highly redundant flight system.
In AIAA Scitech 2019 Forum, page 1922, 2019.

55
This page is intentionally left blank.
Appendix A

Article Reprints

A.1 Paper 1
Paper 1, reference [73], is re-printed in full length with permission from the authors and
the American Institute of Aeronautics and Astronautics, Inc.

I
 JOURNAL OF GUIDANCE, CONTROL, AND DYNAMICS
Vol. 46, No. 4, April 2023

Nonlinear Dynamic Inversion with Actuator Dynamics:

An Incremental Control Perspective

Rasmus Steffensen∗ and Agnes Steinert∗

Technical University of Munich, 85748 Garching, Germany
and
Ewoud J. J. Smeur†
Delft University of Technology, 2629 HS Delft, The Netherlands
[Link]
In this paper, we derive a sensor-based nonlinear dynamic inversion (NDI) control law for a nonlinear system
with first-order linear actuators, and compare it to incremental nonlinear dynamic inversion (INDI), which has
gained popularity in recent years. It is shown that, for first-order actuator dynamics, INDI approximates the
corresponding NDI control law arbitrarily well under the condition of sufficiently fast actuators. If the actuator
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/1.G007079

bandwidth is low compared to changes in the states, the derived NDI control law has the following advantages
compared to INDI: 1) compensation of state derivative terms, 2) well-defined error dynamics, and 3) exact tracking
of a reference model, independent of error controller gains in nominal conditions. The comparison of the INDI
control law with the well-established control design method NDI adds to the understanding of incremental control.
It is additionally shown how to quantify the deficiency of the INDI control law with respect to the exact NDI law for
actuators with finite bandwidth. The results are confirmed through simulation results of the roll motion of a fixed-
wing aircraft.

Nomenclature
C
r
D; R = space of functions differentiable r times with con- Sub/Superscripts
tinuous derivatives, and D is the domain and R the
c = command
range
ref = from reference model
ey = error in the output
Fu = input effectiveness (B matrix in linear system)
Fx = system dependence (A matrix in linear system)
I. Introduction
I = identity matrix
Im×m
Ki
=
=
identity matrix of size m rows and m columns
error gain matrix in alternative error dynamics O NE of the cornerstones of nonlinear control theory is nonlinear
dynamic inversion (NDI), also called output feedback lin-
ki = error gain matrix earization, which has found applications in many different fields
L = Laplace transform [1–5]. The concept is based on inverting the nonlinearities of a
Lp = roll damping system, such that the relation between a virtual control input and
Lp;d = desired roll damping the output behaves as a linear system, in particular a cascade of
Lu = aileron effectiveness integrators. This transformed system is straightforward to control
M† = any right inverse matrix of M with a linear control law. Very complex nonlinear systems can be
p = roll rate controlled perfectly through this method in theory, but the results
pc = roll rate command deteriorate if the model is not accurate, if some of the system states
s = Laplace variable cannot be measured accurately, or actuator dynamics exist that are
u = input vector not considered [6].
ν = pseudocontrol Incremental nonlinear dynamic inversion (INDI) is a control
x = state vector method that uses a local linearization of the model to derive a control
law to control the defined output and its derivatives, by computation
x
r = derivative of x with respect to time r times
of an increment in the control input, neglecting any state-dependent
y = output vector
terms [7]. Using the feedback of derivatives of the system output,
yc = input to reference model
such as angular accelerations, unmodeled effects and disturbances
yref = output reference
are directly measured and compensated for in the next control incre-
Λ = input scaling gain matrix
ment. This led to an increased popularity of the concept in flight
Ω = matrix of actuator bandwidths
control applications [8–14]. In [15] it is shown that under certain
Ωy = matrix of desired inner-loop bandwidths
simplifying assumptions, the closed-loop system responds to dis-
turbances or unmodeled dynamics with the combined dynamics of
the actuators and any filtering that is done on the output. This has also
been observed in practical experiments as well [15].
Received 27 June 2022; revision received 25 October 2022; accepted for It would seem that if one were to include the partial derivatives of
publication 15 November 2022; published online 14 December 2022. the states with respect to the states in the INDI control law, one
Copyright © 2022 by Steffensen, R., Steinert, A., and Smeur, E. J. J. Pub- would obtain better performance in both tracking and disturbance
lished by the American Institute of Aeronautics and Astronautics, Inc., with rejection. Wang et al. [16] suggest keeping a term with the state
permission. All requests for copying and permission to reprint should be change “Δx” over one time step as part of the control law. The
submitted to CCC at [Link]; employ the eISSN 1533-3884 to
initiate your request. See also AIAA Rights and Permissions [Link]/ drawback of this approach is that when actuator dynamics cannot be
randp. neglected, the computed increment in input is not realized within
*Research Associate, Institute of Flight System Dynamics, Bolzmann one time sample. That makes the considered anticipation of the state
Strasse 15. change insufficient. In fact, this problem is hard to solve if one
†
Assistant Professor, Control and Simulation, Kluyverweg 1, Zuid-Holland. follows the “traditional” derivation of the INDI control law, as will
709
 710 STEFFENSEN, STEINERT, AND SMEUR

be further detailed in Sec. III. Li et al. [17] propose to add a large In Sec. II we derive the proposed incremental NDI control law for a
gain, such that if the actuator behaves as a first-order system, the nonlinear system with first-order linear actuators. In Sec. II.B we
system will achieve the desired value within one controller time derive the resulting control law with error dynamics split between
step, effectively removing any actuator dynamics. This approach is actuation and system dynamics. In Sec. III we compare the conven-
not realistic, as controllers typically run at a frequency much higher tional INDI and INDI with input scaling gain to the derived NDI
than the bandwidth of the actuators. control law. In Sec. IV, we compare the derived INDI control law to
Zhou et al. [18] describe a method of including state-dependent existing INDI control laws that take into account the actuator dynam-
terms in the INDI control law in discrete time. The drawback of this ics. In Sec. V we compare the derived NDI with the INDI control laws
approach is that, due to the discrete formulation, a control input is on a simple example of the rolling motion of a fixed-wing aircraft. In
calculated that will solve for the virtual control within one time step. Sec. VI we discuss the results and in Sec. VII we give the concluding
In most cases, the time constant of the actuators is larger than one time remarks.
step of the controller, which would lead to large inputs to the actuators
and possibly hidden oscillations between samples.
Several concepts have been proposed that scale the control effec- II. Nonlinear Dynamic Inversion Control Law with
tiveness matrix, for different reasons. Cordeiro et al. [19] noted that First-Order Linear Actuator Dynamics
an input gain scaling, which was used to reduce the influence of To introduce the concept, consider the general system with first-
noise, can reduce the closed-loop bandwidth. In [20], it was shown order linear actuator dynamics given by
that the input scaling also increases robustness with respect to time
delays. Pfeifle and Fichter [21] proposed an additional gain that x_  f
x; u
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/1.G007079

depends on the sampling time and the actuator time constant based
on an alternative derivation of the incremental dynamic inversion y  h
x (1)
control law. Raab et al. [22] related the meaning and value of an input
gain scaling to the actuator parameters. where x ∈ Rn is the system state, u ∈ Rk is the actuator state, y ∈ Rm
Raab et al. [22] also suggested that actuator dynamics can be is the output, f ∈ C
r1
Rn × Rk ; Rn , and h ∈ C
r1
Rn ; Rm .
included in the derivation of the controller, by taking an additional The actuator dynamics are given by
derivative of the system output. Essentially, this is achieved by
considering the actuator dynamics as a part of the system dynamics; u_  Ω
uc − u (2)
hence the control law then also inverts the actuator dynamics. The
main benefits of this approach are the ability to incorporate actuator where Ω ∈ R
k×k is a diagonal matrix with constant elements rep-
rate constraints in the control allocation and artificially choose faster resenting the bandwidth of the different actuators. Assume the system
or slower effective actuator dynamics, which is especially useful if to have a relative degree of r ∈ N with respect to u for all elements of
actuators with different effective bandwidth are used to control y; i.e., r is the number of times the output y has to be differentiated
coupled outputs. However, the state-dependent effects are not taken with respect to time for the input u to appear explicitly. The rth
into account in Raab et al. [22]. derivative can be expressed as follows:
In an extension, Bhardwaj et al. [23] based a reference model design
on dynamic inversion including the actuator dynamics. From this y
r  F
x; u (3)
physical reference model a feed-forward control signal is generated
that accounts for state-dependent model effects. This theoretically with F ∈ C1
Rn × Rk ; Rm . Traditionally, this expression for the rth
leads to perfect tracking of the reference model if the reference model derivative of y is used to obtain the control law by inverting for u and
equals the plant and if the states of the reference model equal the plant by relating y
r to the virtual control ν; see, e.g., [6]. In those cases,
states, e.g., if there are no model uncertainties and no disturbances. actuator dynamics are not considered, or assumed to be relatively fast
However, the benefit of this approach is deteriorated if the plant is not such that they can be neglected. Instead, in this paper the nonlinear
exactly following the reference model. If the reference model states are dynamic inversion is continued through the actuator dynamics by
different from the plant states, the feed-forward command will not be performing one more differentiation w.r.t. time:
correct, and this will lead to unpredictable error dynamics. Further-
more, the plant states that are not controlled (zero dynamics) might y
r1  Fx x_  Fu u_ (4)
differ from the corresponding states of the reference model in case that
they are directly influenced by certain control effectors. In case this where Fx ≔
∂F
x; u∕∂x and Fu ≔
∂F
x; u∕∂u. Now the
influence is incorporated in the reference model, some sort of feedback actuator relation in Eq. (2) is substituted into Eq. (4) to obtain a
from the control allocation is required to make sure that these uncon- relation for y
r1 that includes the actuator command.
trolled states do not diverge from the respective plant states such that an
accurate feed-forward command can be achieved, which has not been y
r1  Fx x_  Fu Ω
uc − u (5)
shown how to be accomplished.
The contribution of this paper is based on an alternative derivation Note that traditionally the relation for x_ would be substituted into
of a set of incremental control laws using nonlinear dynamic inver- Eq. (5). Here, we keep x_ to allow it to be possibly determined by
sion. In particular, note the following: measurements. Choose now y
r1  ν, where ν is a virtual control
1) This paper formulates a novel sensor-based incremental non- command. If it is assumed that Fu Ω has full row rank, the following
linear dynamic inversion control law taking into account state-
choice of uc will make y
r1  ν:
dependent terms in a consistent and straightforward manner.
2) It formulates error dynamics split into actuation and system
uc 
Fu Ω†
ν − Fx x
_ u (6)
dynamics. This allows specifying actuation dynamics according to
actuator limitations and system dynamics according to vehicle eigen-
response requirements. where
Fu Ω† denotes a right inverse matrix that solves the linear
3) An alternative derivation of the incremental nonlinear dynamic equation system given in Eq. (5). The control signal uc linearizes the
inversion control law is presented based on the above-mentioned response from the virtual control input ν to the output y to a chain of
error dynamics. The derivation shows that in case the actuators can be integrators.
assumed to be fast in relation to the state dynamics the traditional
INDI law approximates the derived exact NDI control law. A. Desired Error Dynamics
4) In addition, an alternative derivation of incremental nonlinear Often a linear controller is used to regulate the output y. In this
dynamic inversion with input scaling gain is presented together with an paper, for the next step, the virtual control input ν is designed using a
associated interpretation of inner-loop bandwidth reduction/increase. linear error controller such that the specified error dynamics are
 STEFFENSEN, STEINERT, AND SMEUR 711

achieved. The order of the error dynamics in y will correspond to the straints in the actuators. It therefore makes sense to design the error
sum of the relative degree and the order of the actuator dynamics, controller such as to include actuator dynamics with a bandwidth
such that the error dynamics could be specified by equal or comparable to the real actuators. Furthermore, as will be
shown in Sec. III, the proposed formulation of the error dynamics
X
r
allows a direct comparison to the INDI control law.
e
r1
y  ki e
i
y 0 (7) The order of the error dynamics in y corresponds to the sum of the
i0
relative degree of y and the order of the actuator dynamics. The
desired dynamics can therefore be interpreted as cascaded dynamics
where ey  yref − y is the error in y, e
i
y is the ith derivative with composed of the slower desired system error dynamics and a faster
respect to time, and y
r1 is chosen as the virtual control ν such that desired inner-loop dynamics with bandwidth Ωy , corresponding to
the actuator dynamics. It can hence make sense to express the desired
X
r
ν  y
r1
ref  ki e
i
y (8) error dynamics Eq. (7) in the Laplace domain as the product of
i0 dynamics due to the system physics and due to the actuators as
follows:
Inserting ν into Eq. (6) results in the control law:  
X
r−1
 X
r  Ey
s s I 
r
Ki s
sI  Ωy   0
i
(11)
uc 
Fu Ω† y
r1
ref  ki e
i
y − F x x
_ u (9) i0
i0
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/1.G007079

where ey  yref − y is the error in y with Laplace transform Ey
s,

The resulting control law is an NDI control law with linear error  P 
controller, where the actuator dynamics were additionally included in sr I  r−1
i0 Ki s
i
describes the desired error dynamics with
the system dynamics and inverted. By inserting Eq. (9) into Eq. (5), respect to the system,
sI  Ωy  is the desired error dynamics due
the desired error dynamics are obtained as designed, because Fu Ω to the first-order actuators, and Ki are the new error controller gains
has full row rank such that
Fu Ω
Fu Ω†  Im×m : defining the
 desired
P
system error
 dynamics. The system error dynam-
 X
r  ics Ey
s sr I  r−1 K
i0 i s i can be formulated in the time domain

r1
i
y
r1 _ †
 Fx x  Fu Ω
Fu Ω yref  ki ey − Fx x
_ as
i0
X
r   X
r−1  X
r−1
 y
r1
ref  ki e
i
y (10) L−1 Ey
s sr I  K i si  e
r
y  Ki e
i
y 0 (12)
i0 i0 i0

This leads to tracking of the corresponding derivative of a reference where Ki are the error gain matrices and L−1 is the inverse Laplace
model y
r1
ref , and the resulting error dynamics will correspond to transform. Using Eq. (12), the combined error dynamics from
Eq. (7). Eq. (11) can be described in the time domain by
This control law can be compared to the control law proposed by
Bhardwaj et al. [23], where the state-dependent term Fx x_ in the    
d
r X r−1 Xr−1
control law in Eq. (9) is generated by a physical reference model ey  Ki e
i
y  Ωy ey
r  Ki e
i
y
using an additional feed-forward term corresponding to Fx x_ ref . In dt i0 i0
that case, the feed forward from the reference model is only correct if Xr  X
r−1 
the system is tracking the reference model and the states correspond-  e
r1
y  Ki−1 e
i
y  Ωy ey 

r
Ki e
i
y 0 (13)
ing to the zero dynamics are matching the respective plant states. In i1 i0
addition, in case of disturbances that lead to a perturbation with
differences between the desired and actual trajectory, the actual error Note that if a formulation as given in Eq. (7) is preferred, Eq. (13)
dynamics will differ from the desired error dynamics, because they can be reformulated as
will be excited by the term Fx
x_ − x_ ref  as can be seen in Eq. (10).‡
However, in the proposed inversion-based control law, where the
state derivatives are used, the error dynamics will correspond to the X
r−1
e
r1
y 
Kr−1  Ωy e
r
y 

Ki−1  Ωy Ki e
i
y   Ωy K0 ey  0
desired error dynamics. A possible drawback of this approach could i1
be reduced stability margins when analyzing the linearized closed-
loop system due to the additional feedback, but this requires further (14)
investigation. It will become apparent in Sec. III that the classical
INDI approximates this NDI law under the condition that the actua- and the gain matrices ki can be obtained directly. Inserting e
r1
y 
tors are fast with equal bandwidth. y
r1 − y
r1 into Eq. (13) and solving for y
r1 provides the
ref
pseudocontrol input ν that can be used in Eq. (6):
B. Consideration of Actuator Dynamics in the Error Dynamics
In this section we propose a particular structure of the error X
r  X
r−1 
dynamics in order to split up the error tracking in terms of actuators ν y
r1
ref  Ki−1 e
i
y

r
 Ωy ey 
i
Ki ey (15)
and system dynamics. This new formulation also serves effective in i1 i0
deducing the approximations required to obtain the popular INDI
control law, as will be shown in Sec. III. The error dynamics in Eq. (7) with reference dynamics given in terms of the respective derivatives
has a generic form and is shaped by choice of the gains ki . Some of yref , and error controller gains given by Ωy and Ki . The final
choices of these gains could lead to error dynamics that cannot be control law is then given by
realistically obtained with real-world actuators, due to physical con-
 X
r

‡ uc 
Fu Ω† y
r1

ref  Ki−1 e
i
y
Another subtle issue that can arise is if the system B matrix and the Fu
i1
matrix used for control allocation are different, because the system contains   
states that are not directly controlled by the inversion law, then the dynamics of X
r−1
the states in the reference model that are not controlled can depend on the  Ωy e
r
y  Ki e
i
y − Fx x_  u (16)
control allocation. i0
 712 STEFFENSEN, STEINERT, AND SMEUR

The resulting error dynamics will correspond to Eq. (11), which not be instantaneously reached by the actuator, due to the actuator
can be interpreted as cascaded dynamics composed of the slower dynamics, the Fx
x − x0  term in the y
r dynamics of Eq. (17) will
desired system dynamics and a faster desired inner-loop dynamics not be exactly canceled by the control signal u
t. The resulting
bandwidth Ωy , due to the actuator dynamics. Since Eq. (16) essen- stability properties of the INDI control law are unclear, neither is it
tially inverts the actuator dynamics as well, the system could be made clear what was neglected from the Taylor expansion. The virtual
arbitrarily fast through the choice of Ωy . However, for practical control ν in Eq. (21) is derived by specifying the desired error
applications, the actuators were designed to operate up to a certain dynamics in y. Usually, the desired error dynamics are specified by
bandwidth, and the choice Ωy should be limited to the design Eq. (12), such that the virtual control is given by
bandwidth of the actuators accordingly. In some cases, the dynamics
of a control effector is limited by the effector dynamics and not the X
r−1

physical actuator. Here an increase in effector bandwidth can be ν  y
r

ref  Ki e
i
y (22)
obtained by selecting Ωy larger than the actuator bandwidth in the i0

considered direction. The second degree of freedom consists of the

choice of the error gain matrices Ki ’s. The Ki matrices should be Combining Eqs. (21) and (22), the following INDI control law is
chosen to satisfy requirements on the error dynamics related to the found:
system dynamics. This will be demonstrated in the example in Sec. V,  X
r−1 
where the roll damping is increased for a fixed-wing aircraft. At last, uc  F†u y
r
r
ref − y0  Ki e
i
y  u0
the closed-loop dynamics from reference input yc to response y is i0
determined by the choice of reference model dynamics, i.e., the  X 
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/1.G007079

r−1
dynamics from yc to yref . †
r
i
 Fu ey  Ki ey  u0 (23)
i0

III. Comparison with Incremental Nonlinear Dynamic By comparing the above control law with the exact NDI law in
Inversion Eq. (16), it is possible to identify which part was neglected. In the
This section compares the NDI control law, which was derived in following, it is shown that under certain conditions relating to the
Sec. II, to an INDI controller, which has been derived, e.g., by Bacon actuators and the NDI control law parameters, the INDI law approx-
and Ostroff [7] and Sieberling et al. [24]. The INDI controller will be imates the exact NDI law. Assume the following:
derived along the same lines, keeping the nomenclature the same as in 1) All actuators have the same bandwidth, i.e., Ω  ωIk×k , where
the previous section, such that the controllers can be effectively ω is a positive scalar.
compared. The classic INDI derivation contains some inaccuracies, 2) Choose Ωy  ωIm×m .
which are pointed out in this section as well. Then
Again, consider the system in Eq. (1), and the rth derivative of the
output as in Eq. (3). Take the Taylor expansion of F
x; u with respect
Fu Ω† Ωy 
Fu ωIk×k † ωIm×m 
Fu ω† ω  Fu †
1
ω  Fu †
to x and u: ω
(24)
y
r
t  F
x
t; u
t  F
x0 ; u0   Fx
x
t − x0 
Hence, the control law from Eq. (16) evaluates to
 Fu
u
t − u0   O
Δx; Δu (17)
 
1 †
r1 X r
where Fx ≔
∂F
x; u∕∂x and Fu ≔
∂F
x; u∕∂u and O denotes uc  Fu yref  Ki−1 e
i
y − Fx x
_
ω
higher-order terms. The term F
x0 ; u0  can also be denoted by y
r
0 . It
i1
should be noted that x and u are functions of time. The Taylor  X
r−1 
expansion in Eq. (17) is performed with respect to x and u. For x0  Fu † e
r
y  K e
i
i y u (25)
and u0 it makes sense to choose a state and control input, correspond- i0
ing to the system, a small time instance ago; i.e., define
It is seen here that, in the limit where the actuator bandwidth ω
x0  x
t − Δt approaches infinity, the NDI law turns exactly into the INDI control
law of Eq. (23):
u0  u
t − Δt (18)
  
1 †
r1 X r

Ify
r is chosen as virtual control ν and u is chosen as the control uc  lim Fu yref  Ki−1 e
i
y − Fx x
_
ω→∞ ω
signal uc , the following relation is obtained: 
i1
 

r
X
r−1

i
†
 Fu ey  Ki ey  u
ν  y
r
0  Fx
x − x0   Fu
uc − u0   O
Δx; Δu (19)
i0
 X
r−1 
It is assumed that the higher-order terms can be neglected. The
 Fu † e
r
y  K e
i
i y u (26)
control law is obtained by solving Eq. (19) for uc :
i0

uc  F†u
ν − y
r
0 − Fx
x − x0   u0 (20) The NDI law of Sec. II perfectly inverts the system, so by comparing
the INDI law given in Eq. (23) with the exact sensor-based NDI law
The term Fx
x − x0  is usually neglected, with several different given by Eq. (25) reveals exactly which part is neglected by applying
arguments involving the bandwidth of the actuator and a sufficiently the INDI control law, namely, the first term in Eq. (25). These missing
small sample time [24,25], leading to the following control law: terms will lead to errors in reference model tracking, and lead to error
dynamics that are different from the designed error dynamics. It is
uc  F†u
ν − y
r
0   u0 (21) seen that the mismatch between the INDI law and the exact sensor
based NDI law is vanishing for sufficiently large actuator band-
If the term Fx
x − x0 , or some approximation of this, is sought to be width ω.
included, it is not clear how Δx in Eq. (20) should be chosen. Wang This comparison shows that the derivation that was performed in
et al. [16] proposed to use Δx  x
t_ − ΔtΔt with Δt chosen as the Sec. II is a useful alternative to arrive at the incremental nonlinear
sample time. However, because the actuator command, Eq. (20), will dynamic inversion law as it does not require any ad hoc arguments.
 STEFFENSEN, STEINERT, AND SMEUR 713

As such, the derivation provides new theoretical support for the INDI This control law equals the NDI control law in Eq. (16) if Ω  ωIk×k ,
control method, while also providing a means to compensate for Ωy  ωIm×m , and Fx x_ is neglected. This approach requires all
model-dependent terms (Fx x) _ in the control law. In INDI, these terms actuators to have the same bandwidth; otherwise Eq. (29) does
are not compensated for, and depending on the system, this may lead not hold.
to significant errors in tracking and unpredictable error dynamics. In Ref. [22] an approach can be found to resolve this problem. With
As shown in the Appendix, the INDI law does not hold in the limit; the error dynamics proposed in Eq. (14), the control law proposed in
i.e., it is only valid for ω < ∞; hence the above equation essentially [22] can be formulated using the notation of this paper as
states that, for a fixed time, the INDI law approximates the true
NDI law arbitrarily well by the choice of sufficiently high actuator  X
r  X
r−1 
bandwidth. uc 
Fu Ω† y
r1
ref  K e
i
i−1 y  Ωy e
r
y  K e
i
i y u
It can further be shown that the input scaling as suggested by i1 i0
Cordeiro et al. [19,20] for reducing the influence of noise and (32)
increasing the robustness with respect to time delays can be inter-
preted as a modification of the innermost bandwidth of y: Ωy  Λω,
This is again equal to the corresponding NDI law in Eq. (16) with Fx x_
where the input scaling gain matrix Λ ∈ Rm×m is diagonal. If this
relation is inserted into the NDI control law in Eq. (16), and it is neglected. Compared to Eq. (31), it does not have the assumption of
assumed that the actuators have equal bandwidth, i.e., Ω  ωIk×k , equal actuators and it offers the option to specify a desired Ωy .
and taking the limit of the bandwidth ω going to infinity, we obtain
the conventional INDI control law with the scaling gain matrix that
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/1.G007079

was proposed by Cordeiro et al. [19]: V. Simple Example

   A. Derivation of Roll Rate Control Law
1 †
r1 X r
uc  lim Fu yref  Ki−1 e
i
y − F x x
_ In the following, the mechanics of deriving the control law is
ω→∞ ω
i1 demonstrated on a simple single-input–single-output (SISO) linear
 Xr−1   system. The roll rate of a fixed-wing aircraft is to be controlled.
1
 Fu † Λω e
r y  Ki e
i
y u Consider the roll dynamics given by
ω i0
 X
r−1  p_  Lp p  Lu u (33)
 Fu † Λ e
r
y  Ki e
i
y u (27)
i0 where p is the roll rate, Lp < 0 is the roll damping, u is the aileron
deflection, and Lu < 0 is the aileron roll effectiveness. Lp is usually
Cordeiro et al. [19] observed that choosing the diagonal elements large and hence contributes significantly to the dynamics. This
of Λ smaller than 1 reduces the closed-loop bandwidth. With this example demonstrates how it is taken into account by the proposed
formulation of the control law, the scaling gain can be directly iden- control concept. The ailerons are driven by an actuator with the
tified as a modification of the desired innermost bandwidth. If the following first-order dynamics:
scaling factor is 1, as in conventional INDI, this corresponds to a
desired innermost bandwidth equal to the actuator bandwidth. u_  ω
uc − u (34)

IV. Comparison with Incremental Nonlinear Dynamic where uc is the actuator command and ω is the actuator bandwidth.
Inversion Including Actuators The output to be controlled is y  p. The relative degree r of the
system is 1 as given by Eq. (33). The control relation as given by
Smeur et al. [15] showed that when actuator dynamics are present Eq. (5) is
and the INDI control law in Eq. (21) is applied, then y
r follows ν
with the dynamics of the actuator, in the case that the term Fx x_ in p  Lp p_  Lu ω
uc − u (35)
Eq. (4) is neglected, and when all actuators have the same bandwidth,
i.e., Ω  ωIk×k . For linear first-order actuators, Eq. (4) can then be
expressed by Solving for uc as explained in Eq. (6) gives the following control
law:
y
r1  Fu u_  Fu ω
uc − u (28)
1
uc  _ u

p − Lp p (36)
where the actuator relation in Eq. (2) was substituted. Inserting the Lu ω d
INDI control law from Eq. (21) into Eq. (28) results in
As depicted in Fig. 1, a reference model is chosen based on the
y
r1
r
 ω
ν − y  (29) desired dynamics:

i.e., y
r follows ν with the dynamics of the actuator. Based on p_ ref  −Lp;d
δ − pref 
Eq. (29) a new control law can be derived. First the virtual control δ_  ωd
pc − δ (37)
is defined by
where Lp;d < 0 is the desired roll damping and δ is a generalized roll
1
r1
ν y  y
r (30) acceleration due to the aileron deflection.
ω The dynamics of δ is given by the desired buildup in roll accel-
Choosing the same error dynamics like in the NDI control law, given eration due to the aileron dynamics ωd ; pc is the pilot roll rate
command, since in steady state, pref  pc . Note that p ref can easily
by Eq. (14) with Ωy  ωIm×m ; solving for y
r1 ; substituting this
be calculated using the reference dynamics in Eq. (37):
relation for y
r1 in Eq. (30); and inserting the resulting ν into the
INDI control law given by Eq. (21) result in the control law
+ -
   1 1
1
r1 X r X
r−1
- +
,
uc  F†u yref  Ki−1 e
i
r
y  ω ey  Ki e
i
y  u0
ω i1 i0
(31) Fig. 1 Block diagram of the reference model.
 714 STEFFENSEN, STEINERT, AND SMEUR

p ref  −Lp;d
ωd
pc − δ − p_ ref  (38)

To choose p d , the desired error dynamics have to be formulated.

Since the actuators are of first order and the system dynamics are first
order, a product of two first-order systems is chosen as follows:

Ep
s  ωd 
s − Lp;d   0 (39)

where ep  pref − p, which in the time domain is

ep  ωd e_ p − Lp;d
e_ p  ωd ep   e p  k1 e_ p  k0 ep  0 (40)

where

k1 
ωd − Lp;d 
k0  −Lp;d ωd (41) Fig. 2 Step response of the INDI and NDI control law.

Solving for p  p d,
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/1.G007079

p d  p ref  k1 e_ p  k0 ep (42)

will lead to the final control law by substituting Eq. (42) into Eq. (36):

1
uc  _ u

p  k1 e_p  k0 ep − Lp p (43)
Lu ω ref

which can be ordered in terms of the contributions:

0 1
1 @
uc  p ref  k1 e_ p  k0 ep − Lp p_ A  u (44)
Lu ω |{z} |{z} |{z}
Feed-forward Error control Model part

This can be compared to the standard INDI control law (23) for the
same example Fig. 3 Comparison of desired and actual error dynamics considering a
5°∕s initial value perturbation.
1
uc 
e_ − Lp;d ep   u (45)
Lu p
exactly to the desired error dynamics given by Eq. (40). The resulting
with reference dynamics:
error dynamics of the INDI controller do not correspond to the
p_ ref  −Lp;d
pc − pref  (46) specified error dynamics e_ p − Lp;d ep  0.
Figure 4 compares the NDI control law with the INDI control law
where actuator dynamics are taken into account in the error controller
B. Simulation Results design. Both laws use the same reference dynamics. This extended
For the example described above, simulations were conducted to INDI controller still does not realize perfect tracking of the reference
validate the approach. In the simulations, it was assumed that state signal of the NDI law, the reason being that in the INDI controller the
information was available without noise, and there were no uncertain- Lp p_ term is neglected. In [23], this issue was fixed for the extended
ties in the parameters. The parameters were Lu  0.25
deg ∕s2 ∕ deg, INDI controller by adding an additional reference-model-based feed-
Lp  −6.6
deg ∕s2 ∕
deg ∕s, Lp;d  2Lp  −13.2
deg ∕s2 ∕ forward term. The resulting control law in [23] can be reformulated

deg ∕s, and ω  20 rad∕s. In the following we will compare
1) NDI as given in Sec. II, for this example given by Eq. (43), with
reference dynamics given by Eqs. (37) and (38),
2) INDI as given in Sec. III, for this example given by Eq. (45) with
reference dynamics given by Eq. (46), and
3) INDI with actuators as given in Sec. IV, for this example given
by Eq. (43) without Lp p,_ with reference dynamics given by Eqs. (37)
and (38).
Figure 2 compares the responses in roll rate p to a step input of
5°∕s in pc for the INDI and NDI control law. The NDI law makes the
plant correctly follow the reference model with reference inputs,
while this is not the case for INDI with its corresponding reference
dynamics.
Figure 3 reveals the error dynamics of the classical INDI control
law in Eq. (45) and the proposed NDI law in Eq. (44). The closed-
loop system response for both control laws was simulated with a
roll rate command pc  0°∕s for an initial value perturbation of
p  5°∕s; i.e., the initial value of the plant was 5°∕s, while the initial
value of the reference model was pref  0°∕s. The simulation shows Fig. 4 Step response of the INDI with actuators in the design and NDI
that the resulting error dynamics of the NDI controller correspond control law.
 STEFFENSEN, STEINERT, AND SMEUR 715

Fig. 7 Difference between NDI error dynamics and design INDI error
Fig. 5 Comparison of desired and actual error dynamics considering a dynamics.
5°∕s initial value perturbation.
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/1.G007079

to resemble the NDI control law given by Eq. (44), with the only design, i.e., e_ − Lp;d ep  0, but for increasing actuator bandwidth
difference that the term Lp p_ is substituted with Lp p_ ref from a phy- these dynamics are approached by the resulting error dynamics.
sical reference model, such as the one given in Eq. (37). Under the 2) The desired NDI error dynamics given by Eq. (40) approach the
condition that p_  p_ ref (i.e., no error exists) and the plant is exactly INDI error dynamics given by e_ − Lp;d ep  0, for increasing actuator
equal to the reference model, this leads to perfect tracking of refer- bandwidth. Because the INDI error dynamics do not explicitly con-
ence inputs, similar to the NDI law. However, the condition that Fx x_ sider actuator dynamics, then when the actuator bandwidth incre-
equals Fx x_ ref might not hold in the following cases: 1) dynamics and ases, the actuator part of the error dynamics in the NDI law becomes
couplings that are not modeled in the reference dynamics but are negligible and hence approaches the INDI error dynamics.
present in the plant; 2) a control allocation mismatch for the reference For statement 1, consider Fig. 6, which reveals the difference
model and the INDI controller due to disturbances, model uncertain- between the error ep resulting from simulation of the closed-loop
ties, and utilization of overactuated channels for secondary control system with the INDI control law, and the initial value response of the
objectives, leading to different responses; and 3) disturbances. desired error dynamics e_ − Lp;d ep  0. It shows that for increasing
In Fig. 5 the discrepancy is shown for the case that the reference bandwidth of the actuator this difference decreases, meaning that the
model state is zero, but the plant state is not. Here, the closed-loop resulting error dynamics approach the specified desired dynamics of
system responses of the NDI control law and the INDI with actuators the INDI law. For statement 2, consider Fig. 7, which depicts the
were simulated for a roll rate command pc  0°∕s with an initial difference between the desired INDI error dynamics and the desired
value of the plant of 5°∕s, while the initial value of the reference NDI error dynamics, for increasing actuator bandwidth. Hence, the
model was pref  0°∕s, such that p_ ≠ p_ ref  0. It is revealed that, for conclusion can be drawn that, with increasing actuator bandwidth,
the NDI law in Eq. (44), the dynamics with which the error pref − p the actuator part of the error dynamics in the NDI law becomes
declines corresponds exactly to the desired error dynamics specified negligible, and the NDI error dynamics approach the INDI error
by Eq. (40). The INDI control law with actuators was designed with dynamics.
the same desired error dynamics like the NDI law, and it is seen that
the resulting error dynamics do not correspond to these desired
dynamics. VI. Discussion
Finally, Figs. 6 and 7 investigate the influence of the actuator One of the potential issues that could be raised with the proposed
bandwidth on the error dynamics of the classical INDI and the NDI sensor-based NDI control law as given in Eq. (6) is the reliance on
control laws. The same simulations as before with pc  0°∕s and an additional state information. The question is if the respective signals
initial condition p  5°∕s was performed, but with varying actuator can be measured or otherwise obtained. This is something that is
bandwidth. The simulations show the following: dependent on the system under consideration. In terms of the system
1) For the INDI law in Eq. (45) the dynamics of pref − p do not output, Eq. (26) shows that the requirement is the same as for an INDI
exactly correspond to the error dynamics that were specified in the controller; the output up until y
r should be available. Additionally,
the NDI control law requires the state derivative information x. _
Depending on the system, this may overlap with the derivative of
the output. In many cases, the requirements on available signals may
therefore be the same, or slightly increased compared to a regular
INDI controller.
In some cases, filtering of these noisy derivative signals may be
required. Filtering leads to delay, which will lead to deterioration of
the controller performance. For an INDI controller, this problem can
be circumvented with filter synchronization of output and input filters
[15], but that particular method cannot be applied to the feedback of x_
in the proposed controller, as there is no signal to synchronize with.
Instead, a complementary filter could be used, where the high
frequencies are coming from a model, and the low frequencies from
the filtered measurement [26,27]. If the system under consideration is
already stable, the proposed method could improve the tracking
behavior, but the feedback of the state derivatives could influence
the stability and robustness in the case of model uncertainty and
measurement errors. How this compares to classical INDI has to be
Fig. 6 Difference between actual INDI error dynamics and design INDI compared on a case to case basis and should be further investigated in
error dynamics. future research. The proposed NDI control law additionally requires
 716 STEFFENSEN, STEINERT, AND SMEUR

information about the actuator bandwidth, which might also be When taking the limit of ω tending to infinity, Eq. (48) does not tend
uncertain. In addition, the actuators might not be first order. In that to the desired error dynamics. Hence, even in the limit, the INDI does
case, an effective bandwidth of the actuator can be used for the control not produce the correct error dynamics. INDI only approximates the
law design. true NDI control law signal uc arbitrarily well as given in Eq. (26).

VII. Conclusions Acknowledgments

The proposed derivation of the NDI control law considering first- The authors would like to thank Stefan Raab and Florian Holzapfel
order linear actuator dynamics, which one could call actuator NDI for the fruitful discussions of incremental nonlinear dynamic inver-
(ANDI), theoretically provides a perfect inversion of a system with sion concept development and their many interesting insights into the
such actuators, also in case the actuators do not all have the same theory behind incremental control laws. The authors would also like
bandwidth. It allows to assign an innermost desired bandwidth for the to thank the reviewers for their qualified comments and suggestions.
tracking variable different from the actuator bandwidth. The ANDI
control law leads to an INDI control law if infinitely fast actuators are References
considered. This provides a better theoretical foundation for INDI
[1] da Costa, R. R., Chu, Q. P., and Mulder, J. A., “Reentry Flight Controller
control, from which it can be readily seen what the impact is of the
Design Using Nonlinear Dynamic Inversion,” Journal of Spacecraft and
terms that are neglected in the INDI derivation. Moreover, the derived Rockets, Vol. 40, No. 1, 2003, pp. 64–71.
NDI control law compensates for state-dependent terms, which are [Link]
not taken into account with the classical INDI formulation. Com- [2] Lane, S. H., and Stengel, R. F., “Flight Control Design Using
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/1.G007079

pared to the reference-model-based feed-forward control from liter- Non-Linear Inverse Dynamics,” Automatica, Vol. 24, No. 4, 1988,
ature, the benefit of the proposed control law is twofold. First, the pp. 471–483.
inversion is based on the actual vehicle states, and therefore provides [Link]
exact tracking performance without the reference model having to [3] Wang, J., Bierling, T., Höcht, L., Holzapfel, F., Klose, S., and Knoll, A.,
exactly model and match all plant states. Secondly, it provides pre- “Novel Dynamic Inversion Architecture Design for Quadrocopter
dictable and consistent error dynamics. Control,” Advances in Aerospace Guidance, Navigation and Control,
Springer, Berlin, 2011, pp. 261–272.
[Link]
Appendix: NDI and INDI Limit [4] Holzapfel, F., and Sachs, G., “Dynamic Inversion Based Control Con-
cept with Application to an Unmanned Aerial Vehicle,” AIAA Guidance,
Here it is shown how the error dynamics of NDI and INDI behave Navigation, and Control Conference and Exhibit, AIAA, Reston, VA,
in the limit of the actuator bandwidth going to infinity. Consider first 2004, pp. 1006–1018; also AIAA Paper 2004-4907, 2004.
the NDI law from Eq. (25) together with the system output dynamics [Link]
as given in Eq. (5), with the same assumptions as in Sec. III: [5] Lombaerts, T., Kaneshige, J., Schuet, S., Hardy, G., Aponso, B. L., and
1) All actuators have the same bandwidth, i.e., Ω  ωIk×k , where Shish, K. H., “Nonlinear Dynamic Inversion Based Attitude Control for
ω is a nonzero scalar. a Hovering Quad Tiltrotor eVTOL vehicle,” AIAA Scitech 2019 Forum,
AIAA, Reston, VA, 2019, pp. 1–26; also AIAA Paper 2019-0134, 2019.
2) Choose Ωy  ωIm×m . [Link]
By inserting Eq. (25) into the system dynamics given in Eq. (5), the [6] Khalil, H. K., Nonlinear Systems, 3rd ed., Prentice-Hall, Hoboken, NJ,
error dynamics can be deduced: 2002, pp. 505–544, Chap. 13.
[7] Bacon, B., and Ostroff, A., “Reconfigurable Flight Control Using Non-
y
r1  Fx x_  Fu Ik×k ω
uc − u linear Dynamic Inversion with a Special Accelerometer Implementa-
  Xr  tion,” Guidance, Navigation, and Control Conference and Exhibit,
1
y
r1  Fx x_  Fu Ik×k ω
Fu  y
r1  K e
i−1 y

i
− F x x
_ AIAA, Reston, VA, 2000, pp. 195–221; also AIAA Paper 2000-4565,
ω ref
i1
2000.
[Link]
 X  
r−1 [8] Steffensen, R., Steinert, A., and Holzapfel, F., “Longitudinal Incremen-

Fu  e
ry  Ki e
i
y u−u tal Reference Model for Fly-by-Wire Control Law Using Incremental
i0 Non-Linear Dynamic Inversion,” AIAA Scitech 2022 Forum, AIAA,
 Xr  Reston, VA, 2022, pp. 1–21; also AIAA Paper 2022-1230, 2022.
1
y
r1  Fx x_  ω y
r1  K e
i
i−1 y − Fx x
_ [Link]
ω ref
i1 [9] Surmann, D., Zrenner, M., and Myschik, S., “Flight Performance
 X
r−1  Evaluation of a Conceptual eVTOL System Using Nonlinear Simula-
tions,” AIAA Scitech 2022 Forum, AIAA, Reston, VA, 2022, pp. 1–13;
 ω e
ry  K e
i
i y also AIAA Paper 2022-2220, 2022.
i0
[Link]
 X
r   X
r−1  [10] Steinleitner, A., Frenzel, V., Pfeifle, O., Denzel, J., and Fichter, W.,
y
r1  y
r1
ref  K i−1 e
i
y  ω e
r
y  K i e
i
y “Automatic Take-Off and Landing of Tailwheel Aircraft with Incremen-
i1 i0 tal Nonlinear Dynamic Inversion,” AIAA Scitech 2022 Forum, AIAA,
X
r  X
r−1  Reston, VA, 2022, pp. 1–19; also AIAA Paper 2022-1228, 2022.
e
r1
y  Ki−1 e
i
r
y  ω ey  Ki e
i
y 0 (A1) [Link]
[11] Gabrys, A. C., Steffensen, R., de Angelis Cordeiro, R., Azinheira, J. R.,
i1 i0
Moutinho, A., and Holzapfel, F., “Integration of Phase Plane Flight
Envelope Protections in Cascaded Incremental Flight Control,” IFAC-
which is the desired error dynamics given in Eq. (13). PapersOnLine, Vol. 52, No. 12, 2019, pp. 429–435.
If the INDI law from Eq. (26) is inserted into Eq. (5), the following [Link]
can be derived: [12] Huang, Y., Pool, D. M., Stroosma, O., and Chu, Q., “Robust Incremental
Nonlinear Dynamic Inversion Controller of Hexapod Flight Simulator
  X
r−1   Motion System,” Advances in Aerospace Guidance, Navigation and
y
r1  Fx x_  Fu Ik×k ω
Fu † e
r
y  K e
i
i y  u − u Control, Springer International Publishing, Cham, Switzerland, 2018,
i0 pp. 87–99.
  X
r−1  [Link]
y
r1  Fx x_  Fu Ik×k ω
Fu † e
r
y  K e
i
i y
[13] Grondman, F., Looye, G., Kuchar, R. O., Chu, Q. P., and Van Kampen,
E.-J., “Design and Flight Testing of Incremental Nonlinear Dynamic
i0
  Inversion-Based Control Laws for a Passenger Aircraft,” Guidance,
X
r−1 Navigation, and Control Conference, AIAA, Reston, VA, 2018,
− y
r1  Fx x_  ω e
r
y  Ki e
i
y 0 (A2) pp. 1–25; also AIAA Paper 2018-0385, 2018.
i0 [Link]
 STEFFENSEN, STEINERT, AND SMEUR 717

[14] Li, M., Li, J., Tang, Y., and Sun, S., “An Extended INDI Approach and [21] Pfeifle, O., and Fichter, W., “Time-Optimal Incremental Nonlinear
Application to Pitch Rate Control Laws Design of an Aircraft,” AIAA Dynamic Inversion Through Deadbeat Control,” AIAA Scitech 2022
Aviation 2021 Forum, AIAA, Reston, VA, 2021, pp. 3005–3018; also Forum, AIAA, Reston, VA, 2022, pp. 1–10; also AIAA Paper 2022-
AIAA Paper 2021-3005, 2021. 1596, 2022.
[Link] [Link]
[15] Smeur, E. J. J., Chu, Q. P., and de Croon, G. C. H. E., “Adaptive [22] Raab, S. A., Zhang, J., Bhardwaj, P., and Holzapfel, F., “Consideration
Incremental Nonlinear Dynamic Inversion for Attitude Control of Micro of Control Effector Dynamics and Saturations in an Extended INDI
Aerial Vehicles,” Journal of Guidance, Control, and Dynamics, Vol. 39, Approach,” AIAA Aviation 2019 Forum, AIAA, Reston, VA, 2019,
No. 3, 2016, pp. 450–461. pp. 1–13; also AIAA Paper 2019-3267, 2019.
[Link] [Link]
[16] Wang, X., Van Kampen, E., Chu, Q. P., and De Breuker, R., “Flexible [23] Bhardwaj, P., Raab, S. A., and Holzapfel, F., “Higher Order Reference
Aircraft Gust Load Alleviation with Incremental Nonlinear Dynamic Model for Continuous Dynamic Inversion Control,” AIAA Scitech 2021
Inversion,” Journal of Guidance, Control, and Dynamics, Vol. 42, Forum, AIAA, Reston, VA, 2021, pp. 1–15; also AIAA Paper 2021-
No. 7, 2019, pp. 1519–1536. 1130, 2021.
[Link] [Link]
[17] Li, X., Jiang, Y., Zhang, J., Shi, S., and Zhao, L., “A Method to [24] Sieberling, S., Chu, Q., and Mulder, J., “Robust Flight Control Using
Compensate Interaction Between Actuator Dynamics and Control Allo- Incremental Nonlinear Dynamic Inversion and Angular Acceleration
cator Under Incremental Nonlinear Dynamic Inversion Controller,” IOP Prediction,” Journal of Guidance, Control, and Dynamics, Vol. 33,
Conference Series: Materials Science and Engineering, Vol. 428, No. 1, No. 6, 2010, pp. 1732–1742.
2018, pp. 1–8. [Link]
[Link] [25] Tal, E., and Karaman, S., “Accurate Tracking of Aggressive Quadrotor
[18] Zhou, Y., Ho, H. W., and Chu, Q., “Extended Incremental Non-
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/1.G007079

Trajectories Using Incremental Nonlinear Dynamic Inversion and Dif-

linear Dynamic Inversion for Optical Flow Control of Micro Air ferential Flatness,” Proceedings of the IEEE Conference on Decision
Vehicles,” Aerospace Science and Technology, Vol. 116, June 2021, and Control (CDC), Inst. of Electrical and Electronics Engineers,
Paper 106889. New York, 2019, pp. 4282–4288.
[Link] [Link]
[19] Cordeiro, R. A., Azinheira, R., and Moutinho, A., “Cascaded Incre- [26] Kim, C.-S., Ji, C.-H., Koh, G.-O., and Kim, B. S., “Stability Margin and
mental Backstepping Controller for the Attitude Tracking of Fixed- Structural Coupling Analysis of a Hybrid INDI Control for the Fighter
Wing Aircraft,” 5th CEAS Conference on Guidance, Navigation and Aircraft,” International Journal of Aeronautical and Space Sciences,
Control (EuroGNC), Council of European Aerospace Societies (CEAS) Vol. 22, No. 5, 2021, pp. 1154–1169.
Paper FrM21.4, 2019. [Link]
[20] Cordeiro, R. A., Marton, A. S., Azinheira, J. R., Carvalho, J. R., and [27] Kumtepe, Y., Pollack, T., and Van Kampen, E.-J., “Flight Control Law
Moutinho, A., “Increased Robustness to Delay in Incremental Control- Design using Hybrid Incremental Nonlinear Dynamic Inversion,” AIAA
lers Using Input Scaling Gain,” IEEE Transactions on Aerospace and Scitech 2022 Forum, AIAA, Reston, VA, 2022, pp. 1–18; also AIAA
Electronic Systems, Vol. 58, No. 2, 2021, pp. 1199–1210. Paper 2022-1597, 2022.
[Link] [Link]
 Appendix A: Article Reprints

A.2 Paper 2
Paper 2, reference [71], is reproduced from authors Rasmus Steffensen, Agnes Steinert and
Florian Holzapfel, Incremental control as an enhanced and robust implementation of gain
scheduled controllers avoiding hidden coupling terms. Aerospace Science and Technology;
141 (October 2023, 108500): Page 1-18 (all pages). Copyright © 2023, Elsevier Masson
SAS. All rights reserved.

XI
 Aerospace Science and Technology 141 (2023) 108500

Contents lists available at ScienceDirect

Aerospace Science and Technology

journal homepage: [Link]/locate/aescte

Incremental control as an enhanced and robust implementation

of gain scheduled controllers avoiding hidden coupling terms
Rasmus Steffensen ∗ , Agnes Steinert, Florian Holzapfel
Institute of Flight System Dynamics, Boltzmannstrasse 15, 85375, Garching bei München, Germany

a r t i c l e i n f o a b s t r a c t

Article history: This paper proposes a new and simple method to combine incremental control with the classical
Received 29 April 2023 proportional-integral (PI) control in order to increase the robustness properties of the resulting control
Received in revised form 23 June 2023 law. The proposed method avoids hidden coupling terms in a similar manner to the differential PI
Accepted 2 July 2023
controller implementation (DPI) and hence allows for fast gain scheduling. The integral action is
Available online 10 July 2023
accomplished by feedback of actuator measurements and therefore avoids integrator windup without
Communicated by Tsourdos Antonios
complicated anti-windup schemes. The control law obtains an increase in robust performance with a
Keywords: response less sensitive to variations in flight dynamic parameters when compared to the equivalent PI
Incremental flight control controller. Improvement in disturbance rejection is also obtained with a faster and more robust response.
Robust performance Increased robustness toward input time delay is observed by example. The trade-off between increased
Gain scheduling robustness and sensitivity to sensor noise and delay can be adjusted with an intuitive tuning parameter.
Hidden coupling terms The novel method, denoted incremental DPI (iDPI), is investigated in the context of fixed-wing flight
control. The concept is verified as a retrofit for an industry-level PI baseline controller for the longitudinal
motion of a fixed-wing aircraft using a high-fidelity simulation model.
© 2023 Elsevier Masson SAS. All rights reserved.

1. Introduction It has also been used in production aircraft, e.g. Lockheed-Martin

F-35 [16]. Since NDI is a nonlinear control law concept, it does
This paper investigates the possibility of using robust incre- not require gain scheduling as a PI control law. However, because
mental control techniques together with gain scheduled classical NDI requires accurate model knowledge, a variation denoted In-
controllers for use in flight control. A majority of closed-loop fly- cremental Nonlinear Dynamic Inversion (INDI) has been proposed
by-wire control laws for fixed-wing aircraft are based on some that reduces the amount of required model knowledge for con-
type of PI control law structure [1–6]. PI controllers are especially trol design. The concept is based on commanding increments to
popular because of their simplicity which is often desired in safety the control signal, rather than absolute values. INDI have become
critical applications. popular amongst many research groups in academia [17–33]. INDI
An implementation variant of PI controllers is the so-called relies on additional sensor measurements/filtering, and has been
Differential-Proportional-Integral controller (DPI) [7,8], not to be shown to be robust against uncertainties in [34–39]. Increased ro-
confused with the Proportional-Integral-Derivative (PID) controller. bustness has also been investigated with adding adaptive elements
The DPI was invented as a method to implement gain scheduled in [40–45]. Additional improvements have been made by consid-
PI controllers without introducing hidden coupling terms (HCT), eration of actuator dynamics in the control law design [46–49].
as these appear when directly implementing gain scheduling in However, if the actuator/surface deflection dynamics are relatively
PI controllers. DPI is used for inner-loop flight control in the Eu- slow compared to the system dynamics, the assumption that INDI
rofighter as described in [9–12] and allows to schedule controller approximates NDI breaks down [49]. If additionally the flight en-
gains over fast varying states such as angle of attack α . velope is large, then in practice INDI control laws often need to
Another promising approach for aircraft inner-loop control is introduce gain scheduling in the error controller to satisfy the de-
nonlinear dynamic inversion (NDI) as used in for example [13–15]. sign requirements over the entire flight envelope.

In order to combine the increased robustness properties of INDI,

* Corresponding author. the simplicity of PI, and the gain scheduling properties of DPI, this
E-mail address: [Link]@[Link] (R. Steffensen). paper introduces the iDPI. The implementation of iDPI is simple

[Link]
1270-9638/© 2023 Elsevier Masson SAS. All rights reserved.
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

and can be directly applied on existing PI control law architectures. Table 1

All the classical and well established analysis tools and methods Example aircraft parameters.
can be used for iDPI. The novel control method is established by Z α = −1.38 Zq = 0 Z η = − 0.1
taking derivatives of the measurements, which are compensated by M α = −4.588 M q = −2.0 M η = −4.0
an equivalent integration with the introduction of an incremental V 0 = 51.54 m/s

loop, where the actuator position is added to the control command

just before it is sent to the actuators. angle of attack dynamics of the longitudinal motion is given by Eq.
The following lists show the advantages and disadvantages of
(1):
the proposed approach. Italic is used to indicate advantages shared
between iDPI and DPI over PI. α̇ K =

( F z )K (1)
Advantages of iDPI over PI  E + q K − tan β K ( p K cos α K + r K sin α K )
1. No hidden coupling terms when scheduling the controller gains. m V KG cos β K
2. Bump less switching when different gain groups are selected.
where α K and β K is the kinematic angle of attack and side slip
3. Easy to avoid integrator windup.
angle, ( F z ) K is the forces in the z direction of the kinematic frame
4. Robust performance with respect to variations in, M α , M q , M η  E
and Z α , for example due to changes in mass, cg and aerody- K , m is the mass, V KG is the kinematic ( K ) speed ( V ) of the
namic characteristics. center of gravity (G ) with respect to the earth frame E. q K is the
5. Enhanced disturbance rejection, having more robust response kinematic pitch rate and r K the kinematic yaw rate. The rotational
with respect to parameter variations. dynamics considering only the longitudinal motion is given by:
1      
6. Increased robustness against time delays at the input.
q̇ K = M GA + M GP + I xz r K 2 − p K 2
Cost/drawbacks I yy B B
 (2)
1. Require measurements of u.
− ( I xx − I zz ) p K r K
2. More sensitive to sensor noise.
3. More sensitive to unknown measurement time delay. where I xx ,I y y ,I zz and I xz is the respective moment and product
of inertia, M GA B is the aerodynamic ( A ) pitching moment around
The drawbacks arise from the incremental loop combined with the
the differentiation of the measurement, and are similar to the  G center of gravity (G ), denoted in the body (B) frame, and
M P B is the propulsive ( P ) moment around the pitch axis. The
cost/drawbacks that one would find in INDI control laws.
linearization around zero flight path angle is given by:
      
The novel contributions of this paper are: α̇ K Zα 1 + Zq α K Zη
= + η
q̇ K Mα Mq q K Mη
1. Introduction of a novel robust PI control law structure based    (3)
on incremental control, namely iDPI. − Z α /V 0 − Zq wg
+
2. Development of an implementable version of iDPI using ap- −M α / V 0 −Mq qg
propriate synchronization filtering. and output equation:
3. Derivation showing that the implementation avoids hidden     
coupling terms and hence is a safe implementation when used α A 1 0 α K
=
with gain scheduling. n z − (V 0/g) Z α − (V 0/g) Zq q K
4. Demonstration of the enhanced robustness and stability char-  
0
acteristics of iDPI by examples. + η (4)
5. Verification as straight forward upgrade to an industry level PI − (V 0/g) Z η
  
controller for the longitudinal motion of the TUM/FSD DA42 −1/ V 0 0 wg
Research Platform using a high-fidelity 6-DoF nonlinear simu- +
Z α / g (V 0/ g) Zq qg
lation model.
where V 0 is the trim velocity, M α , M q , M η , Z α , Z q and Z η are
In Section 2 the plant, actuators and incremental loop is intro- the stability and control derivatives, see Table 1. η is the eleva-
duced, in Sections 3 and 4 the ideal and filtered iDPI is introduced. tor deflection, n z is the load factor, w g and q g is the vertical and
In Section 5 it is shown that iDPI does not lead to hidden coupling rotational gust components respectively. The delta () denotes de-
terms and in Section 6 the iDPI concept is verified using a DA42 viations from the trim state.
baseline controller. In Section 7 some general discussion remarks
on implementation and usability is given and at last in Section 8 2.2. Actuators
we sum up the results.
For the analysis, the actuators are assumed to be linear time
2. Setup invariant, strictly proper with unit DC-gain, i.e.

2.1. Plant U (s) bn−1 sn−1 + bn−2 sn−2 + ... + a0

GA = = , (5)
U C (s) sn + an−1 sn−1 + ... + a0
In this paper, we consider the plant dynamics given by the lon- where G A is the actuator transfer function, U C is the Laplace trans-
gitudinal equations of motion for a fixed-wing aircraft. For simplic- form of the actuator command u c , u is the actuator or effector po-
ity, this paper considers only the short time response, and hence sition with Laplace transform U , and bn , an are the actuator model
the analysis is restricted to the short period approximation, with parameters. As will become clear later, we have to assume that the
velocity and flight path angle fixed. This however does not restrict degree of the numerator is less than the degree of the denomina-
generalization of the results of the paper, and is verified in simula- tor, and a0 has to be equal for the numerator and denominator
tion on a full dynamic model in Section 6. The nonlinear kinematic of the transfer function. For lifter motors in electric vertical take

2
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

Fig. 1. Two equivalent block diagram of the ideal Incremental Loop.

Fig. 2. Two equivalent block diagram of the filtered Incremental Loop.

off and landing aircraft, the actuators are usually modeled as first
order actuators [47], for electro-mechanically actuated surfaces a
second order actuator is usually a good approximation [50,51] and
a0 ωa 2
G̃ A (s) = = , (11)
for hydraulic actuators a fourth order approximation is sometimes s + a1 s + 2ζa ωa
used [52]. Hence all the above mentioned actuator use cases are with ωa and ζa being the natural frequency and damping ratio of
included in the allowable actuator types. the second order actuator.

2.3. Incremental loop

An approximation of the incremental loop is often necessary in
practical implementations, due to synchronization with respect to
A core part of an incremental control law is the incremental
estimating derivatives of the measurements needed in the control
loop for the actuator. In this section, it is shown that introducing
an incremental loop with the actuators, will create a free integra- law [53]. In this paper, a first order filter is applied on the actu-
tor and some residual actuator dynamics, which are faster than ator feedback, such that the filtered incremental loop will result
the original actuator dynamics. The outer part of the control law in a zero with time constant τ f and an integrator as depicted in
will command increments in actuator commands u c . We split the Fig. 2, together with the resulting filtered residual actuator dynam-
analysis in 2 cases: ics. This structure will result in the ideal implementation when the
filter time constant τ f goes to 0, and is hence a suitable approx-
1. The ideal case. imation for implementation. For a first order actuator, the filtered
2. The filtered case. residual actuator dynamics are given by Eq. (12),
1 1
It turns out that the ideal case is often not implementable in re- G̃ A , f =   , (12)
ality, hence an approximation with filtering is introduced. In the τ f + τa τ f τa
s+1
τ f +τa
ideal case, the command to the actuators is constructed as:
where τa = 1/ωa = 1/a0 is the time constant of the first order ac-
u c = u c + u . (6) tuator. The filtered residual actuator dynamics is of the same order
as the original actuators, but faster, i.e. with a lower time constant
In this paper we consider the actuator dynamics for each actu-
than the original actuators. We denote the static gain of the fil-
ator independent of each other and the plant dynamics, i.e. the
tered incremental loop by K̃ a, f . In case of a second order actuator,
total transfer function matrix for the actuators G A (s), is a diag-
the filtered residual actuator becomes:
onal matrix with the transfer function of each actuator dynamic
at the diagonal. When closing the incremental loop as shown in K̃ a, f ω̃a2, f
Fig. 1, the resulting dynamics from u c to u for each diagonal ele- G̃ A , f = , (13)
s2 + 2ω̃a, f ζ̃a, f s + ω̃a2, f
ment can be written as a free integrator plus the residual actuator
dynamics G̃ A : with static gain K̃ a, f , eigenfrequency ω̃a, f and relative damping
U (s) 1 bn−1 sn−1 +bn−2 sn−2 +...+a0 ζ̃a, f given by:
U C (s) =  
s sn−1 +(an−1 −bn−1 )sn−2 + an−2 −bn−2 sn−3 ...+a1 −b1
. (7)
2
K̃ a, f = 2ζ ω ω+a τ ω 2
For a first order actuator as given in Eq. (8) a a f a
 1
2
ωa 2ζa ωa +τ f ωa 2
GA = , (8) ω̃a, f = τf
(14)
s + ωa  
τ f 2ζa ωa +1
the ideal residual actuator dynamics is just a gain, equal to the
ζ̃a, f = 2ω̃a, f τ f
.
actuator bandwidth, as given in Eq. (9). By inspection, it can be verified that the limit in Eq. (15) holds:

G̃ A (s) = a0 = ωa , (9) lim G̃ A , f = G̃ A , (15)

τ f →0
where ωa is the actuator bandwidth. In this paper we denote the
static gain of the incremental loop by K̃ a such that for the first and hence indicate that the filtered implementation is an appro-
order actuator, K̃ a = ωa . For a second order actuator as given in priate approximation of the ideal incremental loop.
Eq. (10)
3. Ideal iDPI and PI/DPI
ωa2
GA = , (10) In this section, the incremental Differential Proportional-Integral
s2 + 2ωa ζa s + ωa2
controller (iDPI) is introduced and compared with the Proportional-
the ideal residual actuator dynamics can be derived resulting in Eq. Integral (PI) and Differential Proportional-Integral (DPI) control
(11): laws. Fig. 3 shows the structure of the ideal DPI control law, and

3
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

Fig. 3. Block diagram of ideal DPI algorithm.

Fig. 4. Block diagram of ideal incremental DPI (iDPI).

Fig. 4 the structure of the ideal iDPI control law for an angle of at- Table 2
tack control system. The gain h f f is the feed-forward gain, kα i is Open and Closed-Loop poles.
the integrator gain, kα is the angle of attack feedback gain and kq Open-Loop Closed-Loop
is the pitch rate feedback gain. G P is the plant transfer function Actuator Pole pa = −1/τa = −62 pa,cl = −1/τa,cl = −55.98
matrix. Short Period Pole Pair ω S P = 2.71 ω S P ,d = 6
For a fixed gain set, the DPI is exactly equal to the PI controller, ζ S P = 0.623 ζ S P ,d = 0.7
hence for the gain design and Linear Time Invariant (LTI) analysis, Integrator Pole - p I = −1
the PI control law structure can be used and is denoted by PI/DPI.
For the analysis of the gain scheduled controller, the PI and DPI
will differ, as will be investigated in Section 5 and compared to For the investigation, pole placement is used since it allows
iDPI. to place the poles of both the PI/DPI and iDPI closed-loop sys-
In the iDPI closed-loop system, there is 1 pole less than in tem at the same locations. For the PI/DPI system, there are 4
the DPI closed-loop system. The reason is that G̃ A is one order poles (2 short period poles, 1 actuator pole and 1 integrator pole)
lower than G A because the integrator is borrowed from the actua- and 3 feedback gains, hence only the integrator and short-period
tors with the actuator incremental loop as shown in Eq. (7). This poles can be placed. The actuator pole will then float to a position
means that iDPI does not have a dedicated integrator inside the pa,cl = −1/τa,cl depending on the other pole locations. The feed-
control law. forward gain h f f = −kα i τ I is used to place a zero on top of the
closed-loop integrator pole p I = −1/τ I , to hide the integral behav-
3.1. Performance properties of ideal iDPI and comparison with PI/DPI
ior from the pilot as is often done in flight control. The resulting
To make a fair comparison, a PI/DPI controller with a feed- closed-loop transfer function is:
forward, designed to cancel the closed-loop integrator pole is con- (τ s+1)(τα s+1)ω S P ,d 2
G ααc , P I / D P I =   I  , (16)
sidered as a benchmark, see Fig. 3. In this case, it is possible to τa,cl s+1 (τ I s+1) s2 +2ζ S P ,d ω S P ,d s+ω S P ,d 2
design the iDPI and the DPI with exactly same closed-loop transfer
where ζ S P is the relative damping of the short period, ω S P is the
function from αc to α in nominal conditions. This makes it straight
eigenfrequency of the short period, τα is the time constant of the
forward to visualize the robustness of the response with variations
alpha dynamics zero, and the subscript d denotes their desired val-
in the model parameters of the plant.
ues. The parameters are given in Table 2.
For the iDPI there are only 3 poles (2 short period poles and
Consider the short-period motion of an aircraft as described by
the system in Eq. (3) with the output given by Eq. (4). For sim- 1 shared integrator-actuator pole). Since there are still 3 feedback
plicity, we consider the actuators to be first order with bandwidth gains, all 3 poles can be placed. The feed-forward is not necessary
ωa . In order to visualize the robustness properties with respect to in this example, as there is no dedicated controller integrator pole
variations in the individual model parameters, each parameter is in (the integrator is borrowed from the actuator) to cancel in the iDPI,
turn varied with a uniform distribution ±80% around the nominal hence h f f = 0. In the gain design, the actuator pole is placed at the
value. same location as the closed-loop actuator pole of the DPI, pa,cl =

4
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

Fig. 5. Performance comparison of PI/DPI with ideal iDPI. (For interpretation of the colors in the figure(s), the reader is referred to the web version of this article.)

−1/τa,cl , such that the closed-loop transfer functions of DPI and responses and the dashed red lines are the varied iDPI responses.
iDPI match when considering the pole-zero cancellation in Eq. (16). It is seen that iDPI features the following advantages when com-
The resulting closed-loop transfer function is: pared to PI/DPI:

(τα s + 1) ω S P ,d 2
G ααc ,i D P I =    (17) 1. Significantly increased time domain response robustness wrt.
τa,cl s + 1 s2 + 2ζ S P ,d ω S P ,d s + ω S P ,d 2 all investigated parameter variations.
In nominal conditions, the closed-loop response from αc to α is 2. Very high robustness in closed-loop short period pole location
exactly the same for PI/DPI and iDPI. However, with variations with respect to variations in M α , M q and M η .
in the plant parameters, significant different robustness proper- 3. Significantly faster disturbance response in nominal condi-
ties of the 2 control laws exist. Fig. 5 shows the variation of the tions, and less spread in response with respect to all parame-
pole/zero locations, complementary sensitivity bode plot, step re- ter variations.
sponse and disturbance response with individual variation of the
parameters M α and M q , M η , Z α , and elevator actuator bandwidth The disturbance response is shown in the last column of Fig. 5.
ωη . The green lines depict the nominal design response, i.e. the Even in nominal conditions, the iDPI disturbance response is faster
response without any parameter variations. In case the nominal than the DPI response. This is because the iDPI disturbance re-
response differs between iDPI and DPI, the iDPI response is de- sponse is one order lower than the DPI response, since the inte-
picted as green dashed lines. The blue lines are the varied PI/DPI grator pole is borrowed from the actuators as described in Eq. (7).

5
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

Table 3 This is a common problem in both DPI control laws [55,56], and
Comparison of nominal SISO Gain, Phase and Time Delay margins between PI (and incremental control laws [53,57]. The solution is to, in one way or
DPI/Filtered DPI), ideal iDPI and filtered iDPI.
the other, create a phase advance after the control signal is calcu-
Cut PI Ideal iDPI Filtered iDPI lated. In DPI it is solved by adding an additional zero to the inte-
uc ∞/69.8◦ /0.17s 36 dB/76.2◦ /0.21s ∞/77.2◦ /0.21s grator in the block before u c as shown in Fig. 7, and in INDI and
u c - ∞/85.5◦ /0.02s ∞/65.5◦ /0.04s iDPI it is solved by adding filtering on the u feedback loop, which
α ∞/70.7◦ /0.29s ∞/79.3◦ /0.12s ∞/62.7◦ /0.10s effectively adds the same zero on the forward path as shown in
q ∞/111◦ /0.23s ∞/102◦ /0.03s ∞/83.0◦ /0.05s the block after u c in Fig. 8.

Comparing the control structure of the filtered DPI and iDPI

This makes the settling time of the iDPI disturbance response sig-
in Fig. 7 and 8 show that the effective difference between the
nificantly lower. For all parameter variations, the spread of the iDPI
control laws is the modified actuator dynamics G̃ A , f , which are
response is lower and hence less sensitive to parameter variations.
significantly faster than the actuator dynamics G A due to the u
feedback. This is in the end what produces the increased robust-
3.2. Stability properties of ideal iDPI and comparison with PI/DPI
ness. It is very important to note that although G̃ A , f is faster than
the open-loop actuator dynamics G A , this does not mean that the
In this section, the stability properties of the iDPI and DPI/PI is closed-loop actuator pole is faster in iDPI compared to DPI, which
compared using classical SISO stability margins and graphical tools depends on the choice of feedback gains. It is often not desired
in terms of Nyquist and Nichols charts. to push the actuators faster than their design operation, and this
is not what iDPI in the end will do, as will be demonstrated in
In terms of flight control, stability margins usually have to the following example. It is seen that by proper choice of control
be satisfied according to the SAE94900 standard [54]. The stan- gains, the iDPI and PI actuator commands will be exactly the same
dard differentiates between aerodynamic-closed-loops (ACL) and in nominal conditions.
non-aerodynamic-closed-loops (N-ACL). Neglecting requirements
related to aero-structural modes and corner cases of airspeed en- It can often make sense to scale the incremental loop to have
velopes, the gain margin has to be more than 6 dB and phase unit DC gain, for example if other non-linear elements are added
margin more than 45◦ for ACL. In addition, with variation of crit- in the control law. This is done in Section 6 for the DA42 base-
ical stability derivatives, e.g. M α , M q , M η and Z α , by up to 20%,
line control law. The inverse of the incremental loop gain 1/ K̃ a, f
the required stability margins shall be at least half of the nominal
is multiplied before the incremental command u c as shown in
stability margin requirements.
Fig. 12. In this case, it is seen that with τ f going to infinity, the
incremental loop is slowed down, and iDPI turns exactly into a ba-
In incremental control laws, including iDPI, the incremental
sic PI. I.e., the control law given in Fig. 8 with the added scaling
loop does not rely on aerodynamics for loop closure, and hence
results in the following:
should be considered as a N-ACL. The loop closure at u c does rely
on aerodynamics for loop closure, and hence should be considered sτ f + 1 1
u =G̃ A , f
as an ACL. This means that both the nominal stability require- s K̃ a, f
ments, and the stability requirements with aerodynamic parameter 
variations have to be considered at the u c loop point, and only s 1
hff αc + kα i (αc − α ) (18)
nominal stability margins have to be considered at u c . τf s + 1 τf s + 1

s s
For the demonstration example, the SISO stability margins are + kα α + kq q .
τf s + 1 τf s + 1
given in Table 3. Both phase and time delay margins are better at
the input u c for the iDPI (and filtered iDPI) than for the PI/DPI. Noting that
The phase and time delay margins at the measurements are less 1
for the iDPI (and filtered iDPI) than for the PI/DPI. If the time lim G̃ A , f = G A , (19)
τ f →∞ K̃ a, f
delays of the measurements are known, this can be synchronized
using the technique presented in [53]. In case the time delays are it can be verified that:
unknown they cannot be synchronized. This indicates that iDPI is  
1
less robust against unknown time delays than PI/DPI, which is also lim u = G A h f f αc + kα i (αc − α ) + kα α + kq q , (20)
τ f →∞ s
known from INDI. Fig. 6 gives an overview of the graphical repre-
sentation of the stability properties under parameter variations. L which is equal to the basic PI. This gives an insight in the conse-
is the loop transfer function calculated at u c or u. In the first 2 quences of different choices of filter time constant τ f . Choosing a
columns, it is seen that the spread of the iDPI due to parameter high τ f recovers the basic PI, but reduces the robust performance
variations in both the Nichols and Nyquist charts are significantly properties of iPDI. By inspection of Fig. 6 and 10, it is seen on the
less in the critical regions, compared to the spread of the PI. The bode plot of the loop transfer function L at u c , that the crossover
SISO margins can be also visualized in the bode plot in the 3rd col- frequency, i.e. frequency for crossing of 0 dB, is lower for the PI
umn of Fig. 6, and the last column shows the sensitivity function than for the ideal iDPI. The filtered iDPI has a crossover frequency
at u c . in-between determined by the choice of τ f . It is additionally noted
that the roll-off of L at high frequency is higher for the filtered
4. Filtered DPI & iDPI for realizable implementation version of iDPI. Choosing a high τ f can be used for safety reasons
in the beginning of a flight test campaign if concerns exist about
The ideal iDPI can usually not be implemented in practice, be- noise amplification, structural couplings or un-modeled high fre-
cause the derivatives of the control law signals are not readily quency dynamics. Choosing a low τ f will recover the ideal iDPI
available. In order to approximate the derivatives, filtering can be behavior with the best robust performance properties. In the end
used to obtain the derivatives. Due to the phase delay introduced τ f will be chosen as a compromise between noise/mode amplifi-
by the filtering, proper synchronization of the signals is required. cation and the desired robust performance characteristics.

6
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

Fig. 6. Stability comparison of PI/DPI with ideal iDPI.

Fig. 7. Block diagram of filtered DPI algorithm.

7
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

Fig. 8. Block diagram of filtered incremental DPI (iDPI).

4.1. Properties of filtered iDPI and comparison with PI/filtered DPI 5. Hidden coupling terms and gain scheduling of iDPI

To make a fair comparison, the example from Section 3.1 is Hidden coupling terms are the additional terms that can occur
used again. Since the filters used in DPI exactly cancel for the lin- in the linearization of a closed-loop control system when compar-
ear analysis, there is no difference between the PI, ideal DPI and ing to the original design model. These can occur when the control
filtered DPI. However, for the filtered iDPI, there is a difference, be- gains of the control law are scheduled with some scheduling vari-
cause the effective actuators are now of the approximated/filtered able. Scheduling variables can be either dependent on the system
type denoted G̃ A , f instead of G̃ A in the ideal case. As shown in or come from external sources. For flight control, the control gains
Section 2.3, the filtered effective actuators are of the same order as are often scheduled with dynamic and static pressure. Since these
the original actuators G A , but with faster dynamics. Hence, in the are fairly slowly varying compared to the feedback signals of the
filtered iDPI, there will be as many poles as in the original system, control law, this usually does not cause any issues. However, if the
in this case 4 poles (2 short period poles, 1 actuator pole and 1 fil- control gains are scheduled over quickly varying states, e.g. angle
ter pole). In order to retain a fair comparison, the poles are placed of attack, the effects of hidden coupling terms cannot be neglected.
as in the PI closed-loop system, see Table 2. The extra filter pole Scheduling with fast varying states is often done in high perfor-
will be floating, ending up to the left (in the complex plane) of the mance aircraft.
actuator pole. To obtain a matching closed-loop response from αc In order for no hidden coupling terms to appear, it is sufficient
to α , the feed-forward zero is placed to cancel the filter pole by to check the linearization property as given in [7,8,56]:
choosing h f f = −kα i τ f ,cl , where τ f ,cl is the time constant of the
closed-loop filter pole. The filter parameter τ f should be chosen Property (Linearization property). The linearization property is satis-
considering the noise characteristics, as these determine possible fied if the linearization l at trim point θ of the gain scheduled controller
noise amplification due to the derivative action on the measure- l(C (θ)) coincides with the linear controller C l (θ) at every trim point θ .
ments. For this analysis, a choice of τ f = 1/40s was made as this
is similar to values often used in comparable INDI control laws.
To investigate the hidden coupling terms of the control law, the
method of evaluating the Linearization Property is used. The prop-
Fig. 9 shows the comparison of the robust performance be- erty is fulfilled if the linear design model of the controller is equal
tween the filtered iDPI and the PI/DPI/filtered DPI. Comparing with to the linearization of the gain scheduled controller. First, we as-
Fig. 5, it is seen that the filtering removes a bit of robustness com- semble the system equations of the gain scheduled filtered iDPI
pared to the ideal implementation, as seen by the slight increase controller. Then the gain scheduled controller is linearized. At last
in spread of the responses due to the parameter variations. Espe- the linearization is checked against the linear design model of the
cially the short period pole location varies significantly more with controller, and it is shown that they are equivalent.
variation in the elevator effectiveness M η .

The system equations of the gain scheduled filtered iDPI con-

The stability margins are given in Table 3, and show that the
troller can be deduced from the block diagram in Fig. 8. The states
phase margin and time delay margin are similar in magnitude to
of the controller xC are:
the ideal implementation. The loop transfer function at u c for iDPI
can be obtained using the loop transfer function calculated for the  
xC = u f , αc , f , αi , f , α f , q f , (21)
design PI control law as shown in Fig. 16, Appendix A.1. Fig. 10
shows the frequency characteristics of the loop transfer function L where subscript f indicate filter states, i.e. u f is state for the filter
at u c . It is seen that the spread due to parameter variations are on the u feedback in the incremental loop, αc , f is the filter state
still less than for the PI control law, but a bit more spread than for the angle of attack command αc , αi , f the state for the filter
in the ideal implementation. In Appendix A.2, Fig. 18 shows the before the integral control gain, α f is for the α measurement, and
graphical stability analysis for the actuator cut u c . It can be seen
q f the pitch rate. The inputs to the controller u C are given by:
that the nominal margins are larger at this cut, however, there is a
much larger spread with parameter variations. This is because the u C = (u , αc , α , q, σ s ) , (22)
incremental loop is also cut, and the integral effect of the actua-
tor feedback is lost, and hence makes the robustness analysis from where σ s are the scheduling variables, and the output of the con-
that perspective less relevant. Because the loop cut at u c is a N- troller y C is the control signal:
ACL with respect to certification [54], robustness with respect to
aerodynamic parameter variations is not to be considered at u c . y C = uc . (23)

8
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

Fig. 9. Performance comparison of PI/DPI with filtered iDPI.

 
The internal states of the controller are all the filter states, with The plant trim point is given by θ = u 0 , α0 , q0 , σ s,0 , which
the following filter differential equations: allow for steady turning, climbing, and steady push/pull ma-
  neuvers to be considered. u c = u 0 is a necessary condition for
u̇ f = ω f u − u f the actuators in Eq. (5) to be in trim. The remaining con-
 
α̇c, f = ω f αc − αc, f troller variables to be solved  for in the trimming procedure are:
  αc , u f , αc, f , αi, f , α f , q f . Except for the state αi, f , the remaining
α̇i, f = ω f αc − α − αi, f (24)
  filter trim states can all be found by setting the derivatives equal
α̇ f = ω f α − α f to 0 in Eq. (24) such that:
 
q̇ f = ω f q − q f ,
αc, f ,0 = αc,0
with ω f = 1/τ f and the output equation given by: u f ,0 = u 0
  (26)
u c =u f + h f f (σ s ) ω f αc − αc , f + kα i (σ s ) αi , f α f ,0 = α0
    (25)
+ kα (σ s ) ω f α − α f + kq (σ s ) ω f q − q f . q f = q0 .

9
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

Fig. 10. Stability comparison of PI/DPI with filtered iDPI.

For the output equation, u c = u 0 holds in trim and hence: Since the controller system equations given by Eq. (24) are al-
    ready linear, the only nonlinear element of the gain scheduled
u 0 =u f ,0 + h f f σ s,0 ω f αc,0 − αc, f ,0
      controller is the output equation (25). The linearized output equa-
+ kα i σ s,0 αi , f ,0 + kα σ s,0 ω f α0 − α f ,0 (27) tion is given by:
   
+ kq σ s,0 ω f q0 − q f ,0 .
∂ uc ∂ uc
This can be used with Eq. (26) to calculate the last filter state trim u c =   x C + u C , (30)
∂ xC xC ,0 ,u C ,0 ∂ uC  
value αi , f ,0 : xC ,0 ,u C ,0

 
0 = kα i σ s,0 αi, f ,0 ⇒ αi, f ,0 = 0. (28) with
Now using Eq. (28) and α̇i, f in Eq. (24), the last trim variable αc,0 ∂ uc  
can be calculated:
= 1 −h f f (σ s ) ω f kα i (σ s ) −kα (σ s ) ω f −kq (σ s ) ω f
∂ xC
   
0 = ωf αc,0 − α0 − αi, f ,0 ⇒ αc,0 = α0 . (29) ∂ uc
= 0 h f f (σ s ) ω f kα (σ s ) ω f kq (σ s ) ω f ∂∂σucs (31)
This concludes the trimming procedure. To check the linearization
∂ uC
property, the controller is now linearized around this trim point. u c = u c − u c,0 .

10
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

Note here that u c is the difference from the trim value, and not
the incremental control signal as used before. The last entry ∂∂σu c is
s
the important one, as this can cause the differences between the
linearized gain scheduled controller and the linear design model.
The entry is given by:
∂ uc ∂  
= h f f (σ s ) ω f αc − αc , f
∂σ s ∂σ s
∂ ∂  
+ kα i (σ s ) αi , f + kα (σ s ) ω f α − α f (32)
∂σ s ∂σ s
∂  
+ kq (σ s ) ω f q − q f .
∂σ s
By inspection, it is seen that the particular choice of controller
structure results in the scheduling term to evaluate to zero in the
trim point:

∂ uc
  = 0, (33)
∂σ s xC ,0 ,u C ,0

and hence does not contribute to the linearization. It is worth to

note that in case the incremental loop is broken, i.e. the u feedback
is removed, then this term is not zero and will produce hidden
coupling terms. Evaluating the Jacobian matrices in Eq. (31) in the
trim point leads to the linear output equation given by: Fig. 11. Non-linear simulation of gain scheduled PI, Filtered DPI and Filtered iDPI.
   
u c =h f f σ s,0 ω f αc − αc, f
      5.1. Simulation comparison demonstrating hidden coupling effects
+kα i σ s,0 αi , f + kα σ s,0 ω f α − α f (34)
    In this section, the effect of hidden coupling terms in the PI
+kq σ s,0 ω f q − q f . controller implementation is compared to the response of the fil-
Comparing this to Eq. (25) or the block diagram in Fig. 8 confirms tered DPI and iDPI implementation without hidden coupling terms.
that the linearized gain scheduled controller corresponds exactly Similar to the previous analysis, all three control laws have the
to the linear design model, and hence no hidden coupling terms exact same linear response from input αc to α in each envelope
exist for the gain scheduled filtered iDPI control law. point, however, it is seen that the actual response on the nonlin-
In contrast, the corresponding PI control law does have hidden ear system is significantly different.
coupling terms. The controller state is given by the integrator state For the investigation, the controllers are simulated on the non-
xC = e α i , controller input u C = (αc , α , q) and output y C = u c . The linear flight dynamics model given by Eq. (1) and Eq. (2), with
equation for the integral state is given by: the velocity fixed. The gain design is done using analytically de-
rived linear models on the form of Eq. (3) and Eq. (4). The control
ė α i = αc − α , (35) law gains are calculated analytically and are continuously adjusted
with online pole placement. The control objective is varied such
and the output equation is given by: that the desired short-period damping ζ S P ,d is scheduled to in-
crease from 0.7 to 1 as the angle of attack approaches the maxi-
u c = h f f (σ s ) αc + kα i (σ s ) e α i + kα (σ s ) α + kq (σ s ) q. (36) mum of 19◦ .
The plant trim is given by (u 0 , α0 , q0 , σ s,0 ). The controller Fig. 11 shows the step response for αc = 18.4◦ . It can be seen
trim point can be calculated to: u c ,0 = u 0 , αc ,0 = α0 and e α i ,0 = that the gain scheduled PI controller is very oscillatory due to the
 hidden coupling terms produced by the scheduling of the con-
u 0 − h f f (σ s,0 )αc ,0 − kα (σ s,0 )α0 − kq (σ s,0 )q0 /kα i . The linearized
gain scheduled PI control law in (36) can be written as: troller gains over the fast changing angle of attack α . Neither the
iDPI or DPI is oscillating as expected. It seems that the iDPI adapts
   
u c =h f f σ s,0 αc + kα i σ s,0 e α i faster to the changing desired response (damping of 1), and further
research will be conducted to see if this can be shown theoreti-
    ∂ uc
+ kα σ s,0 α + kq σ s,0 q +   σ s ,
cally.
∂σ s xC ,0 ,u C ,0

(37) 6. Application of iDPI to high-fidelity simulation of DA42

baseline control system
where
∂ uc ∂ h f f (σ s ) ∂ k (σ s ) In this section the iDPI concept is applied to an existing indus-
  = αc,0 + α i e α i ,0 try level PI control law that was flight tested on the Grob G520
∂σ s xC ,0 ,u C ,0 ∂σ s σ s,0 ∂σ s σ s,0 [58], Elias Electric Aircraft [59], the Dornier Do-228 [60] and Di-
(38)
∂ k α (σ s ) ∂ k q (σ s ) amond DA42 [61]. The control law is explained in detail in the
+ α0 + q0 , above mentioned papers, but for the investigation related to iDPI,
∂ σ s σ s,0 ∂ σ s σ s,0
it is important to note that the control law includes elements of:
which in general is not equal to 0. Comparing Eq. (37) with Eq.
(36) shows that the linearized gain scheduled controller differs 1. Notch filtering to attenuate aeroservoelastic couplings.
from the design model by the last term in Eq. (37) denoted the 2. Roll off filtering of measurements to attenuate measurement
hidden coupling term. noise.

11
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

Fig. 12. Structure of DA42 controller with iDPI. Orange are modifications due to iDPI.

3. Nonlinear compensations for trim rate, turn compensation and measurements u. However, in this investigation, the measure-
coordination, inertia couplings and control allocation. ments were processed directly without an observer.
8. Run gain design routine with modified actuator dynamics
The evaluation is performed on a DA42 flight dynamic model as G̃ A , f .
presented in [62]. The model is a full 6-DoF model with advanced
models for aerodynamics, propulsion, weight and balance, actua- 6.2. Simulation of DA42
tion, signal transmissions and sensor simulation including effects
of aircraft structural modes that propagate to load factor measure- In the following, the performance of PI and filtered iDPI is com-
ments. pared using simulations at one envelope point on the nonlinear
model of the DA42. Fig. 13 depicts the time domain responses con-
6.1. Implementation sidering 20% variation in the respective aerodynamic coefficients.
It is seen that the prediction from the simple example studies in
Fig. 12 shows the structure of the DA42 longitudinal baseline Section 3 and 4 also holds on a realistic industry level control law
control law with the adaptations required to implement the iDPI and simulation model. I.e. the variation of the performance output
structure marked with orange. The control variable is the load fac-  f D gz = −n z around the design response is much smaller for
tor n z , where n z,c is the incremental load factor command with the iDPI than for the PI considering all parameter variations.
respect to the trim value given by n z,trim = cosθ̂/cosφ̂ , where φ̂ The SISO stability margins are compared in Fig. 14, based on
and θ̂ is the estimate of the bank and pitch angle. kn z ,i and kn z is the linearization of the controller/plant design model, including all
the integral and proportional feedback gain respectively. N L in and subsystems and filters, and truncating V and θ from the model.
N L out contains all the nonlinear compensations mentioned in item Both the nominal limit and half of the nominal limit as discussed
(3) above. G no (s) is the notch filtering, and G ro (s) the roll off filter- in Section 3.2 are indicated by the red dashed lines. It is seen that
ing. V O M O contains all the command voting and monitoring, and all requirements are fulfilled for both PI and iDPI. The higher (in-
K F is a Kalman observer to process the surface deflection mea- finite) phase margin (and hence infinite time delay margin) at the
surements. input for iDPI was also verified in simulation by adding input time
In particular, the following modifications were required to the delays, which results in very poor performance, but the system
original controller: remains stable. The other margins are less for iDPI, which seems
reasonable due to the higher robust performance.
1. Add filters before controller gains. Add reset mechanism, reset
to input. 7. Discussion
2. Remove integrator from control law.
3. Add incremental loop. This section discusses several issues and design choices.
4. Add downscaling 1/ K̃ a, f to make incremental loop with unit
gain. The constant K̃ a, f is calculated using a second order ap- If the system is overactuated, there are several strategies to dis-
proximation of the actuator dynamics. tribute the controls. One possibility is to use a MIMO gain design
5. Add roll off filter to the u feedback in the incremental loop, to technique, such that the feedback gains appear as matrices that
synchronize with the roll off filter on q and n z . distribute the control action to the effectors. Another approach is
6. Move notch filter in original control law from N L out to inside a dedicated control allocation. In case of INDI, often a bottleneck
the incremental loop. strategy is used, where the control signals are collapsed into a
7. In case of noise and uncertainties on the measurement of u, it pseudo control, which is then distributed with a control allocation
would be reasonable to add an observer to process deflection algorithm to the appropriate effectors. A similar strategy can be

12
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

Fig. 13. Nonlinear simulation of DA42 baseline with PI and iDPI for 20% variation in model parameters.

Fig. 14. Evaluation of SISO stability margins considering a 20% variation in aerodynamic parameters. The green line represents the nominal margins, the box shows the 25th
to 75th percentiles, the black line the median, while the end markers indicate the range boundaries.

applied in iDPI, by rearranging the structure, as shown in Fig. 15. also allows for secondary control objectives in the null-space of
The feedback of q̇ˆ is moved from the PI control structure and re- the control allocation matrix.
sults in a change in the gain kq by 1. In this structure B † represents A key difference between PI and iDPI is the need for measure-
any pseudo-inverse that solves the control allocation problem. This ments of the effector position u. This means that an additional

13
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

Fig. 16. Calculation of filtered iDPI from PI design model.

8. Conclusion

In this paper a new control law structure denoted iDPI was

introduced, which combines incremental control as known from
incremental nonlinear dynamic inversion, and the concept of
differential-proportional integral control. The proposed method
was shown to have increased robustness properties in terms of
control performance and disturbance rejection considering varia-
Fig. 15. Re-arranged structure for including control allocation.
tions in flight dynamic model parameters. This is accomplished
by filtered derivatives of the measurements and by replacing the
strategy is needed in case that this measurement is incorrect or integral action in the control law with actuator position feed-
lost. In other incremental control laws this problem is often solved back to form an incremental loop that in turn introduces the
by adding an observer for the measurement of u together with ad- integral action. Since the integrator is removed from the control
ditional monitoring and recovery. law, integrator wind-up is avoided and hence removes the neces-
sity of complex anti-windup methods. The control law structure
was shown to avoid hidden coupling terms using the linearization
Another common strategy with INDI controllers is a model fol- property, which allows for gain scheduling over fast varying states
lowing approach. This can be combined with the methods of iDPI such as angle of attack or external variables such as stick deflec-
presented in this paper. Fig. 19 in Appendix A.3 shows a modifi- tion. The use of actuator positions and filtered derivatives of the
cation of the iDPI structure to include model following of a phys- measurements in turn leads to increased sensitivity to noise and
ically inspired reference model, with q̇re f , ṅ z,re f and the control unknown sensor delays. The desired increase in robustness versus
variable C V re f coming from the reference model. The physical ref- the increase in sensitivity to sensor noise/delay can be adjusted
erence model has been used both with NDI and INDI controllers with an intuitive tuning parameter. With a conservative choice of
[19,63,64]. In the following approach, the derivatives of the refer- this tuning parameter, the control law was shown to approach
ence model variables are used for feedback rather than the total the classical PI. The control law structure is of low complexity as
states. This avoids problems where discrepancies between refer- desired in safety-critical applications, and the similarity with the
ence model and plant parameters make the relative distance be- PI structure allows for easy retrofitting of existing controllers as
tween the states incorrect. Instead only the derivative is compared, was shown in simulation with an industry level PI control law
which is zero in steady state, and hence in steady state does not and demonstrated improvements on a practically relevant system.
fight against each other due to uncertainties. This problem was Further research will be conducted with multiple inputs, different
addressed in [19], where a blending strategy in the error con- control allocation structures, secondary control objectives and gain
troller was proposed to accommodate this. The approach proposed design methods.
in this paper with iDPI solves the same issue in a simpler and
more straight forward manner. The derivatives of the reference Funding sources
model have to be synchronized with the τ f filter to match the
estimated measurement derivatives. The integral behavior will be This research is co-funded by the European Union in the scope
accomplished by the feedback of the difference between the con- of INCEPTION project, which has received funding from the EU
trol variables C V re f and C V , and allows for blending of physical Horizon 2020 Research and Innovation Programme under grant
variables over the flight envelope. agreement No. 723515.

14
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

Fig. 17. Stability comparison of PI/DPI with iDPI, with cut at control command u c .

Declaration of competing interest calculated for the design PI control law as shown in Fig. 16. By us-
ing the following results, one can avoid calculating explicitly the
The authors declare that they have no known competing finan- loop transfer function for the implemented controller, and allows
cial interests or personal relationships that could have appeared to the designer to instead assemble the loop transfer function using
influence the work reported in this paper. only the design tools for a basic PI control law. This reduces the
time needed since existing PI control design tools can be used ex-
Data availability clusively.

The authors do not have permission to share data. A.2. Additional loop cut plots

Appendix A In this appendix, additional loop cut figures are provided for
reference in Fig. 17 and Fig. 18.
A.1. Calculation of loop transfer functions from PI design model
A.3. Combination with model following INDI
This section derives how to assemble the loop transfer func-
tions for the iDPI control law by using the loop transfer function See Fig. 19.

15
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

Fig. 18. Stability comparison of PI/DPI with filtered iDPI, with cut at control command u c .

Fig. 19. Combination of iDPI with concept of model following INDI.

16
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

References [28] C.-S. Kim, C.-H. Ji, G.-O. Koh, B.S. Kim, Stability margin and structural coupling
analysis of a hybrid INDI control for the fighter aircraft, Int. J. Aeronaut. Space
[1] F. Holzapfel, M. Heller, M. Weingartner, G. Sachs, O. Da Costa, Development of Sci. 22 (5) (2021) 1154–1169.
control laws for the simulation of a new transport aircraft, Proc. Inst. Mech. [29] T. Han, Q. Hu, H.-S. Shin, A. Tsourdos, M. Xin, Sensor-based robust incremental
Eng., G J. Aerosp. Eng. 223 (2) (2009) 141–156. three-dimensional guidance law with terminal angle constraint, J. Guid. Control
[2] S. Delannoy, S. Oudin, Longitudinal control law for modern long range civil Dyn. 44 (11) (2021) 2016–2030.
aircrafts, Proceedings of the EuroGNC. [30] Y. Song, S. Liang, E. Niu, B. Xu, A perched landing control method based on
[3] F.R. Caldeira, M.V. Campos, R.L. Andrioli Junior, W.d.O. Carvalho, D. Gangsaas, E. incremental nonlinear dynamic inverse, in: 2022 4th International Conference
Camelier, D. Siqueira, L. Rubiano, Flight control system mode and method pro- on Control and Robotics (ICCR), IEEE, 2022, pp. 82–88.
viding aircraft speed control through the usage of momentary on-off control, [31] E. Tal, S. Karaman, Global incremental flight control for agile maneuvering of a
US Patent US20130138274A1, assignee: Embraer SA (05 2013). tailsitter flying wing, J. Guid. Control Dyn. 45 (12) (2022) 2332–2349.
[4] K. Najmabadi, M.R. Evans, E.E. Coleman, R.J. Bleeg, R.S. Breuhaus, D.M. An- [32] S. Zhang, H. Zhang, K. Ji, Incremental nonlinear dynamic inversion attitude con-
derson, T.A. Nelson, Aircraft pitch-axis stability and command augmentation trol for helicopter with actuator delay and saturation, Aerospace 10 (6) (2023)
system, US Patent US6325333B1, assignee: Boeing Co (12 2001). 521.
[5] T. Berger, M. Tischler, S.G. Hagerott, D. Gangsaas, N. Saeed, Longitudinal control [33] D. Surmann, S. Myschik, Gain design of an INDI-based controller for a con-
law design and handling qualities optimization for a business jet flight control ceptual eVTOL in a nonlinear simulation environment, in: AIAA SCITECH 2023
system, in: AIAA Atmospheric Flight Mechanics Conference, 2012, p. 4503. Forum, 2023, p. 1250.
[6] S. Oudin, C. Frot, Optimizing take-off performances using a protected rotation [34] T. Pollack, E.-J. Van Kampen, Robust stability and performance analysis of in-
law, in: AIAA SCITECH 2022 Forum, 2022, p. 0619. cremental dynamic inversion-based flight control laws, in: AIAA SCITECH 2022
[7] I. Kaminer, A.M. Pascoal, P.P. Khargonekar, E.E. Coleman, A velocity algorithm Forum, 2022, p. 1395.
for the implementation of gain-scheduled controllers, Automatica 31 (8) (1995) [35] S. Sieberling, Q. Chu, J. Mulder, Robust flight control using incremental non-
1185–1191. linear dynamic inversion and angular acceleration prediction, J. Guid. Control
[8] W.J. Rugh, J.S. Shamma, Research on gain scheduling, Automatica 36 (10) Dyn. 33 (6) (2010) 1732–1742.
(2000) 1401–1425. [36] R.A. Cordeiro, J.R. Azinheira, A. Moutinho, Robustness of incremental backstep-
[9] R. Osterhuber, M. Hanel, R. Hammon, Realization of the eurofighter 2000 pri- ping flight controllers: the boeing 747 case study, IEEE Trans. Aerosp. Electron.
mary lateral/directional flight control laws with differential pi-algorithm, in: Syst. 57 (5) (2021) 3492–3505.
AIAA Guidance, Navigation, and Control Conference and Exhibit, 2004, p. 4751. [37] R.A. Cordeiro, A.S. Marton, J.R. Azinheira, J.R. Carvalho, A. Moutinho, Increased
[10] N. Moritz, R. Osterhuber, Three-stage gradient-based optimization scheme in robustness to delay in incremental controllers using input scaling gain, IEEE
design of feedback gains within eurofighter primary control laws, in: AIAA Trans. Aerosp. Electron. Syst. 58 (2) (2021) 1199–1210.
Guidance, Navigation, and Control Conference and Exhibit, 2006, p. 6311. [38] X. Wang, E.-J. Van Kampen, Q. Chu, P. Lu, Stability analysis for incremen-
[11] H.-C. Oelker, R. Osterhuber, M. Hanel, Experiences with eurofighter handling tal nonlinear dynamic inversion control, J. Guid. Control Dyn. 42 (5) (2019)
qualities testing, in: AIAA Atmospheric Flight Mechanics Conference, 2009, 1116–1129.
p. 6321. [39] H. Xin, Q. Chen, P. Wang, Y. Wang, Z. Hou, The control performance analysis
[12] F. Hellmundt, A. Wildschek, R. Maier, R. Osterhuber, F. Holzapfel, Comparison of the incremental nonlinear dynamic inverse method and flight test, in: Pro-
of l1 adaptive augmentation strategies for a differential pi baseline controller ceedings of 2022 International Conference on Autonomous Unmanned Systems
on a longitudinal f16 aircraft model, in: Advances in Aerospace Guidance, Nav- (ICAUS 2022), Springer, 2023, pp. 3609–3619.
igation and Control: Selected Papers of the Third CEAS Specialist Conference on [40] V.S. Akkinapalli, G.P. Falconí, F. Holzapfel, Fault tolerant incremental attitude
Guidance, Navigation and Control Held in Toulouse, Springer, 2015, pp. 99–118. control using online parameter estimation for a multicopter system, in: 2017
[13] P. Smith, A. Berry, Flight test experience of a non-linear dynamic inversion con- 25th Mediterranean Conference on Control and Automation (MED), IEEE, 2017,
trol law on the VAAC harrier, in: Atmospheric Flight Mechanics Conference, pp. 454–460.
2000, p. 3914. [41] D.I. Ignatyev, H.-S. Shin, A. Tsourdos, Sparse online Gaussian process adaptation
[14] T. Lombaerts, G. Looye, Design and flight testing of nonlinear autoflight control for incremental backstepping flight control, Aerosp. Sci. Technol. 136 (2023)
laws incorporating direct lift control, in: Advances in Aerospace Guidance, Nav- 108157.
igation and Control: Selected Papers of the Second CEAS Specialist Conference [42] D.I. Ignatyev, H.-S. Shin, A. Tsourdos, Two-layer adaptive augmentation for in-
on Guidance, Navigation and Control, Springer, 2013, pp. 549–568. cremental backstepping flight control of transport aircraft in uncertain condi-
[15] F. Holzapfel, Nichtlineare adaptive regelung eines unbemannten fluggerätes, tions, Aerosp. Sci. Technol. 105 (2020) 106051.
Ph.D. thesis, Technische Universität München, Universitätsbibliothek, 2004. [43] P. Bhardwaj, V.S. Akkinapalli, J. Zhang, S. Saboo, F. Holzapfel, Adaptive augmen-
[16] J.J. Harris, F-35 flight control law design, development and verification, in: 2018 tation of incremental nonlinear dynamic inversion controller for an extended
Aviation Technology, Integration, and Operations Conference, 2018, p. 3516. f-16 model, in: AIAA Scitech 2019 Forum, 2019, p. 1923.
[17] E.J. Smeur, M. Bronz, G.C. de Croon, Incremental control and guidance of hybrid [44] H. Li, S. Myschik, F. Holzapfel, Null-space-excitation-based adaptive control for
aircraft applied to a tailsitter unmanned air vehicle, J. Guid. Control Dyn. 43 (2) an overactuated hexacopter model, J. Guid. Control Dyn. 46 (3) (2023) 483–498.
(2020) 274–287. [45] J. Chang, R. De Breuker, X. Wang, Adaptive nonlinear incremental flight control
[18] J.R. Azinheira, A. Moutinho, J. Carvalho, Lateral control of airship with uncertain for systems with unknown control effectiveness, IEEE Trans. Aerosp. Electron.
dynamics using incremental nonlinear dynamics inversion, IFAC-PapersOnLine Syst. 59 (1) (2022) 228–240.
48 (19) (2015) 69–74. [46] E.J. Smeur, Q. Chu, G.C. De Croon, Adaptive incremental nonlinear dynamic in-
[19] R. Steffensen, A. Steinert, F. Holzapfel, Longitudinal incremental reference version for attitude control of micro air vehicles, J. Guid. Control Dyn. 39 (3)
model for fly-by-wire control law using incremental non-linear dynamic in- (2016) 450–461.
version, in: AIAA Scitech 2022 Forum, 2022, p. 1230. [47] S.A. Raab, J. Zhang, P. Bhardwaj, F. Holzapfel, Consideration of control effec-
[20] F. Grondman, G. Looye, R.O. Kuchar, Q.P. Chu, E.-J. Van Kampen, Design and tor dynamics and saturations in an extended INDI approach, in: AIAA Aviation
flight testing of incremental nonlinear dynamic inversion-based control laws 2019 Forum, 2019, p. 3267.
for a passenger aircraft, in: 2018 AIAA Guidance, Navigation, and Control Con- [48] X. Li, Y. Jiang, J. Zhang, S. Shi, L. Zhao, A Method to Compensate Interaction
ference, 2018, p. 0385. Between Actuator Dynamics and Control Allocator Under Incremental Nonlin-
[21] X. Wang, E. Van Kampen, Q. Chu, R. De Breuker, Flexible aircraft gust load ear Dynamic Inversion Controller, IOP Conference Series: Materials Science and
alleviation with incremental nonlinear dynamic inversion, J. Guid. Control Dyn. Engineering, vol. 428, IOP Publishing, 2018, p. 012048.
42 (7) (2019) 1519–1536. [49] R. Steffensen, A. Steinert, E.J. Smeur, Nonlinear dynamic inversion with actuator
[22] S.A. Raab, J. Zhang, P. Bhardwaj, F. Holzapfel, Proposal of a unified control strat- dynamics: an incremental control perspective, J. Guid. Control Dyn. (2022) 1–9.
egy for vertical take-off and landing transition aircraft configurations, in: 2018 [50] S.P. Scherer, M. Speckmaier, D. Gierszewski, A. Steinert, F. Holzapfel, Compen-
Applied Aerodynamics Conference, 2018, p. 3478. sation of nonlinear transmission effects in electro-mechanical flight control
[23] O. Pfeifle, W. Fichter, Cascaded incremental nonlinear dynamic inversion for systems, in: AIAA SCITECH 2022 Forum, 2022, p. 0621.
three-dimensional spline-tracking with wind compensation, J. Guid. Control [51] D. Teubl, S. Oberschwendtner, M. Hornung, Measurement results of the ACMU
Dyn. 44 (8) (2021) 1559–1571. system in various research UAVs, in: AIAA SCITECH 2023 Forum, 2023, p. 0481.
[24] Z. Liu, Y. Zhang, J. Liang, H. Chen, Application of the improved incremental non- [52] S. Delannoy, Airbus generic control laws, Proceedings, EuroGNC.
linear dynamic inversion in fixed-wing UAV flight tests, J. Aerosp. Eng. 35 (6) [53] R. Steffensen, A. Steinert, Z. Mbikayi, S. Raab, J. Angelov, F. Holzapfel, Filter and
(2022) 04022091. sensor delay synchronization in incremental flight control laws, Aerospace Syst.
[25] D. Milz, G. Looye, Tilt-wing control design for a unified control concept, in: (2023) 1–20.
AIAA Scitech 2022 Forum, 2022, p. 1084. [54] S. AS94900, Aerospace-flight control systems-design, installation and test of
[26] L. Panish, C. Nicholls, M. Bacic, Nonlinear dynamic inversion flight control of a piloted military aircraft, general specification for: aerospace standard, SAE In-
tiltwing VTOL aircraft, in: AIAA SCITECH 2023 Forum, 2023, p. 1910. ternational, Jul.
[27] G. Di Francesco, M. Mattei, Modeling and incremental nonlinear dynamic in- [55] F. Hellmundt, Adaptive flight control for fault tolerance, Ph.D. thesis, Technische
version control of a novel unmanned tiltrotor, J. Aircr. 53 (2016) 73–86. Universität München, 2017.

17
R. Steffensen, A. Steinert and F. Holzapfel Aerospace Science and Technology 141 (2023) 108500

[56] H. Lhachemi, D. Saussié, G. Zhu, An enhanced velocity-based algorithm for safe [61] E. Karlsson, S.P. Schatz, T. Baier, C. Dörhöfer, A. Gabrys, M. Hochstrasser, C.
implementations of gain-scheduled controllers, Int. J. Control 90 (9) (2017) Krause, P.J. Lauffs, N.C. Mumm, K. Nürnberger, et al., Development of an au-
1973–1989. tomatic flight path controller for a da42 general aviation aircraft, in: Advances
[57] E.J. Smeur, G.C. de Croon, Q. Chu, Cascaded incremental nonlinear dynamic in- in Aerospace Guidance, Navigation and Control: Selected Papers of the Fourth
version for MAV disturbance rejection, Control Eng. Pract. 73 (2018) 79–90. CEAS Specialist Conference on Guidance, Navigation and Control Held in War-
[58] A. Steinert, R. Steffensen, D. Gierszewski, M. Speckmaier, F. Holzapfel, R. saw, Poland, April 2017, Springer, 2018, pp. 121–139.
Schmoldt, F. Demmler, U. Schell, M. Ornigg, M. Koop, Experimental results of [62] A.W. Zollitsch, S.P. Schatz, N.C. Mumm, F. Holzapfel, Model-in-the-loop sim-
flight test based gain tuning, in: AIAA SCITECH 2022 Forum, 2022, p. 2296. ulation of experimental flight control software, in: 2018 AIAA Modeling and
[59] S. Scherer, M. Speckmaier, D. Gierszewski, C. Mishra, A. Steinert, R. Steffensen, Simulation Technologies Conference, 2018, p. 0425.
S. Wulf, F. Holzapfel, Automatic take-off and landing of a very light all electric [63] F. Peter, M. Leitão, F. Holzapfel, Adaptive augmentation of a new baseline con-
optionally piloted aircraft. trol architecture for tail-controlled missiles using a nonlinear reference model,
[60] A. Steinert, F. Holzapfel, R. Steffensen, C. Merkl, Flight test based gain tuning in: AIAA Guidance, Navigation, and Control Conference, 2012, p. 5037.
using non-parametric frequency domain methods, in: AIAA Scitech 2021 Fo- [64] F. Zhang, F. Holzapfel, Flight control using physical dynamic inversion, in: AIAA
rum, 2021, p. 1424. Guidance, Navigation, and Control Conference, 2015, p. 1758.

18
A.3 Paper 3

A.3 Paper 3
Paper 3, reference [74], is re-printed under the Creative Commons Attribution (CC BY)
International License, [Link] with no
changes from the original.

XXX
 actuators
Article
Practical System Identification and Incremental Control Design
for a Subscale Fixed-Wing Aircraft
Rasmus Steffensen * , Kilian Ginnell and Florian Holzapfel

Institute of Flight System Dynamics, Technical University of Munich, 85748 Garching near Munich, Germany;
[Link]@[Link] (K.G.); [Link]@[Link] (F.H.)
* Correspondence: [Link]@[Link]

Abstract: An incremental differential proportional integral (iDPI) control law using eigenstructure
assignment gain design is tested in flight on a subscale platform to validate its suitability for fixed-
wing flight control. A kinematic relation for the aerodynamic side-slip angle rate is developed to
apply a pseudo full state feedback. In order to perform the gain design and assessment, a plant
model is estimated using flight test data from gyro, accelerometer, airspeed and surface deflection
measurements during sine-sweep excitations. Transfer function models for the actuators and surface
deflections are identified both in-flight and on the ground for several different actuators and control
surfaces using hall sensor surface deflection measurements. The analysis reveals a large variation
in bandwidth between the different types of servo motors. Flight test results are presented which
demonstrates that the plant model estimates based on tests with good frequency excitation, high
bandwidth actuators and surface deflection measurements can be used to reasonably predict the
closed-loop dynamic behavior of the aircraft. The closed-loop flight test results of the iDPi control
law show good performance and lays the groundwork for further development.

Keywords: incremental flight control; system identification; flight test results

1. Introduction
This paper investigates through flight test the feasibility of using an incremental
Citation: Steffensen, R.; Ginnell, K.;
differential proportional integral (iDPI) control law as introduced in [1] for controlling the
Holzapfel, F. Practical System
Identification and Incremental Control
lateral-directional motion of a fixed-wing aircraft. The investigation covers identification
Design for a Subscale Fixed-Wing
of servo-motor dynamics, system identification of a plant model, setup of the control law
Aircraft. Actuators 2024, 13, 130.
structure, gain design, assessment of stability and robustness, and at last flight test results
[Link] of the control law. The iDPI control concept was developed from the method of incremental
nonlinear dynamic inversion (INDI) [2,3]. The iDPI control law is an implementation of
Academic Editor: André Preumont
a gain scheduled proportional-integral (PI) controller that avoids hidden coupling terms,
Received: 3 February 2024 and inherits the robust performance of INDI against model uncertainties. Avoiding hidden
Revised: 20 March 2024 coupling terms enables the design and assessment of the controller using linear control
Accepted: 27 March 2024 theory, even when scheduling the controller gains makes the control law non-linear. This
Published: 4 April 2024 paper forms the starting point of the investigation of its usefulness for practical flight control.
It provides validation of its usability in a real-world setting and lays the groundwork for
further investigation.
System identification for aircraft is a wide topic, and many branches of the field are
Copyright: © 2024 by the authors. discussed in [4–6]. Recent advances in aircraft system identification include [7–25]. Several
Licensee MDPI, Basel, Switzerland.
methods above use frequency domain identification, which was also studied in [26–29].
This article is an open access article
Also, online identification has been used in e.g., [30–33].
distributed under the terms and
In this paper, surface deflection measurements was used to both identify in-flight
conditions of the Creative Commons
actuator dynamics and for identifying a flight dynamics model. The system consists of a
Attribution (CC BY) license (https://
hall sensor device near the surface hingeline, as developed and tested in [34–36]. In contrast,
[Link]/licenses/by/
4.0/).
internal actuator measurements were used in [37,38].

Actuators 2024, 13, 130. [Link] [Link]

Actuators 2024, 13, 130 2 of 33

Fixed wing flight control is also a wide field with many different approaches and
objectives; an excerpt is discussed in [39–51]. The control method of this paper is based on
incremental methods, which are an active field of research. Incremental control has been
proposed to reduce the dependency of model information for control law design and has
been studied recently in [52–85]. In this paper, the method of eigenstructure assignment is
additionally used for gain design, a method that has been investigated in [86–88].
Flight testing of Unmanned Aerial Vehicles (UAVs) has increased with advancements
in microcontroller hardware and sensors, and the ease of integration with tools such as
MATLAB/Simulink. Recent UAV flight test campaigns include [89–99], which also include
flight testing with incremental type control laws in [100–104]. Subscale flight testing and
system identification was performed in [105–109]. In [110], both system identification
and control law design were performed for the longitudinal motion of a fixed wing UAV.
Complete autopilot systems for UAV have been built or described in, for example, [111–116]
including the PX4 flight stack used in this study.
The goal of this paper is to validate the feasibility of iDPI in a practical setting. In ad-
dition, the goal is to obtain an actuator and plant model of a small fixed wing UAV of
good enough quality to design and predict the closed-loop behavior with the controller.
The contributions of this paper are:
• Frequency domain identification of servo dynamics of different actuators both in flight
and on ground using surface deflection measurements.
• Presentation of experimental results for system identification of a subscale fixed-wing
aircraft using surface deflection measurements.
• Verification that the model resulting from system identification can accurately predict
the closed-loop control law performance.
• Verification that iDPI control laws can obtain good performance on a subscale aircraft.
• Application of eigenstructure assignment for an iDPI lateral/directional control law.
• Derivation of kinematic relations for flow angle rates to avoid derivatives of load
factor or flow angle measurements applicable to differential-integral control laws.

2. Experimental Setup
The aircraft, denoted “The Albadrone”, used for flight testing is the X-UAV Mini Talon
shown if Figure 1, with modifications based on [117]. The basic data are given in Table 1.
The aircraft has a V-tail configuration with pusher propeller. Pitot tubes are mounted
one at each wingtip, with only the left one active for the following experiments. For
accurate airspeed measurements, wingtip mounted pitot tubes are not ideal, but were
set up as such to support another research project investigating the ability to estimate
the wind gradient in banked flight. The flight control computer is a Pixhawk 4, with the
PX4 flight stack [118]. The control laws are implemented with MATLAB/Simulink [119],
and flashed onto the aircraft using the UAV Toolbox Support Package for PX4 Autopilots.
The sensor setup is depicted in Figure 2. The inertial measurement unit (IMU) which
includes gyros and accelerometers is placed almost in the center of gravity. The surface
deflection measurements are sent with I 2 C to the Pixhawk, and a custom-made PX4 driver
is used to relay the measurements to the Simulink Application. Two I 2 C buses are used to
connect the sensors in order to have the required bandwidth and to avoid address clashing.
The I 2 C packets are decoded in a custom-made PX4 application, which forwards the data
to the PX4 Simulink Application using a uORB message. The logging is performed using
a custom-made Simulink Bus packing block that allows for simple logging of all relevant
internal signals stored in a Simulink Bus object. This allows for rapid changes of logged
signals with minimum maintenance of the Simulink model. The signals with a maximum
of 498 bytes are logged with 100 Hz, and saved to an onboard SD card. All maneuvers
for system identification and controller validation are injected automatically via the flight
control software , by using execution and mode switches on the remote controller. Two
maneuver injection systems were used, a custom made system for rapid development,
Actuators 2024, 13, 130 3 of 33

and a well tested system also used for manned aircraft [120]. Both systems automatically
index the maneuvers, to ease the task of post processing and data analysis.

Table 1. Albadrone basic data.

Wingspan 1.3 m
Length 0.83 m
Weight 1–2 kg
Material Polystyrene Foam

Figure 1. Albadrone subscale aircraft.

Flight Control Computer

Pitot Tube
Servomotor
Drive Shaft Position Sensor

GPS & Compass Outer Position Sensor

Inner Position Sensor

Figure 2. Albadrone sensor setup.

3. Identification of Servo+Surface Dynamics and Equivalent Time Delay

This section identifies a simplified servo+surface deflection dynamic model for ground
and in-air operation at a fixed velocity. Different sets of actuators and linkages were tested
and the analysis will uncover the different dynamic behavior. The following list is the
objectives of the analysis:
• Identify servo+surface transfer function models.
• Compare in-flight versus ground performance.
• Compare servo type, servo-surface linkage type and surface type.
In order to obtain a simple model for control design, a second-order plus time delay
model is chosen:
Actuators 2024, 13, 130 4 of 33

2
K A ω0,A
G A (s) = 2
e−Teq s (1)
s2 + 2ω0,A ζ A s + ω0,A
where G A denotes the actuator transfer function, ω0,A is the actuator eigenfrequency, ζ A
is the relative damping, K A is the actuator static gain and Teq is the equivalent time delay.
A second-order response was also observed in [35] for a similar class of servos.

3.1. Servo Motors

Several actuators were tested during the flight test campaign. Table 2 gives an overview
of the actuators and their primary datasheet values.

Table 2. Datasheet values for the tested servos.

Torque (kgf·cm) Speed (s/60◦ , ◦ /s) Voltage (V)

MG90S 1.8 0.1, 600 4.8
M5251H 3.3 0.04, 1500 7.4
M5252H 4.7 0.05, 1200 7.4
MS320 5.5 0.08, 750 7.4

3.2. Excitation
A set of logarithmic sine-sweep maneuvers, as described in [4] (ch. 9), with length
T = 12 s, amplitude A = 10 deg, starting frequency of f 0 = 0.5 Hz (3.1 rad/s) and end
frequency of f 1 = 18 Hz (113 rad/s) was used. The frequencies were chosen to cover the
expected range of operation and capability of the actuators. An example of a time domain
plot is found in Figure 6.

3.3. Coherence
The actuators will have non-linear effects such as rate and acceleration saturations,
dead zone and backlash effects. The magnitude squared coherence function, denoted γ2xy ,
can be used to indicate the degree of linearity as a function of frequency, and can be used
to identify at which frequencies non-linearities have an effect. The control law should be
designed such that there is good separation between the crossover frequencies and the
frequency where the coherence starts dropping. All frequency plots for both actuators and
system dynamics include a measure of coherence, using the MATLAB mscohere() function
with default parameters.

3.4. Model Estimation

A continuous time transfer function model is estimated using the method presented
in [5] [ch. 11], by solving an optimization problem reducing the error between the model
and a set of measured frequency responses. The cost function, which will also be used for
estimation of a plant model, is given by [5]:
" #
1 nTF 20 nω
ω 
2
2 
n TF l∑ nω ω∑
J= Wγ Wg Ĝl (iω ) − | Gl (iω )| + Wp ∠Ĝl (iω ) − ∠Gl (iω ) , (2)
=1 = ω1

where n TF is the number of transfer function matrix entries to estimate (in cases of MIMO
systems) and ω1 to ωnω are the nω frequency points, chosen to be spaced with (almost)
equal distance on a logarithmic scale, such that the model match will be distributed equally
on a standard bode plot. Ĝl is one of the estimated frequency response pairs l (e.g., aileron
to roll rate) from flight test data, and Gl is the evaluation of the model at that frequency.
The weighting in the cost function is chosen based on recommendations from [5] as follows:
Actuators 2024, 13, 130 5 of 33

Wg = 1 and Wp = 0.01745, which will set 1 dB magnitude error equal to 7.57 deg phase
error. The coherence weighting recommended in [5] is used and is given by
  2
2
Wγ = 1.58 1 − e−γxy , (3)

which reduces the weighting of the data points as the coherence decreases; e.g., when
γ2xy = 0.6 the weight on the squared error is reduced by 50%. The frequency response
estimation is performed with MATLAB tfestimate() to obtain Ĝl , using Welch’s estima-
tion method and the H1 estimator. The optimization problem for the cost function in
Equation (2) is solved using the MATLAB fminsearch() function, which uses the Nelder–
Mead simplex direct search method, a derivative-free, unconstrained multivariate opti-
mization method. Equally logarithmically spaced frequency points were chosen on a range
where the crossover frequencies of the final control law were expected to be placed with
a good margin on either side. The expected range of the controller crossover frequencies
is ω ∈ [1 : 45] rad/s, based on engineering judgement by testing hypothetical actuator
performance values and performing the associated control design. At high frequency,
several of the actuators deviated from second-order behavior, and this part was neglected
in the estimation by reducing the highest estimation frequency. In Figures (Figures 3–5),
the frequency points for model estimation are indicated with black dots, the model re-
sponse is indicated with a red line in the region of estimation, and a yellow line for the
extrapolated part. Each figure contains a marking in the top right with relevant actuator
and maneuver setup. “xiLi B” indicates aileron (xi) on the left side (L) using the inner (i)
deflection measurement device. The linkage type is indicated with (B) for ball link and
(SL) for standard plastic bearing. Also, the amplitude (A), start and end frequency (f0 and
f1) and sine-sweep length (T) is indicated. The estimated model parameters are given in
Table 3.

0 0
Amplification [dB]

Amplification [dB]

-2
-5
-4
-10 Data Data
Model -6 Model
Estimation Point Estimation Point
-15 -8
Model Extrapolation Model Extrapolation

0 1 2
10 10 10 10 0 10 1 10 2

0 0
Phase [deg]

Phase [deg]

-100 -100

-200 -200

10 0 10 1 10 2 10 0 10 1 10 2

1 1
Coherence [-]

Coherence [-]

0.5 0.5

0 0
0 1 2
10 10 10 10 0 10 1 10 2
Angular frequency [rad/s] Angular frequency [rad/s]

Figure 3. Comparison of model vs. in-flight test data for actuator frequency response .
Actuators 2024, 13, 130 6 of 33

0 0

Amplification [dB]

Amplification [dB]
-5 -5

-10 Data -10 Data

Model Model
Estimation Point Estimation Point
-15 -15
Model Extrapolation Model Extrapolation

10 0 10 1 10 2 10 0 10 1 10 2

0 0
Phase [deg]

Phase [deg]
-100 -100

-200 -200

0 1 2 0 1 2
10 10 10 10 10 10

1 1
Coherence [-]

Coherence [-]
0.5 0.5

0 0
0 1 2 0 1 2
10 10 10 10 10 10
Angular frequency [rad/s] Angular frequency [rad/s]

Figure 4. Comparison of model vs. in-flight test data for actuator frequency response.

0 0
Amplification [dB]

Magnitude [dB]

-5 -5

-10 Data -10 Data

Model Model
Estimation Point Estimation Point
-15 -15
Model Extrapolation Model Extrapolation

0 1 2 0 1 2
10 10 10 10 10 10

0 0
Phase [deg]

Phase [deg]

-100 -100

-200 -200

10 0 10 1 10 2 10 0 10 1 10 2

1 1
Coherence [-]

Coherence [-]

0.5 0.5

0 0

10 0 10 1 10 2 10 0 10 1 10 2
Angular frequency [rad/s] Angular frequency [rad/s]

Figure 5. Comparison of model vs. in-flight test data for actuator frequency response.
Actuators 2024, 13, 130 7 of 33

Table 3. Comparison of estimated parameters for the actuators on ground and in flight (around
20 m/s ± 2 m/s). Delay is in seconds; eigenfrequency and bandwidth is in rad/s. ω BW denotes the
bandwidth, i.e., the frequency where the magnitude first drops 3 dB from the DC value, and ω∆60 is
the frequency where phase is first dropped by 60◦ .

On Ground Teq ω0,A ζA KA ω BW ω∆60 J

Aileron
MG90S 0.016 31.9 0.45 0.87 42.2 19.8 10.9
MS320 ball link 0.030 43.6 0.78 0.99 39.4 15.8 9.7
M5251H 0.032 95.9 1.00 0.97 71.1 21.2 1.8
M5252H ball link 0.039 92.9 1.00 0.93 92.6 20.8 1.9
Ruddervator
M5252H 0.028 92.4 0.80 0.93 79.9 22.4 7.8
M5252H ball link 0.030 108.2 0.94 0.93 76.3 23.0 10.5
In Flight Teq ω0,A ζA KA ω BW ω∆60 J
Aileron
MG90S 0.014 31.3 0.42 0.81 42.4 20.5 21.9
MS320 ball link 0.032 46.4 0.77 0.87 42.4 15.9 7.4
M5251H 0.028 95.4 0.98 0.89 63.5 22.0 3.3
M5252H ball link 0.028 87.9 0.73 0.85 85.5 23.3 3.8
Ruddervator
M5252H 0.028 88.1 0.75 0.93 82.7 23.0 7.4
M5252H ball link 0.028 85.3 0.73 0.82 82.5 22.8 5.3

3.5. Actuator Discussion

The estimated actuator parameters can be used to make a few statements on the use
for flight control and the considerations necessary in the control design.
1. A second-order system with time delay can be used to reasonably represent the
dynamics for most of the actuators both on ground and in flight, at least up to
approximately the bandwidth frequency. The faster actuators M5251H and M5252H
have the best model fit. It may be considered to increase the model order for the
slower MG90S and MS320 to obtain a better fit.
2. The reduction in bandwidth from ground to in flight at the given conditions is at
maximum ∼10%.
3. There can be significant in-flight bias (DC offset) between actuator/surface command
and surface position.
4. There is noticeable reduction in gain (K A ) between ground and in-air tests.
5. The M5252H actuator, which has both high torque and high speed indicated in the
datasheet, has significantly higher bandwidth than the other actuators.
6. The frequency response of the actuator in flight is reproducible with good accuracy.
7. High coherence values extend to higher frequencies for M5252H and M5251H than
for the MG90S.
As will be shown later, the fastest dynamic mode of the lateral/directional dynamics
for this aircraft is around 32 rad/s (the roll mode). This is basically equal to the bandwidth
of the MS320 and MG90S actuator, and leaves few options to control this dynamic mode.
The bandwidth of the M5252H actuator is twice as fast, and is more suitable for this control
task. It seems to be a general issue for control of subscale aircraft that the frequency
separation between the actuators and the dynamic modes is small. For further flight test
and control law testing, the M5251H and M5252H actuators will be used due to their
higher performance.

4. Identification of Flight Dynamic Model

A linear state space model of the rigid, bare airframe dynamics is sought for control
design. This can be obtained in many ways, but for the purpose of this paper, a single
Actuators 2024, 13, 130 8 of 33

linear model is estimated. This model will be used for control design of the iDPI controller,
in order to demonstrate the applicability of the concept.

4.1. Model Structure: Linear-Directional Lateral Dynamics

The longitudinal and lateral-directional model is assumed to be sufficiently decoupled.
For a given trim point, the lateral-directional dynamics is assumed to be described via a
linear state space model on the form (see [121]):
    
∆ṙK Nr Nβ Np 0 ∆rK
 ∆ β̇ K   Yr − cos α0 g
   Yβ Yp + sin α0 V0 cos θ0   
 ∆β K 
 ∆ ṗK  =  Lr Lβ Lp 0   ∆pK 
∆ϕ̇ tan θ0 0 1 0 ∆ϕ
    (4)
Nξ Nζ  − Nβ /V0 − Np − Nr  
 Yξ Yζ 

  vg
∆ξ −Yβ /V0 −Yp −Yr 
+
 Lξ Lζ 
 + − L β /V0 pg 
∆ζ − L p − Lr 
rg
0 0 0 0 0

with the output equation given by:

 
    ∆rK
∆β A 0 1 0 0  ∆β K 
=  
∆ny (V0 /g)Yr (V0 /g)Yβ (V0 /g)Yp 0  ∆pK 
∆ϕ
  
0 0 ∆ξ (5)
+
(V0 /g)Yξ (V0 /g)Yζ ∆ζ
 
  vg
−1/V0 0 0  pg 
+
−Yβ /g −(V0 /g)Yp −(V0 /g)Yr
rg

where V0 is the trim true airspeed, the states rK , β K , pK and ϕ are the kinematic yaw rate,
side slip angle, roll rate and bank angle, and the outputs β A and ny are the aerodynamic
side slip angle and the lateral load-factor in the aerodynamic frame at the center of gravity.
The lateral load-factor is equal to the specific force in the lateral direction denoted f y /g.
The inputs ξ and ζ are the aileron and rudder deflections, and v g , p g and r g are the side
gust velocity and rotational gust components. Nr , Nβ , Np , Nξ and Nζ are the yaw moment
derivatives; Yr , Yβ , Yp , Yξ and Yζ are the lateral specific force derivatives; Lr , L β , L p , Lξ and
Lζ are roll moment derivatives. The delta (∆) denotes deviations from the trim state.

4.2. Experiment Design

In practice, all maneuvers have to be flown line of sight. To have proper control over
the general flight path of the aircraft, allowing for sufficient space to maneuver at the end
of each maneuver, a maximum maneuver (recording) length is given by Trec = 14 s, which
includes a 1 s straight segment before the excitation starts, a 12 s logarithmic sine-sweep
and then again a 1 s segment with no excitation (the sine-sweep could be improved by
fading in/out the response to avoid the sharp beginning/end). The 1 s segments allow one
to verify correct trim conditions before the beginning of the maneuver, and allow for the
response to fade away after the maneuver. The frequency was varied from 0.5 Hz to 18 Hz,
with an amplitude of 10◦ . This frequency range was found to cover the dynamic modes
of the lateral dynamics and the effects of the actuators. The amplitude was found as a
compromise between obtaining a sufficient signal-to-noise ratio and not forcing the aircraft
too far away from the trim point. Figures 6 and 7 show examples of the used excitation
maneuver. The maneuver was repeated three times (in three different runs), and then
concatenated afterwards to form a sequence with length TF = 42 s. The concatenation
allowed a larger window size when estimating the frequency response functions with
Actuators 2024, 13, 130 9 of 33

Welch’s estimate. This results in an increased frequency resolution with higher coherence.
A relatively high frequency resolution is required to prevent that the peaks from the dutch
roll damping become smeared out in the estimated frequency response, which would
result in an overestimation of the damping ratio. The multisteps were performed with
asymmetric excitation in order to avoid drift away of the aircraft after the maneuver.

4.3. Estimation Data

A single sine-sweep run of the lateral and the directional channel (flown separately) is
shown on the left in Figure 6 and Figure 7, respectively. For the sine-sweep on the (virtual)
rudder, the low dampened dutch-roll mode is clearly visible in the yaw rate and lateral load
factor response. Also, the direct feed-through from the virtual rudder to lateral load-factor
is clearly seen at the high frequency part of the sweep as a constant amplitude response in
the lateral load-factor. The angular acceleration estimates, ṗˆ pp and ṙˆ pp , were obtained via
post processing the angular rates, with a zero-phase forward–backward derivative filter,
and are shown for reference only; the data were not used for estimation. The aircraft was
flown between 15 and 50 m above the ground for the system identification and control
law verification. It was sought to perform the flight test in low turbulence and low wind
conditions. For the data used in this paper, the wind speed was below 4 m/s.

4.4. Parameter Estimation

The model parameters to be estimated are listed in Table 4, in addition to the input–
output time delay Td . The parameter estimation is performed using the cost function in
Equation (2). The model containing the parameters to be estimated is given by
Equations (4) and (5), connected with the sensor filters as described in Appendix C.1. This
allows one to use the surface deflection measurements as input and the sensor measure-
ments as outputs. For the remaining parameters, V0 is taken as the average measured
velocity for the test runs, for θ0 the average of the estimated initial pitch angle is used
and for the trim angle of attack, the relation α0 = θ0 − γ0 is used, where γ0 is taken
as the initial inertial flight path angle estimated from the GPS. The frequency response
estimation and optimization is solved in the same manner as the servo-dynamics estima-
tion in Section 2, using the default options for the Welch estimate. As opposed to the
SISO servo estimation, the estimation for the plant uses a MISO estimate (using MAT-
LAB tfestimate(___,‘mimo’)) of the frequency response pairs; i.e., both virtual rudder and
aileron were used as input, although only the estimation related to the input containing the
sine-sweep was used. The other input is there to compensate for the fact that corrective
aileron was used during some virtual rudder maneuvers and vice versa to keep the aircraft
around the trim point. Five frequency response input–output pairs were chosen for the
model estimation, r/ξ, p/ξ, r/ζ, ny /ζ, p/ζ. Aileron to lateral load factor were left out
as there is almost no coupling, and the two bank angle pairs were left out, as the bank
angle estimate arises from an internal estimation filter, rather than a direct measurement.
Often this can be compensated by flight path reconstruction, but was left out in order to
avoid introducing more hyperparameters. The frequency estimation region was chosen for
the individual pairs according to the quality of the data, i.e., coherence and noise on the
frequency response estimate.

Table 4. Estimated model parameters and an approximate 95% confidence interval based on the
diagonal of the inverse of the Hessian matrix.

Nr −1.825 ±0.744 Yr 0.016 ±0.030 Lr 1.638 ±6.003

Nβ 5.366 ±4.847 Yβ −1.218 ±0.415 Lβ −218.730 ±52.031
Np −3.058 ±0.653 Yp −0.081 ±0.058 Lp −31.440 ±7.043
Nξ −18.301 ±4.396 Yξ −0.518 ±0.532 Lξ −239.410 ±46.642
Nζ −20.024 ±2.499 Yζ 0.271 ±0.040 Lζ 19.672 ±16.865
Actuators 2024, 13, 130 10 of 33

10
10
0
0
cmd
-10
pos cmd
-10
pos
446 448 450 452 454 456 458 460
492 492.5 493 493.5 494 494.5 495 495.5 496 496.5

1000 post process

post process
500 1000
0
0
-500
-1000 -1000

446 448 450 452 454 456 458 460

492 492.5 493 493.5 494 494.5 495 495.5 496 496.5

meas
100 meas
100 model
0
0

-100
-100

446 448 450 452 454 456 458 460

492 492.5 493 493.5 494 494.5 495 495.5 496 496.5

20 meas 0 meas

0 -10
-20
-20
-40
-30
446 448 450 452 454 456 458 460 492 492.5 493 493.5 494 494.5 495 495.5 496 496.5

490
meas meas
480
300
470
460 295

450
290
440
446 448 450 452 454 456 458 460 492 492.5 493 493.5 494 494.5 495 495.5 496 496.5
Time (seconds) Time (seconds)

Figure 6. Flight test of aileron sine-sweep and 3-2-1-1.

cmd
10 pos cmd
10 pos
0
0
-10
-10

566 568 570 572 574 576 578 580

691 692 693 694 695 696 697

post process
500 post process
500

0
0

-500
-500

566 568 570 572 574 576 578 580

691 692 693 694 695 696 697

100
100
meas
meas
50
50 model

0
0
-50 -50

-100 -100
566 568 570 572 574 576 578 580 691 692 693 694 695 696 697

0.6
meas meas
0.5 0.4
model
0.2
0 0
-0.2
-0.5
-0.4
566 568 570 572 574 576 578 580 691 692 693 694 695 696 697

40 meas meas
20
20 10

0 0
-10
-20
-20
566 568 570 572 574 576 578 580 691 692 693 694 695 696 697
Time (seconds) Time (seconds)

Figure 7. Flight test of virtual rudder sine-sweep and 3-2-1-1.

4.5. Estimation Results

The estimated model parameters are given in Table 4. The input–output time delay was
estimated to be 0 s, given by the lower boundary (note that this is the time delay between
surface deflection and flight dynamic response, which does not contain the time delay from
the actuator commands to the surface deflection response, which was estimated in Section 3).
The minimum of the optimization was found at J = 30.9, which is below the guideline
of 50 from [5]. The fit of the model for the considered frequency response pairs is shown
in Figures 8–10. It is seen that the dutch roll is well identified in the virtual rudder
response pairs. In the rudder to lateral load factor frequency response, it was chosen to
Actuators 2024, 13, 130 11 of 33

not include the amplification increase at high frequency (above 50 rad/s), as indicated by
the yellow line. As the controller crossover frequencies are well below this value, it was
not deemed necessary to increase the model order to capture these dynamics. Also, the
aileron to roll-rate frequency response has some discrepancy at high frequency. The flight
test data for this estimation had different actuators for each aileron, namely M5252H and
M5251H with different bandwidth, which may have led to asymmetric excitation at high
frequency. These discrepancies motivate the additional use of non-parametric stability
margin assessment considered in Section 5.6. The non-parametric assessment contains
all the un-modelled dynamics, and hence provides a higher confidence in the stability
and robustness of the control law for the envelope point of the test data. The correlation
between the parameter estimates was calculated using a numerical approximation of the
Hessian matrix, at the optimal solution for parameters. The estimated correlation matrix is
given in Figure 11. It is seen that the correlation is low between most parameters. A general
guideline is that a correlation above 0.9 for a given parameter pair can cause problems due
to collinearity [4]. High correlations are present for the aileron effectiveness and roll rate
parameters, in addition to the high correlation between the side-force coefficient due to
side-slip and roll rate. As can be seen from the time domain multistep shown in Figure 6,
a small oscillation at around 6 Hz is present at all times in the roll rate response. Further
investigation will be conducted into the cause and effect of this, which could arise from
wing structural vibration, turbulence or another source.
The identified model is also validated open loop in the time domain with a multistep
response as shown in Figures 6 and 7 to the right. The model data are an open-loop
simulation with surface deflection measurements as input, and sensor model values as
output. It is seen that the model represents reasonably well the dynamics of the aircraft
also in the time domain.
The identified model is also validated closed-loop in the time domain with a multistep
response on the lateral-load factor command as shown in Figure 25. The model data are
a closed-loop simulation with the plant model, the actuator model and the developed
controller, starting from the trim value. It is seen that the model reasonably well represents
the closed-loop dynamics of the aircraft and controller also in closed loop, validating its
usefulness for control law design.
Magnitude [dB]

20 20
Magnitude [dB]

0 0

-20 -20
10 0 10 1 10 2 10 0 10 1 10 2

-100
-200
Phase [deg]
Phase [deg]

-200
-300

-300
-400

-400
10 0 10 1
10 2 10 0 10 1 10 2

1 1
Coherence [-]
Coherence [-]

Data Data
0.5 Model 0.5 Model
Estimation Point Estimation Point
Model Extrapolation Model Extrapolation
0 0
10 0 10 1 10 2 10 0 10 1 10 2
Angular frequency [rad/s] Angular frequency [rad/s]

Figure 8. Measured and model frequency response for virtual rudder input-output pairs.
Actuators 2024, 13, 130 12 of 33

10 10

Magnitude [dB]
Magnitude [dB]
0
0
-10

-20
-10
10 0 10 1 10 2 10 0 10 1 10 2
-200
-200

Phase [deg]
Phase [deg]
-400
-300
-600
-400
-800
10 0 10 1 10 2 10 0 10 1 10 2

1 1

Coherence [-]
Coherence [-]

Data Data
0.5 Model 0.5 Model
Estimation Point Estimation Point
Model Extrapolation Model Extrapolation
0 0
10 0 10 1 10 2 10 0 10 1 10 2
Angular frequency [rad/s] Angular frequency [rad/s]

Figure 9. Measured and model frequency response for virtual rudder and aileron input-output pairs.
Magnitude [dB]

20

15

10

5
10 0 10 1 10 2
-150
Phase [deg]

-200

-250

-300
10 0 10 1 10 2

1
Coherence [-]

Data
0.5 Model
Estimation Point
Model Extrapolation
0
10 0 10 1 10 2
Angular frequency [rad/s]

Figure 10. Measured and model frequency response for aileron input-output pairs.

1
1

0.127 1 0.9

0.182 0.833 1
0.8
0.137 0.813 0.929 1

0.24 0.329 0.047 0.097 1 0.7

0.462 0.053 0.063 0.068 0.102 1

0.6
0.022 0.041 0.006 0.004 0.002 0.541 1

0.042 0.011 0.039 0.025 0.136 0.529 0.963 1 0.5

0.015 0.002 0.033 0.003 0.114 0.424 0.819 0.851 1

0.4
0.104 0.194 0.008 0.037 0.373 0.134 0.378 0.353 0.302 1

0.86 0.089 0.217 0.186 0.167 0.482 0.051 0.059 0.055 0.165 1
0.3

0.087 0.701 0.61 0.651 0.315 0.041 0.032 0.017 0.249 0.173 0.244 1
0.2
0.032 0.61 0.725 0.735 0.124 0.11 0.029 0.01 0.252 0.055 0.233 0.907 1

0.002 0.524 0.583 0.703 0.144 0.069 0.045 0.053 0.136 0.108 0.191 0.873 0.934 1
0.1
0.02 0.313 0.07 0.12 0.657 0.012 0.154 0.207 0.249 0.512 0.034 0.419 0.219 0.234 1

0.036 0.219 0.242 0.223 0.07 0.119 0.085 0.097 0.022 0.107 0.055 0.349 0.356 0.348 0.16 1
NaN

Figure 11. Estimated parameter correlations based on the inverse of the Hessian matrix.
Actuators 2024, 13, 130 13 of 33

5. Control Design and Evaluation

In the following section, the control design and control evaluation through flight test
will be presented. The control technique will be based on the incremental-differential-
proportional-integral (iDP) technique as introduced in [1]. As described in [1], the advan-
tages of the iDPI control law structure over a classical PI control law are:
• Avoids hidden coupling terms when scheduling the controller gains;
• Bump less transfer between different gain groups;
• Enhanced integrator wind-up prevention;
• Robust performance with respect to variations in model parameters;
• Enhanced disturbance rejection;
• Less sensitive to unknown time delays at the input.

5.1. Control Law Structure

In order to control the lateral-directional motion of the aircraft, a MIMO control law is
designed using eigenstructure assignment for the gain design. The control variables are
chosen to be the bank angle ϕ and the lateral load-factor in the body frame ny,B . As shown
later, a command filter is used to create a rate command attitude hold behavior for the roll
channel such that the piloted command is roll rate pc or ϕ̇c . Steering commands from an
outer-loop can be directly applied as bank angle commands ϕc . In order to apply the iDPI
control law structure, the measurements will have to be differentiated, and the commands
have to be applied to the incremental loop with ūc , as shown in Figure 12. The derivative
filters before the controller gains cancel with the virtual integrator filter after the control
gains in the fixed gain linear analysis, and remove the hidden coupling terms when
analyzing the gain scheduled controller, as shown in [1]. Instead of a filtered derivative of
the side-slip angle, filtering the kinematic relation for the aerodynamic side-slip angle rate
is used as feedback, as derived in Section 5.2. This allows for a pseudo full state feedback
of the flight dynamic variables of the lateral motion described by Equation (4). The control
law block diagram is shown in Figure 12.

𝑠
𝜙𝑐 𝐡𝝓 +
𝜏𝑓 𝑠 + 1 Incremental Loop
+ 1
𝜙 𝐤 𝝓𝒊 + + 𝒖𝒄 𝒖
− 𝜏𝑓 𝑠 + 1 ෩ 𝐴−1
𝐊 𝐆𝐴
𝑠 ഥ𝒄
𝒖 +
𝑛𝑦,𝑐 𝐡𝒏𝒚 + Δ𝒖𝒄
𝜏𝑓 𝑠 + 1 1
+ 1 𝜏𝑓 𝑠 + 1
𝑛𝑦,𝐵 𝐤 𝒏𝒚 𝒊 +
− 𝜏𝑓 𝑠 + 1
Integral Action 𝜏𝑓 𝑠 + 1
𝐆ത𝐴,𝑓 𝐆𝑃
Feedback ഥ𝒄
𝒖 𝑠 𝒖 𝒚
𝑠
Δ𝑟𝐾,𝑡𝑢𝑟𝑛 𝐤𝒓 +
𝜏𝑓 𝑠 + 1
1
𝛽መ𝐴ሶ 𝜏𝑓 𝑠 + 1
𝐤𝜷 +

𝑠
Δ𝑝𝐾,𝑡𝑢𝑟𝑛 𝐤𝒑 +
𝜏𝑓 𝑠 + 1
𝑠
𝜙 𝐤𝝓 +
𝜏𝑓 𝑠 + 1

Figure 12. Block diagram of lateral control law.

The bank angle command is given by ϕc and the lateral load-factor command by
ny,c . ūc is the unscaled input command, which is scaled to the actual actuator command
1
uc via multiplication of the scale factor matrix K̃−
A . The scale factor matrix is calculated
such that the commands from ūc to u have unit gain. The actual implementation of the
incremental loop in this flight test campaign was carried out using the actuator models
Actuators 2024, 13, 130 14 of 33

Ĝ A from Section 3, instead of the actual measurements, as shown in Figure 13. Ĝ A is the
transfer function matrix with the respective SISO actuator models on the diagonal.
Incremental Loop
using actuator model

𝒖𝒄 𝒖 𝒚
ഥ𝒄
𝒖 +
෩ 𝐴−1
𝐊 𝐆𝐴 𝐆𝑃
+

1
𝐆෠𝐴
𝜏𝑓 𝑠 + 1

Figure 13. Block diagram of incremental loop with actuator model.

In order to create a rate command attitude hold behavior, an integrator is used to

transform a roll-rate pc (or ϕ̇c ) command into the bank angle command for the control law,
as given in Equation (6). Additionally, a zero is added to hide the spiral pole from the
roll-rate command to bank angle response, in a similar fashion to [88]. The gain design
moved the spiral pole to the left in order to increase the rejection of disturbances to the bank
angle. The prefilter equation is given in Equation (6) and the block diagram in Figure 14,
where τS is the closed-loop spiral pole.
 
1
τS s + 1
ϕc = pc (6)
s

1/𝜏𝑆

1 +
𝑝𝑐 𝜙𝑐
𝑠 +

Figure 14. Roll rate command filter.

5.2. Flow Angle Derivatives for State Feedback

For the lateral motion, classical controllers often employ feedback of either lateral load-
factor or side-slip angle in order to modify the frequency of the dutch-roll mode. In both
DPI and iDPI control law structures, the feedback signals must be differentiated. In order
to avoid differentiating the lateral load-factor or side-slip angle (which may be be available
through noisy flow angle measurement or through an uncertain model based estimate),
the kinematic relation for the aerodynamic side-slip angle rate can be used, which does not
contain any model parameters. The kinematic relations for the aerodynamic flow angles
were derived in [70], by considering the translational equations of motion, inserting the
relation between aerodynamic, kinematic and wind velocity, and then expanding using the
rotation of the aerodynamic frame with respect to the body frame. The resulting equations
are given by Equation (7):

EA
E 
V̇AG / VAG g

(f) B ( g )O

 
 β̇ A  = M

E AB + M BO
VAG g g
α̇ A cos β A A
 
0
(7)
− rK cos α A − pK sin α A 
pK cos α A sin β A − qK cos β A + rK sin α A sin β A A
1   EO
G
−
E M AB MBO V̇W O
VAG
Actuators 2024, 13, 130 15 of 33

E
where VAG is the magnitude of the aerodynamic velocity (TAS), superscript G refers to
the center of gravity of the aircraft, subscript A denotes that the velocity is the aerodynamic
velocity and superscript E indicates that the velocities are taken with respect to the earth
fixed earth centered reference frame. The scalar time derivative of the true airspeed in

EA
the aerodynamic frame is given by V̇AG , where the second superscript A denotes the
derivative of the velocity is taken in the A frame. M AB is the rotation matrix between the
body frame B and the aerodynamic reference frame A given in Equation (A18), and MBO is
the rotation matrix between the NED frame O and the body frame, given in Equation (A19).
(f) B is the (mass) specific forces denoted in the B frame, which is the quantity measured by


G EO is the change of
the accelerometers. (g)O is the gravity vector in the NED frame. V̇W O
the wind velocity in the NED frame with respect to time. In a constant wind field, this term
is zero. If the wind gusts’ effects on the side-slip angle rate are not considered, the last term
in Equation (7) can be neglected. The resulting side-slip rate estimation is given by:
 
g (f) B ( g )O
β̇ˆ A = T

E ey M AB + MBO − rK cos α A + pK sin α A (8)
VAG g g

with eyT being the unit vector in the y direction of the associated frame (in this case in the A
frame). For many aircraft, including small or subscale aircraft, a measurement or estimate
of α A and β A , as, for example, investigated in [122], may not be feasible. An approximation
may be obtained considering M AB ≈ I, rK cos α A ≈ rK and pK sin α A ≈ 0. This results in a
side-slip angle rate estimate:
g


β̇ˆ A =
E ny B
+ cos θ sin ϕ − rK (9)
VAG

f y,B
where ny B
= g and
 
− sin θ
( g )O
MBO =  cos θ sin ϕ  . (10)
g
cos θ cos ϕ B
A similar expression can be derived for the approximate angle of attack rate:

−g
α̇ˆ =
E ((nz ) B − cos θ cos ϕ) + qK (11)
VAG

It is important to note that the simplifications in Equations (9) and (11) lead to the correct
steady condition (i.e., if β̇ A = 0, then also β̇ˆ A = 0) in a non-accelerated straight flight. This
observation is important for the implementation of iDPI, as this results in correct steady-
state tracking of the control variable in non-accelerated straight flight. The expressions for
Equations (9) and (11) are included in the gain design model, such that the effects of the
simplifications are minimized.

5.3. Turn Coordination

Turn coordination is applied to the body angular rate measurements in order to ac-
count for the non-linear effects caused by the rolling and yawing motion during a roll and a
subsequent turn. This will minimize side-slip excursions due to the non-linearities. The cor-
rection is applied in order to avoid the controller reacting to the non-linear contribution
of the body angular rates caused by a roll/turn. The incremental angular rate feedback is
given by:

∆pK,turn = pK − pK,turn
(12)
∆rK,turn =rK − rK,turn
Actuators 2024, 13, 130 16 of 33

where the turn contributions are given by:

g
pK,turn = −
E tan ϕ sin θ
VAG
g (13)
rK,turn =
cos θ sin ϕ
G E
VA

as derived in Appendix A.

5.4. Gain Design

The controller gain matrices kϕi , kny i , kr , kβ , k p and kϕ are calculated via eigenstruc-
ture assignment as in [88]. The perceived control law integrator behavior was removed in
the command path by placing the zeros introduced by the feed forward gains hϕ , hny on the
respective closed-loop integrator in the relevant command/control-variable input–output
pair. The two sets of gain design parameters in eigenstructure assignment are (1) the
location of the poles and (2) the corresponding eigenvector structure defined by a design
matrix. Several requirements have to be fulfilled, and are verified using the design and
assessment models. Not all requirements have a straightforward relationship with the gain
design parameters, and an iterative process is applied, with variation in the gain design
parameters in the eigenstructure assignment until all requirements are fulfilled.
The design model is constructed using MATLAB state space models for the plant,
sensor, actuator, actuator reference model, digital signal processing and controller filters.
The models are connected to a design model using the MATLAB connect() command to
generate one design state space model with states:
 
x f dm
 xsens/dsp 
 
x=
 xact ,
 (14)
 x 
act,mdl
x f ilters

with input ūc and output being (ϕ, ny,B , ∆rK,turn , β̇ˆ A , ∆pK,turn , ϕ) filtered with the appro-
priate filters in Figure 12. Note that there are no dedicated integrator states, as these are
inherently contained in the incremental loop.
The requirements to be fulfilled consist of:
• Eigenmode characteristics as given in Table 5. The dutch-roll damping was increased
and the spiral pole was further stabilized, as also shown in Figures 15 and 16.
• Eigenmode response decoupling by assigning the desired ϕ/β ratio as given in Table 6
by modifying the desired eigenvector structure given by Table 7. This can be compared
to the open-loop eigenvectors given in Table 8
• Integral behavior by assigning closed-loop integrator poles as given by Table 5.
• Gain margin (>6 dB) and phase margin (>45◦ ). Additionally, it is sought to have the
time delay margin above four samples (>0.04 s). The final design margins are shown
in Figure 17.
• Limit the gain and phase crossover frequencies to below 45 rad/s in order to provide
sufficient margin against model uncertainties at higher frequencies.
• Limitations on maximum actuator rate and position in relation to step commands and
gusts as depicted in Figures 18 and 19. The following limits are empirically found to
work, but should be increased and verified via further flight test. For a step gust of
v g = 3 m/s or command ny,c = 0.2, the actuator rate is sought to be less than 10% of
the unloaded maximum, and the actuator position less than 50% of the maximum.
Several additional requirements could be added, including turbulence response, noise
suppression, attenuation of structural modes, additional robustness metrics, bandwidth,
Actuators 2024, 13, 130 17 of 33

handling qualities, etc.; however, these are not directly relevant for the current investigation,
and could be added for further work.

Table 5. Open-loop and desired (subscript d) closed-loop eigenvalue location (omega in rad/s).

ω ζ ωd ζd
Roll Mode 32.0 – 32.0 –
Dutch-Roll
5.6 0.19 4.0 0.85
Mode
Spiral Mode 0.17 – 1.25 –
Integrator ϕ – – 70.0 –
Integrator ny – – 75.0 –

Table 6. Phi/beta ratio (|ϕ|/| β|).

|ϕ|/| β|
Open-Loop 1.27
Closed-Loop 0.1

Table 7. Assignment of closed-loop eigenvector entries. eϕ and eny are the control law (virtual)
integrator states.

r β p ϕ eϕ eny

0.03 − 1 − − − Roll Mode
 − 1 − 0.1 − − Dutch Roll Mode 1
 
 − 1 − 0.1 − − Dutch Roll Mode 2
 
 − 0.03 − 1 − − Spiral Mode
 
− − − − 1 0 Integrator Bank Angle
− − − − 0 1 Integrator Lateral Load-Factor

Table 8. Open-loop eigenvector entry magnitudes.

r β p ϕ
0.0995 0.0040 0.9945 0.0311 Roll Mode
0.5600 0.1140 0.8078 0.1446 Dutch Roll Mode 1
 
0.5600 0.1140 0.8078 0.1446 Dutch Roll Mode 2
0.4381 0.0296 0.1845 0.8793 Spiral Mode

0.4 0.28 0.18 0.09 8

8

7
7
0.52
6
6

5
5
0.66
4
4

3
3 0.82

2
2
0.94
1
1

-4 -3.5 -3 -2.5 -2 -1.5 -1 -0.5 0 0.5

Figure 15. Pole-zero map of open-loop response from aileron ξ to bank angle ϕ, and closed-loop
response from bank angle command ϕc to bank angle ϕ.
Actuators 2024, 13, 130 18 of 33

8
80.44 0.22

7
7

6
6

5
5 0.62

4
4
0.74
3
3
0.85
2
2
0.92
1
1 0.965
0.99
0

-4 -2 0 2 4 6 8

Figure 16. Pole-zero map of open−loop response from rudder ζ to lateral load-factor ny , and closed-
loop response from lateral load-factor command ny,c to lateral load-factor ny .

20

0
Mag (dB)

Mag (dB)
0
-10

-20 -20

630 720
Phase (deg)

Phase (deg)

630
540
540
450
450
360 360
100 101 102 100 101 102
Frequency (rad/s) Frequency (rad/s)

40

20
Mag (dB)

Mag (dB)

20

0 0

-20 -20
900 900
Phase (deg)

Phase (deg)

720 720

540 540

360 360
100 101 102 100 101 102
Frequency (rad/s) Frequency (rad/s)

0
0
Mag (dB)

Mag (dB)

-50 -20

-100 -40
720 720
Phase (deg)

Phase (deg)

540 540

360 360

180 180
100 101 102 100 101 102
Frequency (rad/s) Frequency (rad/s)

Figure 17. Bode plots for loop breaks of design model. The lowest values of margins are marked in bold.
Actuators 2024, 13, 130 19 of 33

20
Aileron cmd

Actuator command (deg)

Rudder cmd
10 50% of aileron limit
50% of rudder limit

-10

-20
0 0.5 1 1.5 2 2.5 3

100 Aileron rate

Rudder rate
Actuator rate (deg/s)

10% of unloaded actuator rate limit

50

-50

-100

0 0.5 1 1.5 2 2.5 3

5 0.5
Bank angle (deg)

2 0.2

0 0

-2 -0.2
Bank angle
Lateral load factor
-5 -0.5
0 0.5 1 1.5 2 2.5 3
Time (s)

Figure 18. Design model lateral load−factor step response.

20
Aileron cmd
Actuator command (deg)

Rudder cmd
10 50% of aileron limit
50% of rudder limit

-10

-20
0 0.5 1 1.5 2 2.5 3

100 Aileron rate

Rudder rate
Actuator rate (deg/s)

10% of unloaded actuator rate limit

50

-50

-100

0 0.5 1 1.5 2 2.5 3

15 1.5

10 1
Bank angle (deg)

5 0.5

0 0

-5 Bank angle -0.5

Bank angle open loop
-10 Lateral load factor -1
Lateral load factor open loop
-15 -1.5
0 0.5 1 1.5 2 2.5 3
Time (s)

Figure 19. Design model lateral step gust disturbance.

Actuators 2024, 13, 130 20 of 33

5.5. Implementation
The controller was implemented in Simulink in order to create auto-generated code to
be flashed onto the flight control computer. The continuous time control law is converted
to a discrete time implementation. The filters depicted in Figure 12 with time constant
τ f were implemented with a forward Euler discrete integrator. As the time constant
τ f = 0.125 is very high compared to the sampling time Ts = 0.01, the discretization
effects on the frequency response are minor compared to the advantage of the simplicity of
implementation, compared to, for example, a Tustin transform with prewarping. For the
actuator models in Ĝ A depicted in Figure 13, a zero-order hold transform is used. Referring
to Figure 21, it is noted that the controller outputs actuator commands with a zero-order
hold (ZOH) in between commands, and hence this transform is used to represent the
dynamics of the continuous actuators at the sampled time instants.

5.6. Assessment
An automatized assessment is performed to ensure that the control law is safe to fly.
The assessment is performed using three different models:
1. Gain design model (GDM):
Linear continuous time model, possibly with simplifications used to calculate the
controller gains.
2. Linear parametric assessment model (LPAM):
Discrete time linearization of the controller implementation model connected with
the zoh transform of the identified linear continuous time parametric models of the
actuators, plant and sensors. The zero order hold transform is used because the
controller outputs commands that are held between each time update of the controller,
i.e., zero order hold of the commands; using the zoh transform for assessment correctly
represents the dynamics of the sampled closed-loop continuous time system.
3. Non-parametric assessment model (NPAM):
Discrete time linearization of the controller implementation model connected with
the non-parametric frequency response estimates from flight test.
The interconnection between the three models is sketched in Figure 20.

Maneuver injection and data collection From data sheets Based on iDPI and eigenstructure assignment
Sensor Models Control Law
Flight Test Data SM CLAW
Simulink
System identification State
Controller Implementation Model
Space
Plant Model Model CIM
PM slLinearizer
Gain Design Model
Linearized Controller Implementation Model
System identification GDM
Actuator Model
LCIM
AM Differences between GDM and LPAM
- Controller discretization effects.
- Additional signal processing and delays in
Linear Parametric Assessment Model the flight code implementation.
- Implementation errors in the GDM and/or
LPAM LPAM.

Differences between LPAM and NPAM

- Identified model structure, or model
parameters do not correctly catch the
measured frequency response
Frequency Response Functions Non-Parametric Assessment Model
characteristics, e.g. un-modelled dynamics
FRF NPAM such as structural couplings, complex
aerodynamic & propulsive effects etc.
- Downsampling errors.

Figure 20. Relation between the models used for design, implementation and assessment.

The gain design model (GDM) is described in Section 5.4.

The LPAM is created using the continuous time linear models for the actuators GA (s),
plant GP (s) and sensors GS (s), and discrete time linear models for the digital signal
Actuators 2024, 13, 130 21 of 33

processing Cdsp (z) and the discrete time controller C(z). Even though the digital signal
processing model contains the discrete time digital downsampling filters as described
in Appendix C.1, these are considered continuous and embedded in the sensor model
GS (s) in the analysis, which may be reasonable as they are sampled 10× faster than the
controller, as shown in Figure 21 with Ts,1 = 0.001 and Ts = 0.01. An alternative would be
to represent both the controller and dsp at the higher sampling frequency, i.e., upsampling
the controller with no loss of accuracy. No additional dynamics of the sensor was found
in the addition to the digital signal processing (it is common to have a filter in the sensor
for the downsampling; however, the PX4 software moved the filtering to the flight control
computer, which samples at high rate). The dynamics of the estimate of ϕ was neglected.
The continuous time actuator model, plant model and in this analysis also the combined
sensor/dsp model are discretized with the zero order hold transform, using the same
argument as in Section 5.5; i.e., the controller outputs zero order hold commands. This will
produce correct evaluations when considering stability margins, but may lead to slight
errors when considering frequency responses for gusts and turbulence, the inputs of which
are in continuous time. However, since the sampling frequency is much higher than the
frequency of the turbulence, this is also deemed acceptable for this analysis. The LPAM
model creation is fully automatic using Simulink slLinearizer , which automatically linearizes
and discretizes the controller implementation model together with the actuator model,
plant model and sensor model.

𝑝𝐾 𝜉𝐿,𝑐
FCC
𝑟𝐾 𝜉𝑅,𝑐
𝜉𝑐
𝑝𝐾 𝑛𝑦,𝐵 𝜂𝐿,𝑐
𝜁𝑐
𝑟𝐾 𝜙 𝜂𝑅,𝑐
𝐂𝑑𝑠𝑝 (𝑧) 𝑆𝑡𝑎𝑡𝑖𝑐
𝑛𝑦,𝐵 𝐂(𝑧) 𝑍𝑂𝐻 𝐆𝐴 (𝑠) 𝐆𝑃 (𝑠)
𝑇𝑠,1 𝑇𝑠 𝑀𝑖𝑥𝑒𝑟

𝐆𝑆 (𝑠)

Figure 21. LPAM with loop breaks for stability and robustness assessment.

The NPAM is created using the linearized controller implementation model, similar
to the LPAM. However, instead of using the parametric models for the other subsystems,
the frequency response functions (FRF) obtained directly from flight test are used. These
FRFs contain all the effects on the real aircraft, and give the most accurate pointwise
estimate of the stability and robustness characteristics, as depicted in Figure 20. The SISO
loop breaks are calculated using Equations (17) and (18), which ensures only the loop
break under consideration is open; i.e., the other loop break is closed. The structure is
depicted in Figure 22. For each frequency ωk in the frequency response estimate, assemble
the frequency response estimate matrix GFRF (iωk ):
 
Ĝ pK ,ξ c Ĝ pK ,ζ c
 Ĝ ĜrK ,ζ c 
 rK ,ξ c 
ĜFRF =   (15)
 Ĝny,B ,ξ c Ĝny,B ,ζ c 
Ĝϕ,ξ c Ĝϕ,ζ c

Assemble the helper matrix G̃(iωk ):

G̃(iωk ) = C(eiωk Ts )ĜFRF (iωk ) (16)

where the discrete controller transfer function is evaluated in z = eiωk Ts using MATLAB’s fre-
qresp(). The loop transfer function with the rudder ζ c closed can be calculated by inspecting
Figure 22 as shown in Appendix D.1:

−1
Lξ (iωk ) = G̃11 (iωk ) + G̃12 (iωk ) 1 − G̃22 (iωk ) G̃21 (iωk ) (17)
Actuators 2024, 13, 130 22 of 33

and similarly for the rudder ζ c :

−1
Lζ (iωk ) = G̃22 (iωk ) + G̃21 (iωk ) 1 − G̃11 (iωk ) G̃12 (iωk ) (18)

The loop transfer function shapes are shown in Figures 23 and 24 with results from GDM,
LPAM and NPAM.

𝜉𝑐,𝑜𝑢𝑡 𝜉𝑐,𝑖𝑛 𝜉𝑐,𝑜𝑢𝑡 𝜉𝑐,𝑖𝑛

2×4
𝑝𝐾 𝜁𝑐,𝑜𝑢𝑡 𝜁𝑐,𝑖𝑛 𝜁𝑐,𝑜𝑢𝑡 𝜁𝑐,𝑖𝑛
𝑟𝐾 𝐂(𝑧)
𝑛𝑦,𝐵
𝜙

෨
𝐆(𝑧)
𝐆෠ 𝐹𝑅𝐹 (𝑖𝜔)
2×2
4×2
Figure 22. Schematic showing how to assemble the non-parametric loop breaks from frequency
response functions.

10 1

0.5
0

-5
0
-10

-15
-0.5
-20

-25
-1

-30
-400 -350 -300 -250 -200 -150 -100 -50 0 -1.5 -1 -0.5 0 0.5

Figure 23. Stability assessment for SISO loop break at aileron command ξ c .

GDM
10 1
LPAM
NPAM
5

0.5
0

-5
0
-10

-15
-0.5
-20

-25
-1

-30
-400 -350 -300 -250 -200 -150 -100 -50 0 -1.5 -1 -0.5 0 0.5

Figure 24. Stability assessment for SISO loop break at virtual rudder command ζ c .

5.7. Lateral Flight Test Results

The control law was flown in flight test to verify the behavior and applicability to
a subscale remote controlled aircraft. The closed-loop response was compared against
the design model, and showed very accurate prediction of the closed-loop time domain
response. A multistep test for the lateral load-factor is shown in Figure 25, where the
expected design model response is plotted on top of the actual aircraft response. Note that
Actuators 2024, 13, 130 23 of 33

the gains in the flight test slightly differed from those above because of changes in aircraft
configuration between flights and an intermittent crash; however, the model comparison
is made with a matching gain set. This is also the reason why there is no bank angle step
response evaluation. The flight test gain set had a desired ϕ/β ratio of 0.9, which causes the
higher bank angle coupling. There is a slight difference between the bank angle response
and the design model; however, the average trend is correct. This discrepancy may arise
from the oscillations on the roll rate measurements as discussed in Section 4.5, which may
translate into the bank angle estimate. It is also notable that both the virtual rudder and
aileron commands have small high-frequency oscillations, roughly at the same frequency
as the oscillations on the yaw rate and roll rate. These oscillations also exists in open loop;
see Figure 6; hence, they seem to not originate due to the control law. Further investigation
should be conducted to see whether notch filtering can be used to filter out the oscillations
in the measurement, and avoid them propagating to the actuator commands, as done in [1].

20 cmd 20 cmd
10 pos 10 pos

0 0
-10 -10
-20 -20
852 853 854 855 856 857 858 859 860 852 853 854 855 856 857 858 859 860

200 200
post process post process
100 100

0 0

-100 -100

-200 -200
852 853 854 855 856 857 858 859 860 852 853 854 855 856 857 858 859 860

r_turn 20 meas
20
meas design model
design model
0 0

-20 -20

852 853 854 855 856 857 858 859 860 852 853 854 855 856 857 858 859 860

meas 2 meas
0.2 n_y,c design model
design model 0
0
-2
-0.2
-4
852 853 854 855 856 857 858 859 860 852 853 854 855 856 857 858 859 860

2 meas meas
260
design model
0
250
-2
240
-4
852 853 854 855 856 857 858 859 860 852 853 854 855 856 857 858 859 860
Time (seconds) Time (seconds)

Figure 25. Flight test results with predicted and actual response for a lateral load−factor step
command. (*Note that for visualization purposes, the actuator commands have been corrected for in
flight bias from a straight segment run, e.g., ξ c,cor = ξ c + ξ c,bias such that ξ c,cor = ξ meas in the straight
and level segment).

6. Conclusions
This paper investigated the feasibility of practical system identification and incremen-
tal control design for a subscale fixed wing aircraft. A subscale demonstrator platform was
equipped with sensors for system identification and control, in particular surface deflection
measurements, which is uncommon for this class of aircraft. An iDPI control law was
synthesized to control the lateral-directional motion, and it was shown that the control law
performed well in flight test. It was shown that the identified model was able to accurately
predict the closed-loop behavior, motivating the use of the presented process for further
control law development. It was shown that eigenstructure assignment could successfully
be applied together with iDPI to form a multiple input/multiple output incremental control
law. The controller was shown to be easily implemented in discrete time and deployed
on a flight control computer. The control law was flight tested on the subscale aircraft,
and it was demonstrated that the nominal performance was as designed. In addition,
several actuator models were identified using surface deflection measurements, both on
ground and in flight. These models were used to support proper control design. Several
Actuators 2024, 13, 130 24 of 33

improvements can be made for future development, including composite windowing tech-
niques, increased sampling rates, estimation of non-dimensional aerodynamic/propulsion
coefficients, full envelope gain scheduling, investigation of handling qualities, experimental
estimation of actuator non-linearities, estimation of servo dynamic variation with airspeed,
comparison of internal and external servo position measurements, optimization of input
design, identification of structural couplings and notch filter design and incorporation.

Author Contributions: Conceptualization, R.S.; Methodology, R.S. and F.H.; Software, R.S. and K.G.;
Validation, R.S.; Formal analysis, R.S.; Investigation, R.S. and K.G.; Data curation, R.S. and K.G.;
Writing—original draft, R.S.; Writing—review & editing, R.S., K.G. and F.H.; Visualization, R.S.;
Supervision, F.H.; Project administration, R.S. and F.H.; Funding acquisition, F.H. All authors have
read and agreed to the published version of the manuscript.
Funding: This work was partially funded by the Federal Republic of Germany (Project: RAUDY,
FKZ: 20E1910B). Funding organization: Federal Ministry for Economic Affairs and Climate Action on
the basis of a decision by the German Bundestag.

Data Availability Statement: Dataset available on request from the authors.

Acknowledgments: The authors would like to thank Bernd Kartnaller and the model flight club
MCB Bregenz for allowing flight tests at their airfield. The authors would also like to give a special
thanks to Daniel Teubl for his work on the surface deflection measurement hardware which was
a fundamental resource to obtain the results in this paper. The authors would also like to thank
Alexander Zwenig for the help with conducting flight tests and assembling the aircraft, and Agnes
Steinert for discussion and feedback.
Conflicts of Interest: The authors declare no conflicts of interest.

Appendix A
Statement 1 (Trim contributions in side slip angle dynamics). The yaw-rate rK can be split
up into contributions due to maneuvering (pilot command) rK,man , trim compensation of gravity
rK,trim,g and kinematic coupling compensation rK,kc :

rK = rK,man + rK,trim,g + rK,kc (A1)

The specific lateral force f y,A can be split up into a trim lateral load factor f y,trim,A and a maneuver
lateral load factor ∆ f y,A contribution:

f y,A ∆ f y,A f y,trim,A

= + (A2)
g g g

The kinematic coupling compensation yaw rate, trim compensation of gravity and manuever yaw
rate are given by

rK,kc = pK tan α A (A3)

 
g f y,trim,A ( g )O
rK,trim,g =
E − eTj M AO (A4)
VAG cos α A g g
Actuators 2024, 13, 130 25 of 33

 
g ∆ f y,A
rK,man =
E (A5)
cos α A VAG g

where f y,trim,A is the specific trim lateral load factor and ∆ f y,A is the additional maneuver lateral
load factor. e j is the unit vector along the y-axis, and MAO = MAB MBO , where MAB and MBO
are given by Equations (A18) and (A19).

Derivation. The side-slip dynamics is given in Equation (7). By inserting Equation (A1)


G EO = 0, a relation denoted using the relevant contributions can be
and assuming V̇W O
formulated:
1 ∑ (F) A
β̇ A = G
eTj ( ) − rK cos α A + pk sin α A
(VA ) E m
1
= eT (f A + M AB MBO (g)O ) − (rK,man + rK,trim,g + rK,kc ) cos α A + pk sin α A (A6)
(VAG ) E j
g f y,A ( g )O
= ( + eTj M AB MBO ) − (rK,man + rK,trim,g + rK,kc ) cos α A + pk sin α A
(VAG ) E g g

(F) A
where eTj ∑ m = f y,A was used and e j is the unit vector along the y-axis. At trim condition
and with no pilot command, i.e., β̇ A = 0, the side-slip dynamics in Equation (A6), becomes:
 
g ∆ f y,A f y,trim,A ( g )O
0=
+ + e j T M AO − (rK,man + rK,trim,g + rK,kc ) cos α A + pk sin α A (A7)
VA G E g g g

The different yaw rate contribution terms can now be identified as:

rK,kc = pK tan α A (A8)

and
 
g f y,trim,A ( g )O
rK,trim,g =
E + e j T M AO
VAG cos α A g g
g f
y,trim,A (A9)
=
E + cos α A sin β A sin Θ+
VAG cos α A g

cos β A cos Θ sin Φ − sin α A sin β A cos Θ cos Φ

and  
g ∆ f y,A
rK,man =
E (A10)
cos α A VAG g

For design of the flight control laws, the turn contribution of the yaw rate is denoted
as follows:
rK,turn = rK,trim,g + rK,kc (A11)
Note that for zero angle of attack and side-slip angle, i.e., α A = 0, β A = 0, M AB is the
f y,trim,A
identity matrix and if the trim lateral load factor g = 0, then the yaw rate contributions
become

rK,kc = 0 (A12)

and
g
rK,turn = rK,trim,g =
E (cos θ sin ϕ) (A13)
VAG
Actuators 2024, 13, 130 26 of 33

which is similar to the trim yaw rate contribution from the inner-loop presented in [123],
except that the aerodynamic speed is used instead of kinematic speed. Considerations in
addition to the steady wind assumption in this paper can be applied and lead to expressions
containing the kinematic speed. However, in cases of strong winds, this may lead to the
turn contribution tending toward infinity, which is avoided with the chosen form.

Statement 2 (Roll rate in steady turn). Consider a steady state turn where the Euler angle
derivatives θ̇ and ϕ̇ are zero; then, the roll rate pK can be determined as a function of the yaw rate by:

sin θ
pK,turn = − rK,turn (A14)
cos ϕ cos θ

If the turn contribution for the yaw rate is given by the approximation in Equation (A13), then the
roll-rate contribution is given by:
g
pK,turn = −
E tan ϕ sin θ (A15)
VAG

Derivation: First consider the last row of Equation (A20) with θ̇ = 0 for a steady turn;
then:
ψ̇turn cos ϕ cos θ = rK,turn (A16)
Isolate for ψ̇turn :
1
ψ̇turn = rK,turn (A17)
cos ϕ cos θ
Insert the expression for ψ̇turn into the first row of Equation (A20) to obtain the general
relation for the roll rate in a steady turn presented in Equation (A14).

Appendix B
Statement 3 (Rotation from body frame to stability frame). The coordinate rotation matrix
from body frame B to aerodynamic frame A is given by [124]:
 
cos α A cos β A sin β A sin α A cos β A
M AB =  − cos α A sin β A cos β A − sin α A sin β A  (A18)
− sin α A 0 cos α A

Statement 4 (Rotation from NED to body frame). The coordinate rotation matrix from NED
frame O to body frame B is given by [124]:
 
cos ψ cos θ sin ψ cos θ − sin θ
 
MBO =  cos ψ sin θ sin ϕ − sin ψ cos ϕ sin ψ sin θ sin ϕ + cos ψ cos ϕ cos θ sin ϕ  (A19)
cos ψ sin θ cos ϕ + sin ψ sin ϕ sin ψ sin θ cos ϕ − cos ψ sin ϕ cos θ cos ϕ

Statement 5 (Relation between Euler angle rates and body rates). The rotational rate between
the body frame (B) and the NED frame (O) denoted in the body frame is given by [124]:
   
  ϕ̇ − ψ̇sinθ pK
ωOB =  θ̇cosϕ + ψ̇sinϕcosθ  =  qK  (A20)
B
−θ̇sinϕ + ψ̇cosϕcosθ B rK B

Appendix C
Appendix C.1. PX4 Sensor Downsampling Filter
The IMU measurements are filtered internally in the flight control computer at a high rate,
1000 Hz, before they are downsampled to be used in the controller at 100 Hz. The main purpose
is for anti-aliasing with respect to the downsampling, but it may also be used to filter out high
Actuators 2024, 13, 130 27 of 33

frequency noise. The filter is a second-order Butterworth filter (the filter implementation in
PX4 can be found at “Firmware/src/lib/mathlib/math/filter/[Link]”) with
cut off at 40 Hz for the accelerometer and 30 Hz for the gyro (tunable parameter), which is
below the Nyquist frequency. The filter implementation is given by the following equations.
Define the sampling frequency f s and cut of frequency f c in Hertz as:

f s = 1000 Hz
(A21)
f c = 40 Hz or 30 Hz.

Define the following helper variables:

fs
fr = fc  
Ω = tan π (A22)
fr

c = 1 + 2 cos π
4 Ω + Ω2 .

Then, the discrete time filter coefficients can be calculated to be:

b2 = Ω2 /c
b1 = 2b2
b0 = b2
a2 =1 (A23)
2 ( Ω2 −1 )
a1 = c
(1−2 cos( π4 )Ω+Ω2 )
a0 = c

such that the following implements the Butterworth filter:

b2 z2 + b1 z + b0
GB ( z ) = . (A24)
a2 z2 + a1 z + a0

The filter is equivalent to the filter generated via the MATLAB command butter().
The design model contains the continuous version of the filter, and the software implemen-
tation contains the discrete version, which is compared in Figure A1.
0 0

-20 -20
Magnitude (dB)

Magnitude (dB)

-40 -40

-60 -60

-80 -80

-100 -100

0 0

-45 -45
Phase (deg)

Phase (deg)

-90 -90
PX4 butterworth PX4 butterworth
-135 MATLAB butter discrete -135 MATLAB butter discrete
MATLAB butter continous MATLAB butter continous
-180 -180
100 rad/s 101 rad/s 102 rad/s 103 rad/s 104 rad/s 101 rad/s 102 rad/s 103 rad/s 104 rad/s
0 Hz 2 Hz 16 Hz 159 Hz 1592 Hz 2 Hz 16 Hz 159 Hz 1592 Hz

Figure A1. Frequency response of Butterworth filter applied to the accelerometer (left, f s = 30 Hz)
and the gyro (right, f s = 40 Hz).

Appendix D
Appendix D.1. Calculation of Non-Parametric Loop Break
Consider the multi-loop depicted in Figure 22. The loop transfer function Lξ (iω ) with
the ζ loop closed, i.e., ζ c,in = ζ c,out , can be calculated as follows. First write the expression
for ξ c,out :
ξ c,out = G̃11 ξ c,in + G̃12 ζ c,in , (A25)
Actuators 2024, 13, 130 28 of 33

where the numeric subscripts represents the transfer function matrix entries. Now apply
the condition that the rudder loop is closed:

ζ c,in = ζ c,out = G̃21 ξ c,in + G̃22 ζ c,in . (A26)

Isolate for ζ c,in .

I − G̃22 ζ c,in =G̃21 ξ c,in

−1 (A27)
ζ c,in = I − G̃22 G̃21 ξ c,in .

Insert this into Equation (A25) to obtain:

−1
ξ c,out = G̃11 ξ c,in + G̃12 I − G̃22 G̃21 ξ c,in . (A28)

Then it can be seen that:

ξ c,out
−1
(iω ) = G̃11 (iω ) + G̃12 (iω ) I − G̃22 (iω ) G̃21 (iω ) (A29)
ξ c,in ζ −closed

similarly for the rudder ζ c :

ζ c,out
−1
(iω ) = G̃22 (iω ) + G̃21 (iω ) I − G̃11 (iω ) G̃12 (iω ) (A30)
ζ c,in ξ −closed

References
1. Steffensen, R.; Steinert, A.; Holzapfel, F. Incremental control as an enhanced and robust implementation of gain scheduled
controllers avoiding hidden coupling terms. Aerosp. Sci. Technol. 2023, 141, 108500. [CrossRef]
2. Bacon, B.; Ostroff, A. Reconfigurable flight control using nonlinear dynamic inversion with a special accelerometer implementa-
tion. In Proceedings of the AIAA Guidance, Navigation, and Control Conference and Exhibit, Denver, CO, USA, 14–17 August
2000 ; p. 4565.
3. Sieberling, S.; Chu, Q.; Mulder, J. Robust flight control using incremental nonlinear dynamic inversion and angular acceleration
prediction. J. Guid. Control Dyn. 2010, 33, 1732–1742. [CrossRef]
4. Morelli, E.A.; Klein, V. Aircraft System Identification: Theory and Practice; Sunflyte Enterprises: Williamsburg, VA, USA, 2016;
Volume 2.
5. Tischler, M.B.; Remple, R.K. Aircraft and Rotorcraft System Identification; American Institute of Aeronautics and Astronautics:
Reston, VA, USA, 2012.
6. Jategaonkar, R.V. Flight Vehicle System Identification: A Time Domain Methodology; American Institute of Aeronautics and Astronau-
tics: Reston, VA, USA, 2006.
7. Cooper, J.K.; DeVore, M.; Reed, A.; Morelli, E.A. Deterministic and Probabilistic Approaches to Model and Update Dynamic
Systems. J. Aircr. 2023, 60, 1461–1479. [CrossRef]
8. Morelli, E.A.; Grauer, J.A. Advances in Aircraft System Identification at NASA Langley Research Center. J. Aircr. 2023, 60,
1354–1370. [CrossRef]
9. Leshikar, C.; Valasek, J.; McQuinn, C.K. System Identification of Unmanned Air Systems at Texas A&M University. J. Aircr. 2023,
60, 1437–1460.
10. Hosseini, B.; Steinert, A.; Hofmann, R.; Fang, X.; Steffensen, R.; Holzapfel, F.; Göttlicher, C. Advancements in the Theory and
Practice of Flight Vehicle System Identification. J. Aircr. 2023, 60, 1419–1436. [CrossRef]
11. Grauer, J.A.; Boucher, M.J. Aircraft system identification from multisine inputs and frequency responses. J. Guid. Control Dyn.
2020, 43, 2391–2398. [CrossRef]
12. Lampton, A.K.; Klyde, D.H.; Schulze, P.C. Overview of System Identification Tool Advancements at Systems Technology,
Incorporated. J. Aircr. 2023, 60, 1522–1537. [CrossRef]
13. Grauer, J.A. Frequency Response Estimation for Multiple Aircraft Control Loops Using Orthogonal Phase-Optimized Multisine
Inputs. Processes 2022, 10, 619. [CrossRef]
14. Grauer, J.A.; Morelli, E.A. Introduction to the Advances in Aircraft System Identification from Flight Test Data Virtual Collection.
J. Aircr. 2023, 60, 1329–1330. [CrossRef]
15. de Visser, C.C.; Pool, D.M. Stalls and Splines: Current Trends in Flight Testing and Aerodynamic Model Identification. J. Aircr.
2023, 60, 1480–1502. [CrossRef]
16. Caetano, J.V.; De Visser, C.; De Croon, G.; Remes, B.; De Wagter, C.; Verboom, J.; Mulder, M. Linear aerodynamic model
identification of a flapping wing mav based on flight test data. Int. J. Micro Air Veh. 2013, 5, 273–286. [CrossRef]
Actuators 2024, 13, 130 29 of 33

17. Subedi, S.; Hosseini, B.; Diepolder, J.; Holzapfel, F. Online parameter identification and optimal input design using perturbed
nonlinear programming. J. Phys. Conf. Ser. 2023, 2514, 012020. [CrossRef]
18. Hofmann, R.; Hosseini, S.; Fang, X.; Holzapfel, F. Flight Path Reconstruction for a Coaxial Helicopter Equipped with Rotational
Accelerometers. In Proceedings of the AIAA AVIATION 2023 Forum, San Diego, CA, US, 12–16 June 2023; p. 3292.
19. Larsson, R.; Sobron, A.; Lundström, D.; Enqvist, M. A method for improved flight testing of remotely piloted aircraft using
multisine inputs. Aerospace 2020, 7, 135. [CrossRef]
20. Simmons, B.M.; Gresham, J.L.; Woolsey, C.A. Flight-Test System Identification Techniques and Applications for Small, Low-Cost,
Fixed-Wing Aircraft. J. Aircr. 2023, 60, 1503–1521. [CrossRef]
21. Hofmann, R.; Hosseini, S.; Holzapfel, F. Flight-Test Plan Design and Evaluation in a Closed-Loop Framework for a General
Aviation Aircraft. In Proceedings of the AIAA SCITECH 2022 Forum, San Diego, CA, USA, 3–7 January 2022; p. 2170.
22. Bachfischer, M.; Hosseini, S.; Sax, F.; Rhein, J.; Holzapfel, F.; Maier, L.; Barth, A. Linear Model Identification for a Coaxial
Rotorcraft in Hover. In Proceedings of the AIAA SCITECH 2024 Forum, Orlando, FL, USA, 8–12 January 2024; p. 1719.
23. Hosseini, B.; Sax, F.; Rhein, J.; Holzapfel, F.; Maier, L.; Barth, A.; Hajek, M.; Grebing, B. Global Model Identification for a Coaxial
Helicopter. In Proceedings of the Vertical Flight Society 78th Annual Forum, Fort Worth, TX, USA, 10–12 May 2022.
24. Hosseini, S.; Mbikayi, Z.; Bachfischer, M.; Holzapfel, F.; Rauleder, J. Simulation, Flight Dynamics, and Control Design for a
Coaxial Rotorcraft. In Proceedings of the AIAA Scitech 2024 Forum, Orlando, FL, USA, 8–12 January 2024; p. 1717.
25. Hofmann, R.; Hosseini, S.; Fang, X.; Rhein, J.; Sax, F.; Holzapfel, F.; Maier, L.; Barth, A. Center of Gravity Estimation Using
Multiple Accelerometers. In Proceedings of the AIAA Scitech 2024 Forum, Orlando, FL, USA, 8–12 January 2024; p. 1721.
26. Lichota, P.; Szulczyk, J.; Tischler, M.B.; Berger, T. Frequency responses identification from multi-axis maneuver with simultaneous
multisine inputs. J. Guid. Control Dyn. 2019, 42, 2550–2556. [CrossRef]
27. Berger, T.; Tobias, E.L.; Tischler, M.B.; Juhasz, O. Advances and Modern Applications of Frequency-Domain Aircraft and
Rotorcraft System Identification. J. Aircr. 2023, 60, 1331–1353. [CrossRef]
28. Niermeyer, P.; Raffler, T.; Holzapfel, F. Open-loop quadrotor flight dynamics identification in frequency domain via closed-loop
flight testing. In Proceedings of the AIAA Guidance, Navigation, and Control Conference, Kissimmee, FL, USA, 5–9 January
2015; p. 1539.
29. Steinert, A.; Steffensen, R.; Gierszewski, D.; Speckmaier, M.; Holzapfel, F.; Schmoldt, R.; Demmler, F.; Schell, U.; Ornigg, M.; Koop,
M. Experimental Results of Flight Test Based Gain Tuning. In Proceedings of the AIAA SCITECH 2022 Forum, San Diego, CA,
USA, 3–7 January 2022; p. 2296.
30. Li, H.; Myschik, S.; Holzapfel, F. Null-Space-Excitation-Based Adaptive Control for an Overactuated Hexacopter Model. J. Guid.
Control Dyn. 2023, 46, 483–498. [CrossRef]
31. Hafner, S.F.; Hosseini, S.; Holzapfel, F. Excitation Monitoring for Online Parameter Estimation. In Proceedings of the AIAA
SCITECH 2023 Forum, Gaylord National Harbor, MD, USA, 23–27 January 2023; p. 0039.
32. Akkinapalli, V.S.; Falconí, G.P.; Holzapfel, F. Fault tolerant incremental attitude control using online parameter estimation for a
multicopter system. In Proceedings of the 2017 25th Mediterranean Conference on Control and Automation (MED), Valletta,
Malta, 3–6 July 2017; pp. 454–460.
33. Ignatyev, D.I.; Shin, H.S.; Tsourdos, A. Sparse online Gaussian process adaptation for incremental backstepping flight control.
Aerosp. Sci. Technol. 2023, 136, 108157. [CrossRef]
34. Teubl, D.; Bitenc, T.; Hornung, M. Design and Development of an Actuator Control and Monitoring Unit for Small and Medium Size
Research Uavs; Deutsche Gesellschaft für Luft-und Raumfahrt-Lilienthal-Oberth eV: Bonn, Germany, 2021.
35. Teubl, D.; Oberschwendtner, S.; Hornung, M. Measurement Results of the ACMU System in Various Research UAVs. In
Proceedings of the AIAA SCITECH 2023 Forum, Gaylord National Harbor, MD, USA, 23–27 January 2023; p. 0481.
36. Oberschwendtner, S.; Teubl, D.; Hornung, M. Static Test Procedure for Electromechanical Actuators for UAV Applications. In
Proceedings of the AIAA AVIATION 2022 Forum, Chicago, IL, USA, 27 June–1 July, 2022; p. 3558.
37. Mancinelli, A.; van der Horst, E.; Remes, B.; Smeur, E. Autopilot framework with INDI RPM control, real-time actuator feedback,
and stability control on companion computer through MATLAB generated functions. In Proceedings of the 14th Annual
International Micro Air Vehicle Conference and Competition, Aachen, Germany, 11–15 September 2023; pp. 109–116.
38. Mancinelli, A.; Remes, B.D.; De Croon, G.C.; Smeur, E.J. Real-Time Nonlinear Control Allocation Framework for Vehicles with
Highly Nonlinear Effectors Subject to Saturation. J. Intell. Robot. Syst. 2023, 108, 67. [CrossRef]
39. Osterhuber, R.; Hanel, M.; Hammon, R. Realization of the Eurofighter 2000 primary lateral/directional flight control laws with
differential PI-algorithm. In Proceedings of the AIAA Guidance, Navigation, and Control Conference and Exhibit, Providence,
RI, USA, 16–19 August 2004; p. 4751.
40. Wang, X.; Kong, W.; Zhang, D.; Shen, L. Active disturbance rejection controller for small fixed-wing UAVs with model uncertainty.
In Proceedings of the 2015 IEEE International Conference on Information and Automation, Lijiang, China, 8–10 August 2015;
pp. 2299–2304.
41. Barth, J.M.; Condomines, J.P.; Moschetta, J.M.; Join, C.; Fliess, M. Model-free control approach for fixed-wing UAVs with uncertain
parameters analysis. In Proceedings of the 2018 23rd International Conference on Methods & Models in Automation & Robotics
(MMAR), Miedzyzdroje, Poland, 27–30 August 2018; pp. 527–532.
42. Palframan, M.C.; Fry, J.M.; Farhood, M. Robustness analysis of flight controllers for fixed-wing unmanned aircraft systems using
integral quadratic constraints. IEEE Trans. Control Syst. Technol. 2017, 27, 86–102. [CrossRef]
Actuators 2024, 13, 130 30 of 33

43. Oudin, S. Low Speed Protections for a Commercial Airliner: A Practical Approach. In Proceedings of the AIAA Guidance,
Navigation, and Control Conference, Grapevine, TX, USA, 9–13 January 2017; p. 1023.
44. Liu, C.; Chen, W.H. Disturbance rejection flight control for small fixed-wing unmanned aerial vehicles. J. Guid. Control Dyn. 2016,
39, 2810–2819. [CrossRef]
45. Baldi, S.; Roy, S.; Yang, K.; Liu, D. An underactuated control system design for adaptive autopilot of fixed-wing drones.
IEEE/ASME Trans. Mechatronics 2022, 27, 4045–4056. [CrossRef]
46. Reinhardt, D.P. On Nonlinear and Optimization-based Control of Fixed-Wing Unmanned Aerial Vehicles. Ph.D. Thesis, NTNU,
Trondheim, Norway, 2022.
47. Lavretsky, E. Design, analysis, and flight evaluation of a primary control system with observer-based loop transfer recovery and
direct adaptive augmentation for the calspan variable stability simulator learjet-25b aircraft. In Proceedings of the AIAA Scitech
2019 Forum, San Diego, CA, USA, 7–11 January 2019; p. 1081.
48. Weiser, C.; Ossmann, D.; Kuchar, R.O.; Müller, R.; Milz, D.M.; Looye, G. Flight testing a linear parameter varying control law on a
passenger aircraft. In Proceedings of the AIAA Scitech 2020 Forum, Orlando, FL, USA, 6–10 January 2020; p. 1618.
49. Lombaerts, T.; Chu, Q.; Mulder, J.; Joosten, D. Modular flight control reconfiguration design and simulation. Control Eng. Pract.
2011, 19, 540–554. [CrossRef]
50. Lombaerts, T.; Mulder, J.; Voorsluijs, G.; Decuypere, R. Design of a robust flight control system for a mini-UAV. In Proceedings of
the AIAA Guidance, Navigation, and Control Conference and Exhibit, San Francisco, CA, USA, 15–18 August 2005; p. 6408.
51. Weiser, C.; Schulz, S.; Voss, A.; Ossmann, D. Attitude Control for High Altitude Long Endurance Aircraft Considering Structural
Load Limits. In Proceedings of the AIAA SciTech 2023 Forum, Gaylord National Harbor, MD, USA, 23–27 January 2023; p. 0106.
52. Kim, Y.; Kim, S.; Suk, J. Incremental Nonlinear Dynamic Inversion-based Fault-Tolerant Guidance for UAV. In Proceedings of the
AIAA SCITECH 2024 Forum, Orlando, FL, USA, 8–12 January 2024; p. 2564.
53. Stougie, J.; Pollack, T.; Van Kampen, E.J. Incremental Nonlinear Dynamic Inversion control with Flight Envelope Protection for
the Flying-V. In Proceedings of the AIAA SCITECH 2024 Forum, Orlando, FL, USA, 8–12 January 2024; p. 2565.
54. Pfeifle, O.; Fichter, W. Minimum power control allocation for incremental control of over-actuated transition aircraft. J. Guid.
Control Dyn. 2023, 46, 286–300. [CrossRef]
55. Konatala, R.B.; Van Kampen, E.J.; Looye, G.; Milz, D.; Weiser, C. Flight Testing Reinforcement Learning based Online Adaptive
Flight Control Laws on CS-25 Class Aircraft. In Proceedings of the AIAA SCITECH 2024 Forum, Orlando, FL, USA, 8–12 January
2024; p. 2402.
56. Smeur, E.J.; Bronz, M.; de Croon, G.C. Incremental control and guidance of hybrid aircraft applied to a tailsitter unmanned air
vehicle. J. Guid. Control Dyn. 2020, 43, 274–287. [CrossRef]
57. Athayde, A.; Moutinho, A.; Azinheira, J.R. Experimental Nonlinear and Incremental Control Stabilization of a Tail-Sitter UAV
with Hardware-in-the-Loop Validation. Preprint 2024. [CrossRef]
58. Beyer, Y.; Steen, M.; Hecker, P. Incremental passive fault-tolerant control for quadrotors subjected to complete rotor failures. J.
Guid. Control Dyn. 2023, 46, 2033–2042. [CrossRef]
59. Pollack, T.; Van Kampen, E.J. Robust Stability and Performance Analysis of Incremental Dynamic-Inversion-Based Flight Control
Laws. J. Guid. Control Dyn. 2023, 46, 1785–1798. [CrossRef]
60. Azinheira, J.R.; Moutinho, A.; Carvalho, J. Lateral control of airship with uncertain dynamics using incremental nonlinear
dynamics inversion. IFAC-Pap. 2015, 48, 69–74. [CrossRef]
61. Wang, X.; Sun, S. Incremental fault-tolerant control for a hybrid quad-plane UAV subjected to a complete rotor loss. Aerosp. Sci.
Technol. 2022, 125, 107105. [CrossRef]
62. Cordeiro, R.A.; Azinheira, J.R.; Moutinho, A. Robustness of Incremental Backstepping Flight Controllers: The Boeing 747 Case
Study. IEEE Trans. Aerosp. Electron. Syst. 2021, 57, 3492–3505. [CrossRef]
63. Cordeiro, R.A.; Marton, A.S.; Azinheira, J.R.; Carvalho, J.R.; Moutinho, A. Increased robustness to delay in incremental controllers
using input scaling gain. IEEE Trans. Aerosp. Electron. Syst. 2021, 58, 1199–1210. [CrossRef]
64. Jeong, H.; Jeong, J.; Suk, J.; Kim, S. Angular Acceleration Estimation with Off-CG Accelerometers for Incremental Nonlinear
Dynamic Inversion Control. In Proceedings of the AIAA SCITECH 2024 Forum, Orlando, FL, USA, 8–12 January 2024; p. 2566.
65. Smeur, E.J.; Chu, Q.; De Croon, G.C. Adaptive incremental nonlinear dynamic inversion for attitude control of micro air vehicles.
J. Guid. Control Dyn. 2016, 39, 450–461. [CrossRef]
66. De Ponti, T.; Smeur, E.; Remes, B. Incremental Nonlinear Dynamic Inversion controller for a Variable Skew Quad Plane. In
Proceedings of the 2023 International Conference on Unmanned Aircraft Systems (ICUAS), Warsaw, Poland, 6–9 June 2023;
pp. 241–248.
67. Smeur, E.J.; de Croon, G.C.; Chu, Q. Cascaded incremental nonlinear dynamic inversion for MAV disturbance rejection. Control
Eng. Pract. 2018, 73, 79–90. [CrossRef]
68. Steffensen, R.; Steinert, A.; Smeur, E.J. Nonlinear Dynamic Inversion with Actuator Dynamics: An Incremental Control
Perspective. J. Guid. Control Dyn. 2022, 46, 709–717. [CrossRef]
69. Steffensen, R.; Steinert, A.; Mbikayi, Z.; Raab, S.; Angelov, J.; Holzapfel, F. Filter and sensor delay synchronization in incremental
flight control laws. Aerosp. Syst. 2023, 6, 285–304. [CrossRef]
Actuators 2024, 13, 130 31 of 33

70. Steffensen, R.; Steinert, A.; Holzapfel, F. Longitudinal Incremental Reference Model for Fly-By-Wire Control Law using
Incremental Non-Linear Dynamic Inversion. In Proceedings of the AIAA Scitech 2022 Forum, San Diego, CA, USA, 3–7 January
2022; p. 1230.
71. Lu, P.; Van Kampen, E.J.; De Visser, C.; Chu, Q. Aircraft fault-tolerant trajectory control using incremental nonlinear dynamic
inversion. Control Eng. Pract. 2016, 57, 126–141. [CrossRef]
72. Ye, Z.; Chen, Y.; Cai, P.; Lyu, H.; Gong, Z.; Wu, J. Control Design for Soft Transition for Landing Preparation of Light Compound-
Wing Unmanned Aerial Vehicles Based on Incremental Nonlinear Dynamic Inversion. Appl. Sci. 2023, 13, 12225. [CrossRef]
73. Steinleitner, A.; Frenzel, V.; Pfeifle, O.; Denzel, J.; Fichter, W. Automatic take-off and landing of tailwheel aircraft with incremental
nonlinear dynamic inversion. In Proceedings of the AIAA Scitech 2022 Forum, San Diego, CA, USA, 3–7 January 2022; p. 1228.
74. Myschik, S.; Kinast, L.; Huemer, M.; Vicca, D.; Dollinger, D.; Holzapfel, F. Development of a Flight Control System for a
Cyclocopter UAV Demonstrator. In Proceedings of the AIAA AVIATION 2022 Forum, San Diego, CA, USA, 3–7 January 2022;
p. 3282.
75. Kumtepe, Y.; Pollack, T.; Van Kampen, E.J. Flight control law design using hybrid incremental nonlinear dynamic inversion. In
Proceedings of the AIAA SciTech 2022 Forum, San Diego, CA, USA, 3–7 January 2022; p. 1597.
76. Raab, S.A.; Zhang, J.; Bhardwaj, P.; Holzapfel, F. Consideration of Control Effector Dynamics and Saturations in an Extended
INDI Approach. In Proceedings of the AIAA Aviation 2019 Forum, Dallas, TX, USA, 17–21 June 2019; p. 3267.
77. Bhardwaj, P.; Akkinapalli, V.S.; Zhang, J.; Saboo, S.; Holzapfel, F. Adaptive augmentation of incremental nonlinear dynamic
inversion controller for an extended f-16 model. In Proceedings of the AIAA Scitech 2019 Forum, San Diego, CA, USA, 7–11
January 2019; p. 1923.
78. Sun, S.; Wang, X.; Chu, Q.; de Visser, C. Incremental nonlinear fault-tolerant control of a quadrotor with complete loss of two
opposing rotors. IEEE Trans. Robot. 2020, 37, 116–130. [CrossRef]
79. Wang, X.; Van Kampen, E.J.; Chu, Q.; Lu, P. Stability analysis for incremental nonlinear dynamic inversion control. J. Guid.
Control Dyn. 2019, 42, 1116–1129. [CrossRef]
80. Schildkamp, R.; Chang, J.; Sodja, J.; De Breuker, R.; Wang, X. Incremental Nonlinear Control for Aeroelastic Wing Load Alleviation
and Flutter Suppression. Actuators 2023, 12, 280. [CrossRef]
81. Surmann, D.; Myschik, S. Gain Design for an INDI-based Flight Control Algorithm for a Conceptual Lift-to-Cruise Vehicle. In
Proceedings of the AIAA SCITECH 2024 Forum, Orlando, FL, USA, 8–12 January 2024; p. 1590.
82. Wang, X.; Mkhoyan, T.; De Breuker, R. Nonlinear incremental control for flexible aircraft trajectory tracking and load alleviation.
J. Guid. Control Dyn. 2022, 45, 39–57. [CrossRef]
83. Sun, B.; Mkhoyan, T.; van Kampen, E.J.; De Breuker, R.; Wang, X. Vision-based nonlinear incremental control for a morphing
wing with mechanical imperfections. IEEE Trans. Aerosp. Electron. Syst. 2022, 58, 5506–5518. [CrossRef]
84. Di Francesco, G.; Mattei, M. Modeling and incremental nonlinear dynamic inversion control of a novel unmanned tiltrotor. J.
Aircr. 2016, 53, 73–86. [CrossRef]
85. Cordeiro, R.A.; Azinheira, J.R.; Moutinho, A. Addressing actuation redundancies in incremental controllers for attitude tracking
of fixed-wing aircraft. IFAC-Pap. 2019, 52, 417–422. [CrossRef]
86. Kubica, F.; Livet, T.; Le Tron, X.; Bucharles, A. Parameter Robust Flight Control System for a Flexible Aircraft. In Automatic Control
in Aerospace 1994 (Aerospace Control’94); Elsevier: Amsterdam, The Netherlands, 1995; pp. 41–46.
87. SOBEL, K.M.; SHAPIRO, E.Y.; ANDRY JR, A.N. Eigenstructure assignment. Int. J. Control 1994, 59, 13–37. [CrossRef]
88. Holzapfel, F.; da Costa, O.; Heller, M.; Sachs, G. Development of a lateral-directional flight control system for a new transport
aircraft. In Proceedings of the AIAA Guidance, Navigation, and Control Conference and Exhibit, Keystone, CO, USA, 21–24
August 2006; p. 6222.
89. Comer, A.; Bhandari, R.; Putra, S.H.; Chakraborty, I. Design, Control Law Development, and Flight Testing of a Subscale
Lift-Plus-Cruise Aircraft. In Proceedings of the AIAA SCITECH 2024 Forum, Orlando, FL, USA, 8–12 January 2024; p. 2644.
90. Lovell-Prescod, G.H.; Ma, Z.; Smeur, E.J. Attitude Control of a Tilt-rotor Tailsitter Micro Air Vehicle Using Incremental Control.
In Proceedings of the 2023 International Conference on Unmanned Aircraft Systems (ICUAS), Warsaw, Poland, 6–9 June 2023; pp.
842–849.
91. Bronz, M.; Smeur, E.J.; Garcia de Marina, H.; Hattenberger, G. Development of a fixed-wing mini UAV with transitioning flight
capability. In Proceedings of the 35th AIAA Applied Aerodynamics Conference, Denver, CO, USA, 5–9 June 2017; p. 3739.
92. Gray, A.G.; Gonzalez, F.; Vanegas, F.; Galvez-Serna, J.; Morton, K. Design and Flight Testing of a UAV with a Robotic Arm. In
Proceedings of the 2023 IEEE Aerospace Conference, Big Skye, MT, USA, 4–11 March 2023; pp. 1–13.
93. Soal, K.; Volkmar, R.; Thiem, C.; Sinske, J.; Meddaikar, Y.M.; Govers, Y.; Böswald, M.; Teubl, D.; Bartasevicius, J.; Nagy, M.; et al.
Flight Vibration Testing of the T-FLEX UAV using Online Modal Analysis. In Proceedings of the AIAA SCITECH 2023 Forum,
Gaylord National Harbor, MD, USA, 23–27 January 2023; p. 0373.
94. Kaufmann, E.; Bauersfeld, L.; Loquercio, A.; Müller, M.; Koltun, V.; Scaramuzza, D. Champion-level drone racing using deep
reinforcement learning. Nature 2023, 620, 982–987. [CrossRef] [PubMed]
95. Hansen, S.; Blanke, M. Diagnosis of airspeed measurement faults for unmanned aerial vehicles. IEEE Trans. Aerosp. Electron. Syst.
2014, 50, 224–239. [CrossRef]
96. Dantsker, O.D.; Mancuso, R. Flight data acquisition platform development, integration, and operation on small-to medium-sized
unmanned aircraft. In Proceedings of the AIAA Scitech 2019 Forum, San Diego, CA, USA, 7–11 January 2019; p. 1262.
Actuators 2024, 13, 130 32 of 33

97. Sobrón Rueda, A. On Subscale Flight Testing: Cost-Effective Techniques for Research and Development. Ph.D. Thesis, Linköping
University Electronic Press, Linköping, Sweden, 2021.
98. De Marina, H.G.; Sun, Z.; Bronz, M.; Hattenberger, G. Circular formation control of fixed-wing UAVs with constant speeds. In
Proceedings of the 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Vancouver, BC, Canada,
24–28 September 2017; pp. 5298–5303.
99. Condomines, J.P.; Bronz, M.; Hattenberger, G.; Erdelyi, J.F. Experimental wind field estimation and aircraft identification. In
Proceedings of the IMAV 2015: International Micro Air Vehicles Conference and Flight Competition, Aachen, Germany, 15–18
September 2015.
100. Saldiran, E.; Inalhan, G. Incremental Nonlinear Dynamic Inversion-Based Trajectory Tracking Controller for an Agile Quadrotor:
Design, Analysis, and Flight Tests Results. In Control of Autonomous Aerial Vehicles: Advances in Autopilot Design for Civilian UAVs;
Springer: Berlin/Heidelberg, Germany, 2023; pp. 195–230.
101. Tal, E.; Ryou, G.; Karaman, S. Aerobatic trajectory generation for a vtol fixed-wing aircraft using differential flatness. IEEE Trans.
Robot. 2023, 39, 4805–4819. [CrossRef]
102. Mancinelli, A.; Remes, B.D.; de Croon, G.C.; Smeur, E.J. Unified incremental nonlinear controller for the transition control of a
hybrid dual-axis tilting rotor quad-plane. arXiv 2023, arXiv:2311.09185.
103. De Wagter, C.; Remes, B.; Smeur, E.; van Tienen, F.; Ruijsink, R.; van Hecke, K.; van der Horst, E. The NederDrone: A hybrid lift,
hybrid energy hydrogen UAV. Int. J. Hydrog. Energy 2021, 46, 16003–16018. [CrossRef]
104. Pfeifle, O.; Fichter, W. Cascaded incremental nonlinear dynamic inversion for three-dimensional spline-tracking with wind
compensation. J. Guid. Control Dyn. 2021, 44, 1559–1571. [CrossRef]
105. Koeberle, S.J.; Albert, A.E.; Nagel, L.H.; Hornung, M. Flight Testing for Flight Dynamics Estimation of Medium-Sized UAVs. In
Proceedings of the AIAA Scitech 2021 Forum, Virtual Event, 11–15 January 2021; p. 1526.
106. Matt, J.J.; Chao, H. Efficient Frequency Response Identification for Small Fixed-Wing UAS Using Closed-Loop Flight Data. In
Proceedings of the AIAA SCITECH 2023 Forum, Gaylord National Harbor, MD, USA, 23–27 January 2023; p. 0629.
107. Matt, J.J.; Chao, H.; Shawon, M.H.; Hagerott, S.G. Longitudinal System Identification for a Small Flying-wing UAS. In Proceedings
of the AIAA SCITECH 2023 Forum, Gaylord National Harbor, MD, USA, 23–27 January 2023; p. 0628.
108. Sobron, A.; Lundström, D.; Krus, P. A review of current research in subscale flight testing and analysis of its main practical
challenges. Aerospace 2021, 8, 74. [CrossRef]
109. Babcock, J.T.; Osteroos, R.K.; Tischler, M.B. Open and closed loop system identification of the Stryker 200 UAV. In Proceedings of
the AIAA SCITECH 2022 Forum, San Diego, CA, USA, 3–7 January 2022; p. 2405.
110. Sanders, F.C.; Tischler, M.; Berger, T.; Berrios, M.G.; Gong, A. System identification and multi-objective longitudinal control law
design for a small fixed-wing UAV. In Proceedings of the 2018 AIAA Atmospheric Flight Mechanics Conference, Atlanta, GA,
USA, 25–29 June 2018; p. 0296.
111. Theile, M.; Dantsker, O.; Nai, R.; Caccamo, M.; Yu, S. uavAP: A Modular Autopilot Framework for UAVs. In Proceedings of the
AIAA AVIATION 2020 FORUM, Virtual Event, 15–19 June 2020; p. 3268.
112. Meier, L.; Tanskanen, P.; Heng, L.; Lee, G.H.; Fraundorfer, F.; Pollefeys, M. PIXHAWK: A micro aerial vehicle design for
autonomous flight using onboard computer vision. Auton. Robot 2012, 33, 21–39. [CrossRef]
113. Ebeid, E.; Skriver, M.; Terkildsen, K.H.; Jensen, K.; Schultz, U.P. A survey of open-source UAV flight controllers and flight
simulators. Microprocess. Microsystems 2018, 61, 11–20. [CrossRef]
114. Hattenberger, G.; Bronz, M.; Gorraz, M. Using the paparazzi UAV system for scientific research. In Proceedings of the IMAV
2014, International Micro Air Vehicle Conference and Competition, Delft, The Netherlands, 12–15 August 2014; p. 247.
115. Chao, H.; Cao, Y.; Chen, Y. Autopilots for small unmanned aerial vehicles: A survey. Int. J. Control Autom. Syst. 2010, 8, 36–44.
[CrossRef]
116. Gati, B. Open source autopilot for academic research-the paparazzi system. In Proceedings of the 2013 American Control
Conference, Washington, DC, USA, 17–19 June 2013; pp. 1478–1481.
117. Qvale, M. X-UAV Mini Talon Build Guide. 2016. Available online: [Link]
Mini_Talon_Build_Number_1 (accessed on 26 March 2024).
118. PX4 Development Team, PX4 Version 1.12.3, Available online: [Link] (accessed on 26 March
2024).
119. The MathWorks Inc. MATLAB Version: 9.13.0 (R2022b); The MathWorks Inc.: Natick, MA, USA, 2022.
120. Krause, C.; Göttlicher, C.; Holzapfel, F. Development of a generic Flight Test Maneuver Injection Module. In Proceedings of the
ICAS 31st Congress of the International Council of the Aeronautical Science, Belo Horizonte, Brazil, 9–14 September 2018.
121. Nelson, R.C. Flight Stability and Automatic Control; WCB/McGraw Hill: New York, NY, USA, 1998; Volume 2.
122. Myschik, S.; Holzapfel, F.; Sachs, G. Low-cost sensor based integrated airdata and navigation system for general aviation aircraft.
In Proceedings of the AIAA Guidance, Navigation and Control Conference and Exhibit, Honolulu, HI, USA, 18–21 August 2008;
p. 7423.
Actuators 2024, 13, 130 33 of 33

123. Schatz, S.P.; Gabrys, A.C.; Gierszewski, D.M.; Holzapfel, F. Inner loop command interface in a modular flight control architecture
for trajectory flights of general aviation aircraft. In Proceedings of the 2018 5th International Conference on Control, Decision and
Information Technologies (CoDIT), Thessaloniki, Greece, 10–13 April 2018; pp. 86–91.
124. Stevens, B.L.; Lewis, F.L.; Johnson, E.N. Aircraft Control and Simulation: Dynamics, Controls Design, and Autonomous Systems; John
Wiley & Sons: Hoboken, NJ, USA, 2015.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.
A.4 Paper 4

A.4 Paper 4
Paper 4, reference [72], is re-printed under the Creative Commons Attribution 4.0 Interna-
tional License, [Link] with no changes
from the original.

LXIV
Aerospace Systems (2023) 6:285–304
[Link]

ORIGINAL PAPER

Filter and sensor delay synchronization in incremental flight control

laws
Rasmus Steffensen1 · Agnes Steinert1 · Zoe Mbikayi1 · Stefan Raab1 · Jorg Angelov1 · Florian Holzapfel1

Received: 30 August 2022 / Revised: 23 November 2022 / Accepted: 4 December 2022 / Published online: 19 January 2023
© The Author(s) 2023, corrected publication 2023

Abstract
This paper introduces and investigates several methods of filtering and delay synchronization in incremental control law
structures. It is shown that proper filter and delay synchronization is of great importance for both the performance and
stability of the control laws. It is derived how to synthesize complementary filters in case of sensor dynamics, sensor delays
and internal control law filtering such as roll-off filters and notch filters. The complementary filters are compared to common
synchronization approaches previously applied in incremental control laws. It is shown that the proposed methods are beneficial
for Multiple Input Multiple Output (MIMO) systems in case that different delays and filtering is present in the different
measurement signals. In case of strongly coupled input effectiveness, the synchronization on the actuator feedback may
become even unstable, where the complementary filter approach recovers the design closed-loop behavior. In addition, it is
derived how to initialize the complementary filters for a transient-free engagement in non-steady-state conditions. The results
are compared in simulation with a roll rate control law for a fixed wing aircraft.

Keywords Incremental control laws · Complementary filters · Sensor dynamics · Sensor delays

1 Introduction plant to be controlled, INDI simply uses measurements of

derivatives of the output and the input effectiveness. This
In the past decade, the control concept Incremental Non- is because the knowledge of the state dependent dynam-
linear Dynamic Inversion (INDI) gained a lot of interest, ics can be approximated by an appropriate measurement.
especially in the domain of flight control. While control tech- INDI can, therefore, be interpreted as a sensor-based control
niques such as Non-linear Dynamic Inversion (NDI) [1–4] method. This reduced model-dependency has been proven to
and Backstepping (BS) [5] rely on accurate models of the be advantageous when controlling aerial vehicles because of
multiple reasons:
Rasmus Steffensen and Agnes Steinert have contributed equally to this
work.
1. Novel aerial vehicle configurations might be complex to
B Rasmus Steffensen model or the models might be inaccurate or lacking the
[Link]@[Link]
influence of dominating physical effects.
Agnes Steinert 2. Aircraft models depend on aerodynamic parameters,
[Link]@[Link]
which are expensive to identify, e.g., by performing
Zoe Mbikayi extensive flight test campaigns;
[Link]@[Link]
3. Changes in the vehicle design require a repetition of the
Stefan Raab identification process, and therefore the flight test cam-
[Link]@[Link]
paigns;
Jorg Angelov 4. Robustness against a wide range of faults or structural
[Link]@[Link]
damages;
Florian Holzapfel 5. Robustness against disturbances;
[Link]@[Link]
6. Suitability for automatic take-off and landing, especially
1 Institute of Flight System Dynamics, Technical University of for ultra-light aerial vehicles, where ground effects have
Munich, Boltzmannstrasse 15, Garching 85748, Germany a large impact and are difficult to model;

123
286 Aerospace Systems (2023) 6:285–304

7. Unified control strategies without the necessity of switch- mended because of its sensitivity to sensor noise leading
ing or blending between different control laws for differ- to an amplification of noise on the control signal. In [24],
ent flight phases. the angular accelerations were estimated with this technique,
using backward finite differences, and a first-order low-pass
Incremental non-linear flight control techniques, such as filter was introduced before the actuators to attenuate high-
INDI, have been successfully demonstrated on several aerial frequency oscillations caused by numerical noise. With the
platforms and for different applications: the Cessna Citation sensor dynamics and measurement delays present, a low-
II [6,7], VTOL transition UAVs with movable tilts [8–10], amplitude high-frequency oscillation still remained in the
quadrotor UAVs [10–12], hexacopter [14], a hybrid tail-sitter control commands. [28] proposed to filter the measurement,
UAV [15,16], the automatic take-off and landing of a fixed using a second-order low-pass filter, in combination with a
wing tail-wheel aircraft [17], a helicopter [18], an airship discrete differentiation. This allows to estimate the deriva-
[19], a piloted all attitude fixed wing turboprop demonstra- tive for low frequencies, while attenuating high-frequency
tor simulated aircraft [20], and a commercial civil aircraft sensor noise. Since this kind of second-order filter differen-
simulation model [21,22]. tiator introduces lag and produces a delayed estimate of the
It is commonly known that one major challenge of INDI relevant output derivative which degrades the controller per-
is its sensitivity towards measurement delays in the output formance and stability, several techniques were proposed to
derivative feedback [23–26] and actuator command delay accommodate this issue. [29] proposes the use of an input
[EINDI]. In van’t Veld et al. [27], the authors show that the scaling gain to increase the controller robustness to delays
stability of INDI controllers subject to measurement time in the feedback loop, when using the second-order filter dif-
delay is an issue, especially, when the transmission delay of ferentiator. Other approaches are based on the importance
the actuator measurements and that of the relevant output of delaying the actuator and state derivative measurements
derivative measurement are not equal. Delays are usually equally, which has been demonstrated in van’t Veld et al.
encountered for several reasons: [27]. [6,11,27] synchronize the actuator measurement in time
with respect to the angular acceleration by a synchronization
1. Control signal transmission delay; filter. This synchronization corresponds to the second-order
2. Sensor signal delays; low-pass filter and a time delay, corresponding to the rela-
3. Processing and scheduling delays. tive delay between the angular rates and actuator deflection
measurements. Besides the second-order filter differentiator,
In van’t Veld et al. [27], it is observed that INDI is more complementary filters have been proposed for estimating the
sensitive to output derivative measurement delay than actu- angular accelerations, where the measured rate is low-pass
ator delay. Additionally, signal filtering can introduce lag to filtered and differentiated and the high-frequency part of the
the system which might be an issue for the controller perfor- rate derivative is obtained using a high-pass filtered model-
mance and stability as demonstrated in van’t Veld et al. [27] based rate acceleration estimate [14,30]. The latter further
as well. The use of signal filtering in control laws is of high proposed a synchronization scheme on the actuator feed-
importance, for performance, stability, safety and robustness. back in combination with the complementary filter. In [23],
There are several reasons for introducing filtering of the used a linear predictive filter is designed to predict the angular
signals. The most common reasons are: rate derivatives from angular rates and references thereof.
The coefficients of the predictive filter are computed using
1. Anti aliasing filtering;
least squares estimation [31] designs a nondelay differen-
2. Roll-off filtering for noise attenuation;
tiator based on a precise-delay differentiator, to acquire the
3. Notch filtering for attenuation of structural mode excita-
output derivatives without delay.
tion;
In this paper, it is investigated how to compensate for filter
4. Filtering for estimation of certain variables, e.g., comple-
and sensor dynamics in the implementation of different types
mentary filtering;
of incremental control laws. We start with the Non-Linear
5. Filtering for bias removal.
Dynamic Inversion (NDI) control law through the actuators
Another common issue related to INDI for aircraft control introduced as Actuator-NDI (ANDI) in Steffensen et al. [32].
applications is that measurements of the angular accelera- From this control law, several other derived control laws
tions are often required. Although sensors for measuring can be obtained by applying appropriate assumptions. The
the angular rate derivative exist, they are not part of the most basic form is Incremental Non-linear Dynamic Inver-
standard equipment of aerial vehicles. Hence, the rate deriva- sion (INDI) [28]. This control law does not explicitly take into
tive needs to be estimated in most applications. Obtaining account the actuator dynamics. The Extended INDI (E-INDI)
the rate derivative by calculating the discrete derivative of [9] takes the actuators into account, but neglects the state
the rate measurement using a differentiator, is not recom- dependent terms. The filter synchronization is derived ana-

123
Aerospace Systems (2023) 6:285–304 287

lytically for the ANDI control law, using a complementary paper further provides insights into the relations between
filter exploiting relevant model information to compensate these different approaches. Additionally, the cascaded com-
for the negative effects of the filters and sensors. Analysis is plementary filter is demonstrated for an ANDI control law,
made for the initialization of the filter, to investigate methods introduced in Steffensen et al. [32], and compared to the E-
of transient-free activation of the control law. An analytical INDI law [9].
comparison is made of synchronization techniques used in
literature, and a re-formulation of these synchronization tech-
niques is performed in terms of equivalent complementary 2 Control laws
filters with several benefits for MIMO systems when the fil-
ters applied to the measurements differ. The following section provides a derivation of the ANDI
The contribution of this paper consists of: (Actuators in Non-Linear Dynamics Inversion), E-INDI
(Extended Incremental Non-linear Dynamic Inversion) and
• Formulation of a complementary filter structure that INDI (Incremental Non-linear Dynamic Inversion) control
can be used for ANDI, E-INDI and INDI control laws. laws, as a basis for the investigation in Sects. 4 and 5.
The benefits are 1) an inherent synchronization which
increases the robustness with regard to phase delays 2) 2.1 Summary of ANDI control law
It allows to apply different filtering on the different mea-
surement signals. For example to attenuate structural We consider the ANDI control law introduced in Steffensen
modes in the various axes of the aircraft, different notch et al. [32], as depicted in Fig. 1. The ANDI is based on NDI,
filters might be required on the different measurement hence includes the stability guarantees in terms of Lyapunov
signals. and exact inversion for tracking as well as specified error
• Formulation of a proper initialization method for these dynamics. With the assumption of first-order actuators from
complementary filters as used in incremental control ANDI, the approximate concepts INDI and E-INDI can be
laws, for transient-free engagement in non-steady-state deduced. Hence, the ANDI is taken as a starting point, and
conditions. the concepts of synchronization will be extended directly to
• Derivation of alternative formulations of popular syn- INDI and E-INDI. As was shown in Steffensen et al. [32], the
chronization techniques in the actuator feedback path as ANDI control law corresponds to an INDI control law if the
modified complementary filters. These alternative formu- bandwidth of the actuators tends to infinity, and corresponds
lations have the benefit that they can recover the design to the E-INDI [9] if the Fx ẋ term is neglected as explained
closed-loop transfer function structure in MIMO systems later in this section. The ANDI control law allows to actively
when different sensor filtering in the various feedback account for the actuator and state dependent dynamics and is
channels are used. It is shown that the synchronization in summarized in the following. Consider the system
the feedback can become even unstable in case of strongly
coupled input effectiveness, where the equivalent com- ẋ = f (x, u),
plementary filter exactly recovers the design closed-loop y = h(x), (1)
behavior.
• Application of appropriate filters on the output of the with system state x ∈ Rn , actuator state u ∈ Rk , output y ∈
reference model based on the filtering on the measured Rm , f ∈ C (r +1) (Rn × Rk ; Rn ) and h ∈ C (r +1) (Rn ; Rm ).
signals for use in the error controller feedback. Note, that the time dependence (t) was here and in the fol-
• Derivation of an un-delayed state estimate using a cas- lowing omitted for better readability. The actuator dynamics
caded complementary filter taking into account signal are given by:
filtering and delays. This un-delayed state estimate can
be used to obtain an un-delayed ẏmdl . This cascaded filter u̇ =  (u c − u) , (2)
additionally provides an ẋ estimate that can be used in
control laws based on ANDI. with diagonal matrix  ∈ R(k×k) with its entries representing
the bandwidth of the different actuators. In Steffensen et al.
Based on a general linear MIMO system, the closed-loop [32] the ANDI control law is derived assuming the system to
dynamics is analyzed analytically for (1) vanilla (perfect have a relative degree of r ∈ N, i.e., the r th derivative of y
knowledge of all variables) INDI (2) INDI with complemen- with respect to time is the first derivative of y which explicitly
tary filter and ẏmdl (3) filter differentiator approach without depends on u. For the sake of simplicity, we assume here
synchronization (4) filter differentiator approach with syn- r = 1:
chronization (5) Hybrid INDI, i.e., complementary filter
approach with synchronization. The analysis done in this ẏ = F(x, u), (3)

123
288 Aerospace Systems (2023) 6:285–304

with F ∈ C 1 (Rn × Rk ; Rm ). The second derivative of y is In addition, ẋ needs to be estimated, which is covered in
then Sect. 4.6. Note, that in some cases ẋˆ is the same as ẏ,
ˆ and the
same complementary filtered signal can be used.
ÿ = Fx ẋ + Fu u̇, (4)
2.3 ANDI and INDI
∂ F(x,u) ∂ F(x,u)
where Fx = ∂x and Fu = ∂u . Substituting Eq. (2)
into (4) results in To arrive at the INDI control law from the ANDI control
law, a special choice of error dynamics have to be made as
ÿ = Fx ẋ + Fu  (u c − u) , (5) shown in Steffensen et al. [32]. The error dynamics shall be
chosen as a product of actuator and system error dynamics
as shown in Eq. (10). The actuator loop of the ANDI error
which is solved for u c , assuming Fu  to have full row rank
dynamics, i.e., the second part of Eq. (10), shall be chosen
and choosing ÿ = ν as the virtual command:
with the same bandwidth as the actuators, ω. The assumption
necessary on the system is that all actuators need to have the
u c = (Fu )† (ν − Fx ẋ) + u, (6) same bandwidth, that is  = ωIk×k .

where (Fu )† denotes a right inverse matrix that solves the E y (s) (s + K 0 ) (s I + ωI ) = 0, (10)
linear equation system given in Eq. (5). The virtual control
input ν is designed using a linear error controller as with the errors defined as:

ν = ÿref + k1 ( ẏref − ẏ) + k0 (yref − y). (7) e y = yr e f − y,

(11)
ė y = ẏr e f − ẏ.
By applying the control law to the system, the desired error
dynamics are obtained as designed, because Fu  has full In the time domain, this can be expressed as:
row rank such that (Fu )(Fu )† = Im×m :
ë y + k1 ė y + k0 e y = ë y + K 0 ė y + ω(ė y + K 0 e y ) = 0. (12)
 
ÿ = Fx ẋ + Fu  (Fu )† (ν − Fx ẋ) + u − u Then, the pseudo-control ν is chosen as:
(8)
= ÿref + k1 ( ẏref − ẏ) + k0 (yref − y).
ν = ÿr e f + K 0 ė y + ω(ė y + K 0 e y ), (13)
The block diagram of the ANDI controller is depicted in Fig.
and by inserting ν into Eq. (6), the ANDI control law can be
1.
expressed as:
2.2 ANDI control law with measured output 1  
uc = (Fu )† ÿr e f + K 0 ė y − Fx ẋ
ω (14)
It is often not possible to directly measure the time derivative  
+ (Fu )† ė y + K 0 e y + u,
ẏ, which is needed to calculate the pseudo-control ν in Eq.
(7), which is then used in the ANDI control law in Eq. (6). which was first shown in Steffensen et al. [32]. If the actuators
We propose to estimate ẏ using the complementary filter have a high bandwidth ω, the first term of Eq. (14) can be
which will be introduced in Sect. 3.3, that uses a model-based neglected, and the remaining part is equivalent to the INDI
estimate of ẏ and the measured and filtered output signal y f control law:
as depicted in Fig. 2. In addition, due to the filtering of y, it
 
is proposed to filter the reference signal yr e f with the same u c = (Fu )† ė y + K 0 e y + u. (15)
filtering as the measurement to synchronize the error term.
This filtering includes the filtering by F, which is any filtering Similar to Sect. 2.2, the INDI law can be implemented as
applied to the signal, e.g., roll-off filters or notch filters, the depicted in Fig. 3 using a complementary filtered estimate of
measurement delay D y , as well as the sensor dynamics S. ẏ and synchronizing yr e f and y f .
The control law is, hence, modified as such: It shall be noted that sometimes the first term in Eq. (14)
is included using variables from a reference model [20,33],
u c = (Fu )† (ν − Fx ẋ) + u such that the command path exactly corresponds to the ANDI
 control law. If the system is (sufficiently) linear, this will
ˆ
= (Fu )† ÿr e f + k1 ( ẏr e f − ẏ) (9)
only affect the feed-forward part of the system, and hence

+ k0 (yr e f , f − y f ) − Fx ẋ + u. not change the closed-loop poles.

123
Aerospace Systems (2023) 6:285–304 289

Fig. 1 ANDI with known

outputs and state and output
derivatives

Fig. 2 Block diagram of ANDI

with synchronization

Fig. 3 Block diagram of INDI

with synchronization

2.4 ANDI and E-INDI 3 Complementary filters and initialization

In literature, several extensions to the vanilla INDI control In the following section, the complementary filter design and
law can be found. Two approaches that are taking explicitly initialization is presented, and a formal initialization method
the actuator dynamics into account are [9,11]. While [11] is derived for transient-free engagement in non-steady-state
assumes that all actuators have the same bandwidth, i.e.,  = conditions. As will be shown later, this design of the comple-
Ik×k ω, [9] considers the more general case of actuators with mentary filters will inherently contain the synchronization
different bandwidth. As was shown in Steffensen et al. [32], properties that have earlier been shown to be crucial for
choosing the error dynamics given in Eq. (12), the respective increasing performance and stability of INDI control laws.
control laws from [9,11], can be re-formulated as The use of this type of complementary filter design, allows
for the use of different filters in the individual measurement
   channels, which is not directly achievable using common
u c = (Fu )† ÿref + K 0 ė y + y ė y + K 0 e y +u, (16) synchronization techniques on the actuator feedback.

which in case of  = Ik×k ω,  y = Im×m ω, can be expressed

as 3.1 Basic complementary filter concept and
initialization
1   
uc = (Fu )† ÿr e f + K 0 ė y + ω ė y + K 0 e y + u. (17) Figure 4 depicts a simple complementary filter, which will
ω
be used as an introductory example for the derivation of the
initialization in non-steady-state conditions.
This control law equals the ANDI control law in Equation First consider the zero-valued steady-state case. Then, the
(14), if Fx ẋ is neglected. relation between the inputs and outputs of the filter can be

123
290 Aerospace Systems (2023) 6:285–304

Fig. 4 Block diagram of complementary filter Fig. 6 Block diagram of complementary filter with sensor dynamics

In the Laplace domain, accounting for non-zero initial con-

ditions, this results in

Ẏ (s) = s X 1 (s) − x1 (0) + U2 (s) − X 2 (s). (22)

Substituting Eq. (20) into (22) results in

 
ω 1
Ẏ (s) = s U1 (s) + x1 (0) − x1 (0) + U2 (s)
s+ω s+ω
 
ω 1
− U2 (s) + x2 (0)
s+ω s+ω
sω ω
Fig. 5 Block diagram of complementary filter with reset = U1 (s) − x1 (0)
s+ω s+ω
ω 1
+ U2 (s) − U2 (s) − x2 (0).
derived as follows: s+ω s+ω
(23)
Ẏ (s) = (1 − H (s) + H (s))Ẏ (s)
(18) Substituting U1 (s) = Y (s) and U2 (s) = Ẏ (s) = sY (s) −
= (1 − H (s)) Ẏ (s) + H (s)sY (s),
y(0), results in

where U1 (s) = Y (s), U2 (s) = Ẏ (s) and assuming that y(0) sω ω

and ẏ(0) are zero. In case that the model ẏ is accurate, the Ẏ (s) = Y (s) − x1 (0)
s+ω s+ω
filter perfectly distributes the high-frequency part of the ẏ ω 1
estimate to the model-based ẏ and the low-frequency part to + U2 (s) − (sY (s) − y(0)) − x2 (0)
s+ω s+ω
the differentiated y measurement. ω ω 1
In case of non-steady-state initial conditions, the filter needs =− x1 (0) + U2 (s) + y(0) − x2 (0).
s+ω s+ω s+ω
to be carefully initialized. In the following, a method to obtain (24)
the correct initialization is derived. For now, consider H (s) as
a first-order filter such that the differential equations describ- Initializing the integrators with x1 (0) = y(0) = u 1 (0) and
ing the x1 and x2 dynamics, as depicted in Fig. 5, are given x2 (0) = 0 as depicted in Fig. 5 by the red dotted lines, results
by: in

ẋ1/2 = ω(u 1/2 − x1/2 ), (19) Ẏ (s) = U2 (s) = Ẏ (s). (25)

where u 1 = y and u 2 = ẏ. In the Laplace domain, accounting Hence, in case of an accurate model of ẏ, the output of the
also for non-zero initial conditions, the following relation is filter will exactly match the true ẏ, also at the initial time
obtained: instance when engaging the control law.

ω 1 3.2 Complementary filter and sensor dynamics

X 1/2 (s) = U1/2 (s) + x1/2 (0). (20)
s+ω s+ω
Usually, the output y of the plant is measured by a sensor
The complementary filter calculates ẏ as shown in Fig. 5 by with corresponding sensor dynamics S(s). These dynamics
may include an anti-aliasing filter because of the (down) sam-
ẏ = ẋ1 + u 2 − x2 . (21) pling of the sensor signal for the flight control computer. For

123
Aerospace Systems (2023) 6:285–304 291

now consider these dynamics to be described by a first-order The estimate of ẏ is calculated by the complementary filter
filter such that the measured plant output that is used in the as depicted in Fig. 7 by
complementary filter corresponds to

Ys (s) = S(s)Y (s). (26) ẏ = ẋ1 + ũ 2 − x2 , (31)

The complementary filter in Fig. 6 is given by which results in the Laplace domain to

Ẏ (s) = (1 − H (s)S(s)) Ẏmdl (s) + s H (s)Ys (s). (27)

Ẏ (s) = s X 1 (s) − x1 (0) + Ũ2 (s) − X 2 (s). (32)
To show the proper synchronization, now assume that ẏmdl =
ẏ. The complementary filter for zero initial conditions results
For X 1 (s) and X 2 (s), the relation in Eq. (20) is inserted. The
in:
input U1 (s) and U2 (s) are the states of the sensor filter and
sensor filter model as depicted in Fig. 7, i.e., U1 (s) = X̃ 1 (s),
Ẏ (s) = (1 − H (s)S(s)) Ẏ (s) + s H (s)Ys (s). (28) U2 (s) = X̃ 2 (s):

Hence, with an accurate model of ẏ, the filter again exactly  

distributes the high-frequency part of the ẏ estimate to the ω 1
Ẏ (s) = s X̃ 1 (s) + x1 (0) − x1 (0) + Ũ2 (s)
model-based ẏ and the low-frequency part to the differenti- s+ω s+ω
 
ated y in the presence of the sensor dynamics S(s). ω 1
− X̃ 2 (s) + x2 (0) . (33)
To determine the initial values of the integrator state in the s+ω s+ω
sensor filter model used in the complementary filter, we start
with considering the differential equations of the sensor filter
and the sensor filter model, as depicted in Fig. 7, given by Substituting X̃ 1 and X̃ 2 in Eq. (33) by relation (30) results in

x̃˙1/2 = ω̃(ũ 1/2 − x̃1/2 ), (29) 

sω ω̃ 1 s
Ẏ (s) = Ũ1 (s) + x̃1 (0) + x1 (0)
s + ω s + ω̃ s + ω̃ s+ω
where ũ 1 = y and ũ 2 = ẏ. The resulting relation in the
− x1 (0) + Ũ2 (s)
Laplace domain is 
ω ω̃ 1 1
− Ũ2 (s) + x̃2 (0) − x2 (0).
ω̃ 1 s + ω s + ω̃ s + ω̃ s+ω
X̃ 1/2 (s) = Ũ1/2 (s) + x̃1/2 (0). (30) (34)
s + ω̃ s + ω̃

Fig. 7 Block diagram of complementary filter with reset

123
292 Aerospace Systems (2023) 6:285–304

Using that Ũ1 (s) = Y (s) and Ũ2 (s) = Ẏ (s) = sY (s) − y(0)
results in

sω ω̃ 1 s
Ẏ (s) = Y (s) + x̃1 (0) + x1 (0)
s + ω s + ω̃ s + ω̃ s+ω
− x1 (0) + Ẏ (s)

ω ω̃ 1
− (sY (s) − y(0)) + x̃2 (0)
s + ω s + ω̃ s + ω̃
1
− x2 (0)
s+ω Fig. 8 Block diagram of complementary filter with sensor dynamics
sω s and filtering
= x̃1 (0) + x1 (0)
(s + ω)(s + ω̃) (s + ω)
s+ω signal used by the complementary filter corresponds to
− x1 (0) + Ẏ (s)
(s + ω)
 Y f (s) = F(s)D y (s)S(s)Y (s), (37)
ω ω̃ 1 1
− − y(0) + x̃2 (0) − x2 (0)
s+ω s + ω̃ s + ω̃ s+ω
sω where F(s) is the transfer function of the roll-off or notch
= x̃1 (0) filters, D(s) is the transmission delay of the measurement
(s + ω)(s + ω̃)
ω(s + ω̃) and S(s) is the sensor dynamics. The complementary filter
− x1 (0) + Ẏ (s) needs then to be updated as shown in Fig. 8, such that
(s + ω)(s + ω̃)
ωω̃ ω  
+ y(0) − x̃2 (0) Ẏ (s) = 1 − H (s)F(s)D y (s)S(s) Ẏ (s)
(s + ω)(s + ω̃) (s + ω)(s + ω̃)
−
1
x2 (0). + H (s)F(s)D y (s)S(s)Ẏ (s)
s+ω  
= 1 − H (s)F(s)D y (s)S(s) Ẏ (s) + s H (s)Y f (s).
(35)
(38)
Initializing the integrators as discussed before with with
x1 (0) = u 1 (0) = x̃1 (0) and x2 (0) = 0 as depicted in Fig. 7 The initialization of the filters can be derived in a similar man-
by the red dotted lines, results in ner as in the previous sections. In the following, we denote
the series of the filters as
ωω̃
Ẏ (s) = (y(0) − x̃1 (0)) + Ẏ (s) Fcy (s) = F(s)D y (s)S(s). (39)
(s + ω)(s + ω̃) (36)
ω
− x̃2 (0) The complementary filtered estimate of the derivative of the
(s + ω)(s + ω̃)
output signal y is, hence, given by
Hence, the sensor model state needs to be initialized to
x̃2 (0) = ω̃ (y(0) − x̃1 (0)) which corresponds to the time Ẏˆ (s) = (1 − H (s)Fcy (s))Ẏmdl (s) + s H (s)Y f (s), (40)
derivative x̃˙1 (0) of the sensor filter state as depicted in Fig.
7 by the red dotted lines. where Ẏmdl (s) is the model-based estimate of Ẏ .
Note that in case the sensor state x̃1 is initialized to y(0),
x̃˙1 (0) will be 0. This will be the case if the sensor is reset,
and then the integrator state x̃2 of the sensor model could be 4 Synchronization techniques with related
initialized with 0. Otherwise, alternatively as an approxima- complementary filters
tion of x̃˙1 the model based ẏ could be used as initialization
value for x̃2 . In the following, the proposed synchronization is compared
to other synchronization approaches in literature. For the
3.3 Complementary filter, sensor dynamics and analysis, we consider a MIMO linear time invariant system:
additional filtering
ẋ(t) = Ax(t) + Bu(t) + Bd d(t),
(41)
Often the measured output ys of the plant is additionally y (r −1) (t) = C x(t),
filtered to attenuate noise, vibrations from engines, pickup
of structural modes, etc. Commonly roll-off filters or notch where r is the relative degree of the system, A is the system
filters are used for this purpose. Furthermore, transmission matrix, B is the input effectiveness, Bd d(t) is the disturbance
delays might be present on the measurements such that the term, C the output matrix, and with the restriction that:

123
Aerospace Systems (2023) 6:285–304 293

Table 1 Effect of synchronization on C AX in closed-loop

Synchronization CAX contribution
 
Vanilla INDI I − Ḡ A
 
INDI with complementary filter and ẏmdl I − Ḡ A
 
INDI with derivative filter and synchronization I − Ḡ A H Fcy
Fig. 9 Vanilla INDI control law
Hybrid INDI (I − Ḡ A Fcy )

 −1
C B(I − G A (s))−1 G A (s)(C B)† = I − Ḡ A (s) Ḡ A (s),
(42)

where G A (s) are the actuator dynamics, Ḡ A (s) is a diagonal

transfer function matrix, and the matrix C B has full row rank.
Fig. 10 INDI with complementary filter
Equation (42) is true if either:

Basic assumptions (either has to hold): 4.1 Vanilla INDI

1. System is SISO, then Ḡ A = G A . The vanilla INDI with perfect knowledge of the state deriva-
2. C B matrix and actuator transfer function matrix G A are tive is depicted in Fig. 9. The closed-loop transfer function
square, only have elements on the diagonal, and they have is given by:
the same dimension, such that Ḡ A = G A .    
3. Actuator transfer function matrix G A only has elements sY = I − Ḡ A C AX + Ḡ A V + I − Ḡ A C Bd D, (43)
on the diagonal and all the elements are the same, i.e.,
the actuators operate independently and have the same as shown in Appendix A.1. It is seen that ẏ follows ν with the
dynamics. Then, Ḡ A is a diagonal transfer function actuator dynamics, plus the state dynamics and disturbances
matrix with elements equal to the elements in G A but filtered through a high-pass filter given by I − Ḡ A , i.e., if the
the matrix have a different dimension. actuator dynamics are much faster than the state dynamics
and disturbances, the influence is filtered out.
For the following, we consider without loss of generality a
4.2 INDI with complementary filter and ẏmdl
system with relative degree 1, which is common in baseline
flight control laws. In addition it improves the readabil-
If ẏ is estimated with a complementary filter as derived in
ity. Substituting sY (s) with s r Y (s) for the following results
Sect. 3.3, and depicted in Fig. 10, assuming that there are no
recover the general case. Note that for better readability the
model uncertainties or noise, then ẏmdl = ẏ holds, and the
arguments (s) and (t) are omitted.
closed-loop transfer function is given by:
In the following, it is shown that the different synchro-
nization techniques only vary in their influence on the state    
dependent term C AX of the closed-loop dynamics, summa- sY = I − Ḡ A C AX + Ḡ A V + I − Ḡ A H Fcy C Bd D
rized in Table 1. The influence on the direct disturbance term (44)
C Bd D and pseudo-control term V is exactly the same for all
synchronization techniques. as shown in Appendix A.2, which shows that the complemen-
It is additionally shown that for methods 3 and 4 as pro- tary filter recovers the input–output dynamics of the vanilla
posed in [6,11,27] and [30], for MIMO systems, additional INDI.
 It is seen that the disturbance is high-pass filtered by
restrictive assumptions, e.g., same filtering of all measure- I − Ḡ A H Fcy , such that disturbances below the cut-off of
ments, are necessary to obtain the relations given in Table 1. the combined actuator and filter dynamics are rejected.
The respective assumptions will be detailed in the related sec- Note, ẏmdl = ẏ might seem like a crude assumption but is
tions. A re-arrangement of the synchronization is proposed, made to highlight the synchronization effects i.e., ẏmdl gives
such that the design closed-loop transfer function can be an un-delayed estimate of ẏ. For example ẏmdl = C(A x̂ +
recovered for MIMO systems, relaxing the above-mentioned Bu) would be such an estimate, if x̂ is obtained as an un-
assumptions. This re-arrangement allows different filtering delayed estimate, for example from a model or using again
in the measurement feedback channels. a complementary filter as depicted in Fig. 16.

123
294 Aerospace Systems (2023) 6:285–304

Fig. 11 INDI with derivative filter

Fig. 13 Re-formulation of INDI with derivative filter and synchroniza-

tion

   
sY = I − Ḡ A H Fcy C AX +Ḡ A V+ I − Ḡ A H Fcy C Bd D,
(46)

Fig. 12 INDI with derivative filter and synchronization as used in [6, assuming that:
11,27]
 −1  −1
C B I − G A H̄ F̄cy G A (C B)† = I − Ḡ A H Fcy Ḡ A ,
Note that ẏmdl will not be ẏ if there are model uncertain- (47)
ties (e.g., in A, B or C) or measurement noise. The cut-off
frequency of the filter H in the complementary filter has to which hold if either of these additional assumptions are sat-
be chosen as a compromise, in general as high as possible isfied, respectively:
but low enough such that the noise is not amplified by the
derivative. Respective additions to the basic assumptions:

1. No additional assumption if system is SISO.

4.3 INDI with derivative filter and no
2. H̄ and F̄cy are square, only have elements on the diagonal
synchronization
and both have the same dimension as H and Fcy , such
that H̄ F̄cy = H Fcy .
If ẏ is estimated with a filter from a sensor measurement as
3. H and Fcy only has elements on the diagonal and all the
depicted in Fig. 11, then the closed-loop transfer function is
elements of the respective matrix are the same, i.e., each
given by:
measurement channel is filtered separately and equally.
 −1   Then, H̄ F̄cy is a diagonal transfer function matrix with
sY = I − Ḡ A + Ḡ A H Fcy I − Ḡ A C AX elements equal to the elements in H Fcy but the matrix
 −1 has a different dimension.
+ I − Ḡ A + Ḡ A H Fcy Ḡ A V (45)
 −1  
+ I − Ḡ A + Ḡ A H Fcy I − Ḡ A C Bd D, It can be seen that this filtering technique results in a closed-
loop transfer behavior which is similar to the result of the
as shown in Appendix A.3. It can be seen that this filter- vanilla INDI, i.e., ẏ follows ν with the actuator dynam-
ing technique does not correspond to the dynamics of the ics, plus the state dynamics filtered through a high-pass
vanilla INDI, i.e., INDI with perfect state derivative knowl- filter.
 Whilefor vanilla INDI this high-pass filter is given
edge. It is shown in Section 5 by simulation that having no by I − Ḡ A , here the high pass results in (I − Ḡ A H Fcy ),
synchronization can very easily lead to unstable closed-loop which allows signals with lower frequencies to pass through
dynamics. compared to (I − Ḡ A ), which means that more dynamics
from the model dependent part C AX will pass through in
4.4 INDI with derivative filter and synchronization the closed-loop dynamics. The direct influence of the distur-
bances are also high-pass filtered by (I − Ḡ A H Fcy ), which
If ẏ is estimated, as proposed in [6,11,27], and depicted in is similar to the complementary filter discussed in Sect. 4.2.
Fig. 12, i.e., with a filter from a sensor measurement and In Appendix A.4, it is shown that if the block diagram in
synchronized with the same filters H̄ and F̄cy (with possi- Fig. 13 is implemented instead, then the same closed-loop
bly different dimensions as indicated by the bar), as used for behavior as given in Eq. (46) is obtained, even without the
the derivative and sensor filtering in the control signal feed- additional assumptions given above.
back path, then, as shown in Appendix A.4, the closed-loop It is additionally seen that the synchronization in Fig. 13
transfer function is given by: corresponds to the INDI with complementary filter depicted

123
Aerospace Systems (2023) 6:285–304 295

Fig. 15 Re-formulation of hybrid INDI with complementary filter

Fig. 14 Hybrid INDI with complementary filter and synchronization (I − Ḡ A ), and the INDI with derivative filter and synchro-
as proposed in Kumtepe et al. [30]
nization (Sect. 4.4), (I − Ḡ A H Fcy ). If the state dynamics are
slow enough ẏ, hence follows ν with the actuator dynamics
in Fig. 10 with ẏmdl = Bu, i.e., only the high-frequency part Ḡ A . The direct influence of the disturbances are high-pass
of ẏ is considered in the estimate ẏmdl . This re-formulation filtered by (I − Ḡ A H Fcy ), which is similar to the comple-
can be attractive if different filters in Fcy are used in the mentary filter discussed in Sect. 4.2 and the derivative filter
separate control channels of y, e.g., that notch filters with with synchronization discussed in Sect. 4.4.
different notch frequencies are used for the individual axes The assumptions above can again be relaxed if the syn-
of the aircraft. Having the filter synchronization in the u feed- chronization on the feedback is moved to the complementary
back path does not allow for consistent separate filtering in filter and the filter chain Fcy (s) is split into sensor filters S(s)
the different axes. and notch/roll-off filters F(s) as shown in Fig. 15. Different
filtering F(s) in the control law can now be applied in the
4.5 Hybrid INDI feedback paths. With this modification the only assumption is
the sensor filters Sx (s) and S y (s) are diagonal and identical in
If ẏ is estimated, as proposed in Kumtepe et al. [30], with each channel in order to obtain the relation given by Eq. (50).
a complementary filter and synchronized as depicted in Fig.
14 (with the On-board plant model (OBPM) being ẏmdl = 4.6 Un-delayed estimate of model-based output
C Ax f +C Bu, then the closed-loop transfer function is given derivative ẏmdl
by:
In Section 4.2, the analysis was based on an un-delayed model
sY = I − Ḡ A H Fcy C AX − Ḡ A [I − H ] C AFcx X estimate, ẏmdl , of ẏ. This can for this example be obtained by
a cascaded complementary filter for ẋ and then x as depicted
+Ḡ A V + I − Ḡ A H Fcy C Bd D, (48) in Fig. 16. Since ẏ = C(Ax + Bu), this provides the nec-
essary estimate for the use in Section 4.2. Furthermore, this
as shown in Appendix A.5, under the assumption that complementary filter can provide the estimate ẋˆ for the ANDI
control law.
−1
C B I − G A H̄ F̄cy − G A (I − H̄ ) G A (C B)†
−1
= I − Ḡ A H Fcy − Ḡ A (I − H ) , (49)
5 Simulation
which holds if the basic assumptions and the additional
5.1 Roll dynamics example
assumptions mentioned in Sect. 4.4 are satisfied, respectively.
If for example Fcy and Fcx are diagonal with the same ele-
This subsection compares the different presented synchro-
ments on the diagonal, then:
nization approaches for the simplified roll motion dynamics
of an aircraft, given by
sY = (I − Ḡ A Fcy )C AX + Ḡ A V
+ I − Ḡ A H Fcy C Bd D. (50) ṗ = L p p + L ξ ξ, (51)

In this case, it can be seen that this filtering technique pro- with roll damping L p = −2.7 1/s, effectiveness L ξ = −14
duces a high-pass filter on the states which is (I − Ḡ A Fcy ) and 1/s 2 , control input ξ , which is the aileron deflection and mea-
hence with a cut-off frequency which is in between the cut-off sured output p, which is the roll rate. The actuator dynamics
frequencies of the high-pass filter in vanilla INDI (Sect. 4.1), are given by

123
296 Aerospace Systems (2023) 6:285–304

Fig. 16 State complementary

filter for un-delayed model
estimate of ẋ and x

Fig. 17 ANDI roll control

ωA
GA = , (52)
(s + ω A )

with ω A = 50 rad/s. The sensor dynamics are given by

ωS
S(s) = , (53)
s + ωS

with bandwidth ω S = 100 rad/s. The delay is given by

D y (s) = e−T s , (54)

with T = 0.03s. For this example, the filter F(s) = 1. Fcy (s)
is given by Eq. (39). The filter H is given by

ωH
H (s) = , (55)
s + ωH Fig. 18 Comparison of vanilla INDI and INDI with derivative filter
without synchronization
with bandwidth ω H = 30 rad/s.
Figure 18 shows the desired roll acceleration signal ṗdes ,
which shall be tracked, and the closed-loop roll acceleration (Fig. 14), and INDI with the cascaded complementary fil-
ṗ, for the vanilla INDI law depicted in Fig. 9, and the INDI ter (Figs. 10 and 16). For all cases, the pseudo-control is
with derivative filter and no synchronization depicted in Fig. again ν = ṗdes . It is seen that INDI with derivative filter
11. The pseudo-control is in both cases ν = ṗdes . Figure and synchronization improves the response compared to the
18 shows that in case, the derivative of p is obtained by INDI with derivative filter and without a synchronization, by
a differentiation filter without a synchronization technique, removing the oscillations. The hybrid INDI with the com-
the resulting response might be oscillatory. With increasing plementary filter and synchronization improves the response
phase loss of the filter Fcy (s), the closed-loop system will further (the resulting response is closer to the response
become unstable as happened in this case. obtained by vanilla INDI) compared to the pure derivative
Figure 19 compares the responses of vanilla INDI (Fig. 9), filter with synchronization. Compared to that, the INDI with
INDI with derivative filter and synchronization (Fig. 12), the proposed cascaded complementary filter improves the
hybrid INDI with complementary filter and synchronization response further by recovering the vanilla INDI response.

123
Aerospace Systems (2023) 6:285–304 297

Fig. 19 Comparison of vanilla INDI and INDI with different synchro- Fig. 21 Comparison of control signal with noise on the measurement
nization variants of y

results in exact tracking of the reference signal ṗr e f , at the

cost of additional necessary model information.
In Fig. 21, the control input ξ are depicted for the different
controller structures for the case that noise with a standard
deviation of 0.1deg is added to the measured roll rate p. It
can be seen that the resulting jitter on ξ is for all variants in a
similar magnitude range, although it seems slightly less for
ANDI.

5.2 Linearized lateral motion example

This subsection compares the different presented synchro-

nization approaches for the linearized lateral dynamics of an
aircraft given by
Fig. 20 Verification of the behavior of ANDI /E-INDI with cascaded
synchronization ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
ṙ Nr Nβ N p 0 r Nξ Nζ 
⎢ β̇ ⎥ ⎢−1 + Yr Yβ 0 g ⎥ ⎢ β ⎥ ⎢ Yξ Yζ ⎥ ξ
⎢ ⎥=⎢ V0 ⎥ ⎢ ⎥ + ⎢ ⎥
However, if the sensor dynamics are fast and the measure- ⎣ ṗ ⎦ ⎣ L r Lβ L p 0 ⎦ ⎣ p ⎦ ⎣Lξ Lζ ⎦ ζ
ment delay very low, the differences in the responses might be  ˙ 0 0 1 0  0 0
⎡ ⎤⎡ ⎤
minor. In this case, the hybrid INDI or INDI with differentia- −0.520 3.488 −0.628 0 r
tion filter and synchronization might be the preferred solution ⎢−0.987 −0.199 0 ⎥ ⎢
0.130⎥ ⎢ β ⎥
because of its lower complexity. If the sensor dynamics are =⎢
⎣ 0.472 −14.408 −6.624 0 ⎦ ⎣ p ⎦
⎥
slow or the measurement delay non-negligible, the cascaded 0 0 1 0 
complementary filter might be preferred. ⎡ ⎤
0.539 −2.005 
Figure 20 compares the E-INDI control law given by Eq. (17), ⎢ −0.012 0.040 ⎥ ξ
with the ANDI control law, given by Eq. (9), and depicted in +⎢ ⎥
⎣−10.700 2.899 ⎦ ζ , (56)
Figs. 2 and 17. Both the E-INDI and ANDI are implemented
0 0
with K 0 = k0 = 0, i.e., without roll rate error feedback.
Figure 20 reveals that E-INDI recovers the phase of the ref-
erence signal. This is because of the feed forward of ÿr e f in with y = [r , p]T and x = [r , β, p, ]T . All states are
E-INDI. ANDI also uses this feed forward and additionally filtered by the sensor dynamics S(s) given by Eq. (53)
the model dependent term Fx ẋ, which corresponds to L p ṗˆ in with ω S = 100 rad/s and are delayed with Eq. (54) with
the considered example. The closed-loop response of ANDI T = 0.03s. Only the rates p and r are additionally filtered

123
298 Aerospace Systems (2023) 6:285–304

Table 2 Values of notch filters ζN ω N (rad/s) gmin

p 0.7 2 π· 2 0.1
r 0.7 2 π· 5 0.3

Fig. 23 Comparison of vanilla INDI and INDI with different synchro-

nization variants ( ṗ command to ṙ response)

Fig. 22 Comparison of vanilla INDI and INDI with different synchro-

nization variants ( ṗ command to ṗ response)

by a notch filter given by

s 2 + 2gmin ζ N ω N + ω2N
F(s) = , (57)
s 2 + 2ζ N ω N + ω2N

with parameters summarized in Table 2 for p and r , respec-

tively. The filter H is given by Eq. (55) with bandwidth
ω H = 30 rad/s. For the different INDI synchronization
structures ν represents ṙ and ṗ. For the E-INDI and ANDI
structures ν represents r̈ and p̈. Fu = C B is given by Fig. 24 Comparison of vanilla INDI and INDI with different synchro-
 nization variants (ṙ command to ṗ response)
Nξ Nζ
Fu = , (58)
Lξ Lζ
responses for the given command in ṗdes . It can be seen
and Fx = C A by that with the proposed cascaded complementary filter the
response is equivalent to the vanilla INDI response. The other

Nr Nβ N p 0 techniques slightly deteriorate the response, especially with
Fx = . (59)
Lr Lβ L p 0 regard to the decoupling of the axes. If a pure roll command
is given the response in the r channel is desired to be as
The actuator dynamics are given by low as possible, but the excursion is much higher in the con-
ventional synchronization schemes as shown in Fig. 23. The
G A = (s I + )−1 , (60) same holds for the responses to a yaw step command, which
are depicted in Figs. 24 and 25.
with  = I ω A and ω A = 50rad/s. Figures 22 to 25 com- The simulations done until now show that the cascaded
pare the response of ideal vanilla INDI (i.e., with perfect filter has improvement over the synchronization on u, but
knowledge of ẏ as depicted in Fig. 9) with the different syn- the synchronization on u still works reasonably well. The
chronization techniques depicted in Figs. 12, 14 and 10, with synchronization approach does not require any additional
the notch filters placed on the dominant feedback channel model information; hence, it might be the preferred choice.
of the input. Figures 22 and 23 thereby show the ṗ and ṙ The above simulations were made with a weakly coupled

123
Aerospace Systems (2023) 6:285–304 299

Fig. 25 Comparison of vanilla INDI and INDI with different synchro- Fig. 27 Verification of the behavior of ANDI /E-INDI with cascaded
nization variants (ṙ command to ṙ response) synchronization ( ṗ command to ṗ response)

Fig. 26 Comparison of INDI with synchronization on u with the re- Fig. 28 Verification of the behavior of ANDI /E-INDI with cascaded
formulated synchronization on the feedback ( ṗ command to ṗ response) synchronization ( ṗ command to ṙ response)

C B matrix; hence, the assumptions given in Eqs. (47) and tary filter approach is not affected by the couplings. Figure
(49) almost hold. For conventional fixed wing aircraft, this 26 also shows the same behavior for the Hybrid approach
is usually the case, but, for advanced configurations this is with synchronization in u as depicted in Fig. 14, and should
not necessarily the case. For example in a V-tail aircraft be compared to the re-formulated Hybrid INDI depicted in
the control inputs are heavily coupled in the r and p chan- Fig. 15.
nel. In this case, having the synchronization on u can have
adverse effects as shown in the following. As derived in 
1.8 −2
Sects. 4.4 and 4.5, re-formulations of the synchronization C Bcoupled = . (61)
as complementary filters allows for separate filtering in the −10.7 8
different channels. The following simulations are performed
with the exact same lateral system only changing the C B At last, we compare the ANDI and E-INDI approaches used
matrix to be strongly coupled as given in Eq. (61). In Fig. with the proposed cascaded complimentary filter in Figs.
26 the synchronization from Fig. 12 is compared with the 27, 28, 29 and 30. It can be seen that the response for the
re-formulated filter from Fig. 13. It is seen that the syn- ANDI technique with cascaded complementary filter per-
chronization technique on u is even unstable in this case, fectly tracks the reference dynamics and perfectly decouples
where the performance of the re-formulated complemen- the roll and yaw channel.

123
300 Aerospace Systems (2023) 6:285–304

in non-steady state conditions. At last, it was shown that a

cascaded complementary filter can be used to obtain an un-
delayed model-based derivative estimate. Future work will
be focused on comparing the performance and robustness in
practical applications.
Acknowledgements The authors would like to thank Tijmen Pollack,
Haichao Hong and Ewoud Smeur for valuable discussions and com-
ments.
The authors would also like to thank the reviewers for their very relevant
comments.

Funding This research is co-funded by the European Union in the scope

of INCEPTION project, which has received funding from the EU Hori-
zon 2020 Research and Innovation Programme under grant agreement
No. 723515. Open Access funding enabled and organized by Projekt
DEAL.

Fig. 29 Verification of the behavior of ANDI /E-INDI with cascaded Declarations

synchronization (ṙ command to ṗ response)
Conflict of interest The authors declare that they have no conflict of
interest.

Open Access This article is licensed under a Creative Commons

Attribution 4.0 International License, which permits use, sharing, adap-
tation, distribution and reproduction in any medium or format, as
long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons licence, and indi-
cate if changes were made. The images or other third party material
in this article are included in the article’s Creative Commons licence,
unless indicated otherwise in a credit line to the material. If material
is not included in the article’s Creative Commons licence and your
intended use is not permitted by statutory regulation or exceeds the
permitted use, you will need to obtain permission directly from the copy-
right holder. To view a copy of this licence, visit [Link]
[Link]/licenses/by/4.0/.

Appendix A Derivation of closed-loop trans-

Fig. 30 Verification of the behavior of ANDI /E-INDI with cascaded fer functions
synchronization (ṙ command to ṙ response)

Note that in the following the arguments (s) and (t) are omit-
6 Conclusion ted for better readability.

In this paper, we investigated different ways to compensate A.1 Vanilla INDI closed-loop transfer function
and synchronize for filtering and delays of the measurements
in several types of incremental control laws. It was shown Given the system in Eq. (41) and the vanilla INDI control law
how to synthesize complementary filters, taking into account with block diagram depicted in Fig. 9, then the closed-loop
the filtering and delays. It was shown that the use of comple- transfer function can be derived as follows.
mentary filters in contrast to synchronization on the feedback The transfer function from u to u in the Laplace domain
of u can be especially useful in MIMO systems, if separate can be deduced from the block diagram to be
filtering is necessary in the different measurements. It was
shown that in case of strongly coupled input effectiveness, the U = (I − G A )−1 G A U . (A1)
synchronization on the actuator feedback may become even
unstable, where the equivalent complementary filter formu-
lation recovers the design closed-loop behavior. A method Substituting this relation into
for deriving the initialization of the complementary filters
was presented, which allows for transient-free engagement sY = C(AX + BU + Bd D), (A2)

123
Aerospace Systems (2023) 6:285–304 301

   
and replacing U by sY = I − Ḡ A C AX + I − Ḡ A C Bd D
 
+Ḡ A V + I − H Fcy C Bd D . (A11)
U = (C B)† (V − sY ) (A3)
Finally,
results in
   
sY = I − Ḡ A C AX + Ḡ A V + I − Ḡ A H Fcy C Bd D
−1
sY = C AX + C B(I − G A ) G A (C B) (V − sY )
†
(A12)
+C Bd D.
(A4) is obtained.

Using the assumption in Eq. (42) and rearranging the above A.3 INDI with derivative filter and no
results in synchronization
  −1 
I + I − Ḡ A Ḡ A sY Given the system (41) and the INDI control law with block
 −1 diagram depicted in Fig. 11, the closed-loop transfer function
= C AX + I − Ḡ A Ḡ A V + C Bd D. (A5) can be derived as follows. The transfer function from u
  to u in the Laplace domain can be deduced from the block
Multiplying with I − Ḡ A from the left, results in diagram to be (A1). Substituting this relation into (A2) and
    replacing U by
sY = I − Ḡ A C AX + Ḡ A V + I − Ḡ A C Bd D. (A6)  
U = (C B)† V − s H Fcy Y (A13)
A.2 INDI with complementary filter from Section 3.3
results in
Given the system (41) and the INDI control law with block
diagram depicted in Fig. 10, then the closed-loop transfer sY = C AX + C Bd D
 
function can be derived as follows, if ẏmdl is calculated by +C B(I − G A )−1 G A (C B)† V − s H Fcy Y .
(A14)
ẏmdl (t) = C(Ax(t) + Bu(t)). (A7)
Using the assumption in Eq. (42) and rearranging the above
The transfer function from u to u in the Laplace domain can results in
be deduced from the block diagram to be (A1). Substituting   −1 
this relation into (A2) and replacing U by I + I − Ḡ A Ḡ A H Fcy sY
 −1 (A15)
 = C AX + C Bd D + I − Ḡ A Ḡ A V.
U = (C B)† V − I − H Fcy (C AX + C BU )

−s H Fcy Y (A8) Finally,
 −1  
results in sY = I − Ḡ A + Ḡ A H Fcy I − Ḡ A C AX
 −1
sY = C AX + C Bd D+ + I − Ḡ A + Ḡ A H Fcy Ḡ A V (A16)
 −1  
C B(I − G A )−1 G A (C B)† (A9) + I − Ḡ A + Ḡ A H Fcy I − Ḡ A C Bd D
(V − I − H Fcy (C AX + C BU ) − s H Fcy Y ).
is obtained.
Adding and subtracting C Bd D in the term from Ẏmdl and
A.4 INDI with derivative filter and synchronization
using the structure of sY from (A2) gives:
Given the system (41) and the INDI control law with block
sY = C AX + C Bd D
diagram depicted in Fig. 12, the closed-loop transfer function
+C B(I − G A )−1 G A (C B)† (V − sY can be derived as follows.
+ I − H Fcy C Bd D). (A10) The transfer function from u to u in the Laplace domain
can be deduced from the block diagram to be
Applying the assumption given in Eq. (42), arranging for sY  −1
and multiplying by (I − Ḡ A ) from the left gives: U = I − G A H̄ F̄cy GA U. (A17)

123
302 Aerospace Systems (2023) 6:285–304

 
Substituting this relation into (A2) and replacing U by sY = I − Ḡ A H Fcy C AX + Ḡ A V
 
  + I − Ḡ A H Fcy C Bd D. (A26)
U = (C B)† V − s H Fcy Y (A18)
Note that in contrast to the original implementation in
results in Fig. 12, with the proposed modification in Fig. 13 it is not
 −1 required to assume that H̄ F̄cy are square, only have elements
sY = C AX + C B I − G A H̄ F̄cy G A (C B)† on the diagonal and u and y have the same dimension, or
  (A19)
V − s H Fcy Y + C Bd D. that H Fcy only has elements on the diagonal and all the
elements are the same. This means that u and y can have
Under the assumption given by Eq. (47), we obtain different dimensions and different filtering can be applied to
  the different measurement signals.
 −1
I + I − Ḡ A H Fcy Ḡ A H Fcy sY
 −1 (A20) A.5 Hybrid INDI with complementary filter using
= C AX + I − Ḡ A H Fcy Ḡ A V + C Bd D. synchronization
 
Finally, multiplying with I − Ḡ A H Fcy from the left results Given the system (41) and the INDI control law with block
in diagram depicted in Fig. 14, the closed-loop transfer function
  can be derived as follows, if ẏmdl is calculated by
sY = I − Ḡ A H Fcy C AX + Ḡ A V
  (A21)
+ I − Ḡ A H Fcy C Bd D. ẏmdl (t) = C(Ax f (t) + Bu(t)), (A27)

If the synchronization is implemented according to the where x f is x filtered by Fcx (s).

block diagram in Fig. 13, then the closed-loop transfer func- The transfer function from u to u in the Laplace domain
tion can be calculated as follows: can be deduced from the block diagram to be

U = (I − G A )−1 G A (C B)† −1
    U = I − G A H̄ F̄cy − G A (I − H̄ ) GA U. (A28)
V − s H Fcy Y − I − H Fcy C BU . (A22)
Substituting the transfer function in Eq. (A28) into Eq. (A2)
Applying the assumption given in Eq. (42), and adding/
and replacing U by
subtracting the terms C AX and C Bd D results in

 −1 U = (C B)† V − [I − H ] C AFcx X
sY = C AX + C Bd D + I − Ḡ A Ḡ A 
   − [I − H ] C BU − s H Fcy Y (A29)
V − s H Fcy Y − I − H Fcy
(A23)
(C BU + C AX + C Bd D) results in
  
+ I − H Fcy (C AX + C Bd D) .
sY = C AX + C Bd D
Using the structure of sY from Eq. (A2) gives: −1
+ C B I − G A H̄ F̄cy − G A (I − H̄ ) G A (C B)†
 −1  
sY = C AX + I − Ḡ A Ḡ A V − sY V − [I − H ] C AFcx X
   
+ I − H Fcy (C AX + C Bd D) + C Bd D. − [I − H ] C BU − s H Fcy Y .
(A24) (A30)

 
Isolating for sY and multiplying by I − Ḡ A from the left Using the assumption given in Eq. (49), adding and sub-
results in tracting [I − H ] C Bd and [I − H ] C AX and rearranging the
above results in
 
sY = I − Ḡ A C AX
    −1
+Ḡ A V + I − H Fcy (C AX + C Bd D) sY = C AX + C Bd D + I − Ḡ A H Fcy − Ḡ A (I − H ) Ḡ A
  
+ I − Ḡ A C Bd D. (A25) V − [I − H ] C AFcx X + [I − H ] C AX

− I − H + H Fcy sY + [I − H ]C Bd D .
Finally, the closed-loop relation can be obtained: (A31)

123
Aerospace Systems (2023) 6:285–304 303

Isolating for sY and multiplying by I − Ḡ A H Fcy − Ḡ A (I − 12. Smeur EJ, de Croon GC, Chu Q (2018) Cascaded incremental non-
H ) from the left results in linear dynamic inversion for MAV disturbance rejection. Control
Eng Pract 73:79–90
13. Yang J, Cai Z, Zhao J, Wang Z, Ding Y, Wang Y (2023) INDI-based
aggressive quadrotor flight control with position and attitude con-
sY = I − Ḡ A H Fcy − Ḡ A (I − H ) (C AX + C Bd D) straints. Robot Auton Syst 159:104292. [Link]
+ Ḡ A (V − [I − H ] C AFcx X + [I − H ] C AX robot.2022.104292
14. Akkinapalli VS, Holzapfel F (2018) Incremental dynamic inversion
+ [I − H ] C Bd D) based velocity tracking controller for a multicopter system. In: 2018
AIAA Guidance, Navigation, and Control Conference; p. 1345
= I − Ḡ A H Fcy C AX − Ḡ A [I − H ] C AFcx X
15. Smeur EJ, Bronz M, de Croon GC (2020) Incremental Control
+ Ḡ A V + I − Ḡ A H Fcy C Bd D. and Guidance of Hybrid Aircraft Applied to a Tailsitter Unmanned
Air Vehicle. J Guid Control Dyn 43(2):274–287. [Link]
(A32) 10.2514/1.G004520
16. Tal E, Karaman S (2022) Global incremental flight control for
agile maneuvering of a tailsitter flying wing. J Guid Control Dyn
Assuming that Fcy and Fcx are diagonal with the same ele-
45(12):2332–2349
ments on the diagonal, then 17. Steinleitner A, Frenzel V, Pfeifle O, Denzel J, Fichter W. In: Auto-
matic Take-Off and Landing of Tailwheel Aircraft with Incremental
Nonlinear Dynamic Inversion;. Available from: [Link]
sY = (I − Ḡ A Fcy )C AX + Ḡ A V + I − Ḡ A H Fcy C Bd D. 2514/6.2022-1228
18. Pavel M, Shanthakumaran P, Chu Q, Stroosma O, Wolfe M,
(A33) Cazemier H (2020) Incremental Nonlinear Dynamic Inversion
for the Apache AH-64 Helicopter Control. J Am Helicopter Soc
65(2):1–16
is obtained. 19. Azinheira JR, Moutinho A, Carvalho J (2015) Lateral control
of airship with uncertain dynamics using incremental nonlinear
dynamics inversion. IFAC-PapersOnLine 48(19):69–74
20. Steffensen R, Steinert A, Holzapfel F (2022) Longitudinal Incre-
References mental Reference Model for Fly-By-Wire Control Law using
Incremental Non-Linear Dynamic Inversion. In: AIAA SCITECH
1. Lombaerts T, Looye G (2011) Design and flight testing of manual 2022 Forum; p. 1230
nonlinear flight control laws. In: AIAA Guidance, Navigation, and 21. de Angelis Cordeiro R, Azinheira JR, Moutinho A (2019) Cas-
Control Conference; p. 6469 caded incremental backstepping controller for the attitude tracking
2. Lombaerts T, Kaneshige J, Schuet S, Hardy G, Aponso BL, Shish of fixed-wing aircraft. In: 5th CEAS Conference on Guidance, Nav-
KH (2019) Nonlinear dynamic inversion based attitude control for igation and Control
a hovering quad tiltrotor eVTOL vehicle. In: AIAA Scitech 2019 22. Gabrys AC, Steffensen R, de Angelis Cordeiro R, Azinheira JR,
Forum; p. 0134 Moutinho A, Holzapfel F (2019) Integration of Phase Plane Flight
3. Harris JJ (2018) F-35 flight control law design, development and Envelope Protections in Cascaded Incremental Flight Control.
verification. In: 2018 Aviation Technology, Integration, and Oper- IFAC-PapersOnLine 52(12):429–435
ations Conference; p. 3516 23. Sieberling S, Chu Q, Mulder J (2010) Robust flight control using
4. Holzapfel F, Sachs G In: Dynamic Inversion Based Control Con- incremental nonlinear dynamic inversion and angular acceleration
cept with Application to an Unmanned Aerial Vehicle;. Available prediction. J Guid Control Dyn 33(6):1732–1742
from: [Link] 24. Simplício P, Pavel M, Van Kampen EJ, Chu Q (2013) An accelera-
5. Krstic M, Kokotovic PV, Kanellakopoulos I (1995) Nonlinear and tion measurements-based approach for helicopter nonlinear flight
adaptive control design. John Wiley & Sons Inc, US control using incremental nonlinear dynamic inversion. Control
6. Grondman F, Looye G, Kuchar RO, Chu Q, Van Kampen EJ Eng Pract 21(8):1065–1077
In: Design and Flight Testing of Incremental Nonlinear Dynamic 25. Huang Y, Zhang Y, Pool DM, Stroosma O, Chu Q (2022) Time-
Inversion-based Control Laws for a Passenger Aircraft;. Available delay margin and robustness of incremental nonlinear dynamic
from: [Link] inversion control. J Guid Control Dyn 45(2):394–404
7. Pollack T, Looye G, Van der Linden F (2019) Design and flight test- 26. Li M, Li J, Tang Y, Sun S (2021) An extended INDI approach
ing of flight control laws integrating incremental nonlinear dynamic and application to pitch rate control laws design of an aircraft. In:
inversion and servo current control. In: AIAA Scitech 2019 Forum; AIAA aviation 2021 forum. [Link]
p. 0130 [Link]
8. Raab SA, Zhang J, Bhardwaj P, Holzapfel F In: Proposal of a Uni- 27. van’t Veld R, Van Kampen EJ, Chu Q (2018) Stability and robust-
fied Control Strategy for Vertical Take-off and Landing Transition ness analysis and improvements for incremental nonlinear dynamic
Aircraft Configurations;. Available from: [Link] inversion control. In: 2018 AIAA Guidance, Navigation, and Con-
6.2018-3478 trol Conference; p. 1127
9. Raab SA, Zhang J, Bhardwaj P, Holzapfel F (2019) Consideration 28. Bacon BJ, Ostroff AJ, Joshi SM (2001) Reconfigurable NDI con-
of Control Effector Dynamics and Saturations in an Extended INDI troller using inertial sensor failure detection & isolation. IEEE
Approach. In: AIAA Aviation 2019 Forum; p. 3267 Trans Aerosp Electron Syst 37(4):1373–1383
10. Milz D, Looye G (2022) Tilt-wing control design for a unified 29. de Angelis Cordeiro R, Marton AS, Azinheira JR, Carvalho JR,
control concept. In: AIAA Scitech 2022 forum, p 1084 Moutinho A (2021) Increased Robustness to Delay in Incremental
11. Smeur EJ, Chu Q, de Croon GC (2016) Adaptive incremental non- Controllers using Input Scaling Gain. IEEE Transa Aerosp Electron
linear dynamic inversion for attitude control of micro air vehicles. Syst
J Guid Control Dyn 39(3):450–461

123
304 Aerospace Systems (2023) 6:285–304

30. Kumtepe Y, Pollack T, Van Kampen EJ. In: Flight Control Law 32. Steffensen R, Steinert A, Smeur EJ (2022) Nonlinear Dynamic
Design using Hybrid Incremental Nonlinear Dynamic Inversion;. Inversion with Actuator Dynamics: An Incremental Control Per-
Available from: [Link] spective. J Guid Control Dyn. [Link]
31. Liu Z, Zhang Y, Liang JJ, Chen H (2022) Application of the 33. Bhardwaj P, Raab SA, Holzapfel F (2021) Higher Order Refer-
improved incremental nonlinear dynamic inversion in fixed-wing ence Model for Continuous Dynamic Inversion Control. In: AIAA
UAV flight tests. J Aerosp Eng 35(6):04022091 Scitech 2021 Forum; p. 1130

123
 Appendix A: Article Reprints

A.5 Paper 5
Paper 5, reference [70], is re-printed in full length with permission from the authors and
the American Institute of Aeronautics and Astronautics, Inc.

LXXXV
 AIAA SciTech Forum 10.2514/6.2022-1230
January 3-7, 2022, San Diego, CA & Virtual
AIAA SCITECH 2022 Forum

Longitudinal Incremental Reference Model for Fly-By-Wire

Control Law using Incremental Non-Linear Dynamic Inversion

Rasmus Steffensen∗ , Agnes C. Steinert† , Florian Holzapfel ‡

Institute of Flight System Dynamics, Technical University of Munich, Garching bei München, 85748, Germany

In this paper a control architecture including an incremental physical reference model is

presented for the longitudinal channel of an incremental non-linear dynamic inversion flight
control law to provide fly-by-wire capabilities for a piloted fixed wing turboprop demonstrator
aircraft. The reference model is formulated in terms of dynamics with respect to the maximum
angle of attack in the angle of attack domain, and with respect to trim values for pitch rate and
load factor. Together with the incremental dynamic inversion this relieves the use of uncertain
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

trim tables for the control law. A blended control variable is proposed to provide intuitive
control over the whole flight envelope. A compensation algorithm adjusts the reference model
trajectories in case of actuator limits or rate limits predicted by the control allocation. An error
controller formulated in terms of equivalent load factor dynamics is proposed, together with a
conversion to pitch rate acceleration for the incremental non-linear dynamic inversion.

I. Introduction
Flight control laws providing fly-by-wire capabilities have been around since the last half of the 20th century. Over
the time, many different paradigms of design and synthesis have evolved. Since the 1990 in particular methods
using non-linear dynamic inversion have been used in research [1] but also in production aircraft [2]. This method
relies on accurate model information, hence many methods exchanging the model dependency of the inversion
with sensor information have emerged, [3, 4]. A popular method called incremental non-linear dynamic inversion
[5–7] (INDI), solves the problem of model dependency with sensor measurements, by linearising the inversion step
and substituting parts of the model dependency with sensor measurements and only relying on model information
for the control effector influence on the system. As a linear relationship between the so called pseudo controls 𝜈
and the 𝑟’th derivative of the controlled output 𝑦, where 𝑟 is the relative degree of the system, is established, any
control design method can be used from here on to control the system. In [8], a feedback control law is used to
control the output. In [9] and [10] a model following controller is used to control a VTOL transition vehicle. An
advantage of model following controllers is the possibility to design the desired behaviour directly as a reference model
with the desired dynamic behaviour. This was also shown with a reference model based on the aircraft physics by [11, 12].

In this paper, an INDI type longitudinal control law architecture with an incremental reference model is proposed for a
fixed wing all attitude turboprop demonstrator aircraft. A blended control variable is proposed adapting to the current
envelope boundary for high stick deflections and consisting of a 𝐶 ∗ type control variable at low stick deflections. In
addition at low speeds, a transient free transition to a pseudo speed stable aircraft response is introduced by varying
the pseudo trim value command of the system. This speed stability also provides a natural flare tendency relieving
the necessity of a separate flare law. An incremental physical reference model is proposed, with the dynamics of
the reference angle of attack described in relation to the limit value, hence removing the need for trim tables with
corresponding uncertainties. An algorithm for adjusting reference model commands in case of actuator limit or rate
limits is proposed achieving an intuitive and physically feasible reference response. At last, a dynamic inversion type
conversion of equivalent incremental load factor to pitch acceleration command is proposed to connect the underlying
INDI control law with the incremental reference model and corresponding error dynamics.

The outline of this paper is as follows. Section III introduces the overall control law architecture. Section IV
shows the base value calculation for the incremental reference model. The novel incremental reference model is
∗ Research Associate, Ph.D. Student, [Link]@[Link]
† Research Associate, Ph.D. Student, [Link]@[Link]
‡ Professor, Head of the Institute Flight System Dynamics, [Link]@[Link], Senior Member AIAA

Copyright © 2022 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved.
 introduced in Section V. Section VI presents the error controller and Section VII shortly describes the control allocation
and INDI control law. Finally Section VIII provides some concluding remarks.

II. Nomenclature
In this paper, vectors and matrices are noted in bold, and scalars in italic. In addition, geometric vectors are denoted
with the following sub and super scripts:
 𝑑
a𝑏𝑐 (1)
𝑒
where 𝒂 is the type of vector, 𝑏 denotes a reference point, 𝑐 denotes type of velocity or acceleration addressed, 𝑑 refers
to which reference frame possible derivatives are taken with respect to and 𝑒 indicates which frame the coordinates of
the vector is denoted in. Coordinate transformation matrices are given with the following subscripts:

M 𝐴𝐵 (2)
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

where M is the matrix rotating a vector from frame 𝐵 to frame 𝐴.

1) Reference point
𝐺 Center of gravity of the aircraft.
2) Velocity and acceleration type
𝐾 Kinematic, 𝐴 Aerodynamic.
3) Reference frames
𝐼 Inertial frame, 𝐸 Earth centered earth fixed, 𝑂 North-East-Down frame, 𝐾 Kinematic frame, 𝐵 Body fixed frame,
𝑆 Stability frame.

Some
examples are given by:

• 𝒓 𝐺 𝐵 denotes a position vector for the reference point 𝐺 denoted in the body fixed frame 𝐵.

𝐸
• 𝑽𝐾𝐺 𝐴 is the kinematic velocity (𝐾) with respect to the earth centered earth fixed coordinate system (𝐸) of reference
point 𝐺,

denoted in the aerodynamic reference frame 𝐴.
• 𝝎𝑂𝐵 𝑂 is the rotation vector of the body fixed frame 𝐵 with respect to the NED frame 𝑂, denoted in the NED frame 𝑂.

III. Control Law Architecture

Hard Protections
&
Special Functions

𝛿𝑠,𝑙𝑜𝑛 Control Aircraft

Error Input
Reference Model Allocation &
Controller Channel
INDI

Base Calculation
& NL Feedforward
Filtering & Output
Datafusion Channel
Soft Protections
&
Pilot Relief Functions

Fig. 1 Controller Components

The proposed controller architecture is depicted in Fig. 1. The main components are an incremental physical reference
model that shapes the pilot commands in terms of desired dynamics, an error controller to track these commands and a
Control Allocation to distribute the commands to the control effectors together with an INDI type inner-loop to track the
desired body angular accelerations. The proposed structure allows for straight forward integration of hard protections
like low energy protection, angle of attack and load factor protection, additional functions like automatic upset recovery,
as well as soft protections, like overspeed protection and pilot relief functions. The functions and protections themself
are not within the scope of this paper, as indicated in Fig. 1. The Filtering and Datafusion contains roll-off filters, notch
filters and estimators for different entities like angle of attack. However, these estimators are also not in the scope of this
paper.

2
 IV. Base Value Calculation
The reference model described in Section V.A is formulated in terms of increments to reference values, which are
introduced in this section.

A. Base load factor

The pilot commands increments in equivalent load factor from a defined base load factor 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 . These increments
are given in the stability frame. The incremental reference model hence runs in values incremental to this base load
factor as described in Section V.A. The base load factor can be thought of as an operational point usually associated
with a trim value. The base / reference value is comprised of the gravity trim value, the turn compensation, load factor
additions 𝑛 𝑧, 𝑝𝑟 𝑓 ,𝑆 from the pilot relief functions, and a dynamic saturation based on the soft protection load factor
limits 𝑛 𝑧, 𝑝𝑟 𝑜𝑡 ,𝑠,ℎ𝑖,𝑆 and 𝑛 𝑧, 𝑝𝑟 𝑜𝑡 ,𝑠,𝑙𝑜,𝑆 , which for this control law is an over-speed protection, as depicted in Fig. 2. The
base value 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 is calculated according to
𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 = 𝑠𝑎𝑡 (𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑢𝑛𝑠𝑎𝑡 ,𝑆 , 𝑛 𝑧, 𝑝𝑟 𝑜𝑡 ,𝑠,𝑙𝑜,𝑆 , 𝑛 𝑧, 𝑝𝑟 𝑜𝑡 ,𝑠,ℎ𝑖,𝑆 ) (3)
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

where the unsaturated load factor reference value is

𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑢𝑛𝑠𝑎𝑡 ,𝑆 = min(𝑛 𝑧,𝑔,𝑡𝑢𝑟 𝑛,𝑆 , 𝑛 𝑧, 𝛼𝐶,min ,𝑆 ) + 𝑛 𝑧, 𝑝𝑟 𝑓 ,𝑆 (4)
with
¯ 𝐿, 𝛼𝐶,min
𝑞𝑆𝐶
𝑛 𝑧, 𝛼𝐶,min ,𝑆 = (5)
𝑚𝑔
where 𝐶 𝐿, 𝛼𝐶,min is the lift coefficient at 𝛼𝐶,min at the measured dynamic pressure 𝑞,
¯ and 𝑛 𝑧,𝑔,𝑡𝑢𝑟 𝑛,𝑆 is calculated by Eq.
(9).
The min function in Eq. (4) automatically provides a natural speed stability and fading out of turn compensation by
overriding the trim load factor at lower speeds. Because of the min function 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑢𝑛𝑠𝑎𝑡 ,𝑆 will be continuous and
hence provides a seamless transition to the speed stable domain of the control law. The pilot relief function load factor
additions 𝑛 𝑧, 𝑝𝑟 𝑓 ,𝑆 and the soft protection load factor limits 𝑛 𝑧, 𝑝𝑟 𝑜𝑡 ,𝑠,ℎ𝑖,𝑆 , 𝑛 𝑧, 𝑝𝑟 𝑜𝑡 ,𝑠,𝑙𝑜,𝑆 are not within the scope of this
paper. The turn compensation can be calculated conveniently in the body frame by
cos Θ
𝑛 𝑧,𝑡𝑢𝑟 𝑛,𝐵 = (6)
cos Φ𝑙𝑖𝑚
because of the availability of the measurements Φ and Θ. Φ𝑙𝑖𝑚 is the measured bank angle Φ, limited to ±50◦ . When
the bank angle increases from ±60◦ to ±90◦ , Φ𝑙𝑖𝑚 is faded back to 0. Hence, the turn compensation assures, that for
angles less than 50◦ , 𝛾¤ 𝐾 will be approximately 0. The gravity trim value in body fixed frame z-direction is chosen to be
the level flight body trim value:
𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑔,𝐵 = cos Θ (7)
This is a common choice, which provides intuitive behavior for high bank angles and inverted flight. For the considered
aircraft, the thrust is almost acting in the body-x direction, and since the pilot controls the thrust and hence the body-x
acceleration, it is not intended to super-impose control of the acceleration in the body x-direction. Hence, the measured
𝑛 𝑥,𝐵 is also the desired one to be used in Eq. (9). The combined turn and gravity trim value are calculated as a blend by:
𝑛 𝑧,𝑔,𝑡𝑢𝑟 𝑛,𝐵 = (𝑛 𝑧,𝑡𝑢𝑟 𝑛,𝐵 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑔,𝐵 )𝜆 𝑡 𝑐 + 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑔,𝐵 (8)
𝜆 𝑡 𝑐 is a fading parameter that allows to fade the turn compensation out in case of high roll rate commands. The fading
parameter 𝜆 𝑡 𝑐 is calculated with a look-up table, where based on the lateral stick deflection 𝜆 𝑡 𝑐 takes values between 0
and 1. This provides intuitive behavior for high roll rates. If the turn compensation was not faded out, it would produce
rapidly oscillating load factors in rapid roll maneuvers.
The combined turn and gravity contribution can be represented in the stability frame by:

© ª
𝑛 𝑥,𝐵
­
M𝑆𝐵 ­ ®
0 ® (9)
𝑇
𝑛 𝑧,𝑔,𝑡𝑢𝑟 𝑛,𝑆 = −e 𝑘
« −𝑛 𝑧,𝑔,𝑡𝑢𝑟 𝑛,𝐵 ¬𝐵
where M𝑆𝐵 is given in Eq. (102) and e 𝑘 is the unit vector along the 𝑧-axis. An immediate advantage of using body
frame for the trim calculation is, that errors in MSB will cancel, because the final load factor tracking is performed with
load factor measurements in the body frame.

3
 Base Calculation & NL Feedforward 𝑛𝑧,𝑝𝑟𝑜𝑡,𝑠,ℎ𝑖,𝑆

𝑛𝑧,𝛼𝐶,𝑚𝑖𝑛,𝑆 𝑛𝑧,𝑡𝑟𝑖𝑚,𝑢𝑛𝑠𝑎𝑡,𝑆
+ 𝑛𝑧,𝑡𝑟𝑖𝑚,𝑆
𝑛𝑥,𝐵 𝑚𝑖𝑛 𝑛𝑧,𝑡𝑟𝑖𝑚,𝑆
𝐌𝑆𝐵
0 + 𝑞𝑡𝑟𝑖𝑚
𝐌𝐵𝑂 −𝒆𝑇𝑘 𝐌𝑆𝐵 cos Θ 𝑞ሶ 𝑡𝑟𝑖𝑚
𝑛𝑥,𝐵 − − cos Θ 𝜆𝑡𝑐 − cos Θ
Θ cos Φ𝑙𝑖𝑚 𝐵
𝑛𝑧,𝑝𝑟𝑓,𝑆 𝑛𝑧,𝑝𝑟𝑜𝑡,𝑠,𝑙𝑜,𝑆
Φ𝑙𝑖𝑚
𝛽መ𝐴
𝑝𝐾 𝑛𝑧,𝑡𝑟𝑖𝑚,𝑆 𝑞𝑡𝑟𝑖𝑚
𝑟𝐾 + 𝑔 𝑞𝑡𝑟𝑖𝑚,𝑔
𝐞𝑘 𝑇 𝐌𝑆𝐵 𝐌𝐵𝑂 𝐠 𝑂 /𝑔 𝑉𝐴𝐺 𝐸
- + 𝑞ሶ 𝑡𝑟𝑖𝑚
𝐺𝑑 (𝑠)
𝑞𝑘𝑐 +
tan𝛽መ𝐴 𝑝𝐾 cos𝛼𝐴 + 𝑟𝐾 sin𝛼𝐴

Fig. 2 Block diagram of trim value calculation

Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

B. Base pitch rate

The base pitch rate 𝑞 𝑡𝑟 𝑖𝑚 is the reference value for which the incremental pitch rate of the incremental reference model,
described in Section V.A, is referenced to. This pitch rate is based on the steady state relation for an angle of attack rate
of 0, i.e. 𝛼¤ 𝐴 = 0. It is given by
𝑞 𝑡𝑟 𝑖𝑚 = 𝑞 𝑡𝑟 𝑖𝑚,𝑔 + 𝑞 𝑘𝑐 (10)
The base pitch rate is calculated using the pitch rate trim contributions as depicted in Fig. 2 and as as calculated in Eq.
(84) and Eq. (83), assuming small sidelip angles, i.e. cos 𝛽 = 1:
𝑔  

𝐸 𝑧,𝑡𝑟 𝑖𝑚,𝑆 e M M (11)
𝑇
𝑞 𝑡𝑟 𝑖𝑚,𝑔 = 𝑛 − 𝑘 𝑆𝐵 𝐵𝑂 (g) 𝑂
𝑉 𝐴𝐺

𝑞 𝑘𝑐 = tan 𝛽 ( 𝑝 𝐾 cos 𝛼 𝐴 + 𝑟 𝐾 sin 𝛼 𝐴) (12)

The base load factor 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 is calculated by Eq. (3). Note, this means that the steady state pitch rate for a constant
load factor maneuver with decreasing speed is not considered in this base term, however, this is negligible (𝑞 < 1𝑑𝑒𝑔/𝑠)
compared to the base pitch rate required in rapid rollings (𝑞 > 15𝑑𝑒𝑔/𝑠).

V. Physical Incremental Reference Model

The physical reference model consists of five main parts: a reference control law, phase plane based protections,
equivalent actuator dynamics or pseudo control dynamics (in this case pitch acceleration dynamics) and an incremental
short period plant model as shown in Fig. 3. The phase plane based protections, [13, 14], are dynamic saturations on the
pitch acceleration calculated using the margin to the reference limit values, which is not further discussed in this paper.
𝑛𝑧,𝑡𝑟𝑖𝑚,𝑆
Hard protections
Upset recovery 𝑛𝑧,𝑡𝑟𝑖𝑚,𝑆 𝑛𝑧,𝑡𝑟𝑖𝑚,𝑆

Reference Model

Δ𝑞ሶ 𝑟𝑒𝑓,𝑓𝑓 Δ𝑞ሶ 𝑟𝑒𝑓,𝑓𝑓

Δ𝜈𝐶,𝑑𝑓𝑐
Δ𝑞ሶ 𝑡𝑟𝑖𝑚,𝑟𝑒𝑓
Reference Δ𝑞ሶ 𝑟𝑒𝑓,𝐶 Pitch Incremental
PPB 𝜈-deficiency Δ𝑞ሶ 𝑟𝑒𝑓,𝐶,𝑠𝑎𝑡
Model Acceleration Short Period
Protections Adjustment
𝛿𝑠,𝑙𝑜𝑛 Controller Dynamics Plant Model Δ𝑞𝑡𝑟𝑖𝑚,𝑟𝑒𝑓
Δ𝛼𝑙𝑖𝑚,𝑟𝑒𝑓
Δ𝑞ሶ 𝑑𝑓𝑐 Δ𝑛𝑧,𝑡𝑟𝑖𝑚,𝑟𝑒𝑓,𝑆
Δ𝒚𝑟𝑒𝑓 Δ𝛼ሶ 𝑙𝑖𝑚,𝑟𝑒𝑓

Fig. 3 Components of the Physical Reference Model

The reference model controller, described in Section V.D, is designed as a classical PI controller, that shapes the
closed-loop dynamics of the incremental short period model to satisfy the handling quality requirements for the pilot

4
 commands. PI Type controller designs have been extensively used in aircraft control [15, 16] . Therefore, it is well
understood, how certain closed-loop performance requirements and characteristics can be imposed using classical
gain design techniques. However, stability and robustness need to be taken into account in the control design [17, 18],
because of uncertainties in the plant model, making the control design more conservative and limiting the achievable
performance. For the reference model controller however, stability and robustness do not have to be considered because
the incremental short period plant model is exactly known. The pseudo-control deficiency adjustment allows to account
for actuator position and rate saturations. It is described in Section V.C. The equivalent pseudo control dynamics are a
filter containing equivalent linear dynamics corresponding to the estimated actuator dynamics and are presented in
Section V.B. The incremental short period plant model is discussed in Section V.A.

A. Incremental Short Period Plant Model

The classical short period approximation can be written as follows

Δ𝛼¤ 𝑡𝑟 𝑖𝑚 = 𝑍 𝛼 Δ𝛼𝑡𝑟 𝑖𝑚 + Δ𝑞 𝑡𝑟 𝑖𝑚
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

Δ𝑞¤ 𝑡𝑟 𝑖𝑚 = 𝑀 𝛼 Δ𝛼𝑡𝑟 𝑖𝑚 + 𝑀𝑞 Δ𝑞 𝑡𝑟 𝑖𝑚 + 𝑀 𝜂 Δ𝜂𝑡𝑟 𝑖𝑚

(13)
𝑉
Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 = − 𝑍 𝛼 Δ𝛼𝑡𝑟 𝑖𝑚
𝑔
where the force contribution of both the elevator 𝑍 𝜂 and the pitch-rate 𝑍 𝑞 are neglected.
The new proposed linear reference model structure is based on the short period approximation. However, instead of
defining the reference wrt. the trimmed angle of attack 𝛼𝑡𝑟 𝑖𝑚 , as usually done, the angle of attack 𝛼𝐶𝐿,max at 𝐶 𝐿,𝑚𝑎𝑥 , is
used in the angle of attack dynamics. Here 𝐶 𝐿,𝑚𝑎𝑥 is the lift coefficient at the design maximum angle of attack, which
is slightly less than the absolute maximum lift coefficient. In the pitch rate dynamics the contributions of 𝑀 𝛼 and 𝑀𝑞 is
assumed to be perfectly cancelled by inversion, and only the pilot commands are accounted for in the pitch acceleration
dynamics. Further the pitch rate increment is defined, as usually, with respect to the trim pitch rate 𝑞 𝑡𝑟 𝑖𝑚 , see Figure 4.
The proposed structure has the advantage that it makes the reference model independent of trim tables including 𝛼𝑡𝑟 𝑖𝑚 .

Incremental Plant Model

𝑡𝑟𝑖𝑚 𝑙𝑖𝑚
𝑍𝛼
Δ𝑞𝑡𝑟𝑖𝑚,𝑟𝑒𝑓
Δ𝑞ሶ 𝑡𝑟𝑖𝑚,𝑟𝑒𝑓 Δ𝑞ሶ 𝑡𝑟𝑖𝑚,𝑟𝑒𝑓 Δ𝛼𝑙𝑖𝑚,𝑟𝑒𝑓
1
Δ𝑞𝑡𝑟𝑖𝑚 → Δ𝑞𝑙𝑖𝑚 +
1 Δ𝛼𝑙𝑖𝑚,𝑟𝑒𝑓
𝑛𝑧,𝑡𝑟𝑖𝑚,𝑆 𝑠 𝑠 Δ𝑞𝑡𝑟𝑖𝑚,𝑟𝑒𝑓
Δ𝑛𝑧,𝑡𝑟𝑖𝑚,𝑟𝑒𝑓𝑆 Δ𝑛𝑧,𝑡𝑟𝑖𝑚,𝑟𝑒𝑓,𝑆
𝜕𝑛𝑧
Δ𝛼𝑡𝑟𝑖𝑚 ← Δ𝛼𝑙𝑖𝑚 Δ𝛼ሶ 𝑙𝑖𝑚,𝑟𝑒𝑓
𝜕𝛼

Conversion is connected with

(𝑛𝑧,𝑡𝑟𝑖𝑚,𝑆 , 𝑛𝑧,𝐶𝐿,𝑚𝑎𝑥 )

Fig. 4 Sketch of Incremental Short Period Plant Model

Statement 1 (Transformed Short Period) The angle of attack dynamics defined in Eq. (13) in terms of trim values,
can equally be represented by a combination of increments in the trim and limit domain as depicted in Fig. 4 and given
by the following equations:

Δ𝛼¤ lim = 𝑍 𝛼 Δ𝛼lim + Δ𝑞 lim

𝑉 (14)
Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 = − 𝑍 𝛼 Δ𝛼𝑡𝑟 𝑖𝑚
𝑔
with
𝑔

Δ𝛼𝑡𝑟 𝑖𝑚 = Δ𝛼lim − 𝑛 𝑧,𝐶𝐿,max ,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆
𝑉 𝑍𝛼
𝑔
(15)
Δ𝑞 lim = Δ𝑞 𝑡𝑟 𝑖𝑚 − 𝑛 𝑧,𝐶𝐿,𝑚𝑎𝑥 ,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆
𝑉

5
 and the load factor produced by the maximum lift coefficient as
¯ 𝐿,max
𝑞𝑆𝐶
𝑛 𝑧,𝐶𝐿,max ,𝑆 = (16)
𝑚𝑔
The definition of the delta states used in the linear short period dynamics expression in Eq. (14) and Eq. (15) are given
in Table 1

Lim domain Trim domain

Δ𝑞 lim = 𝑞 − 𝑞 𝐶𝐿,max Δ𝑞 𝑡𝑟 𝑖𝑚 = 𝑞 − 𝑞 𝑡𝑟 𝑖𝑚
Δ𝑛 𝑧,lim,𝑆 = 𝑛 𝑧,𝑆 − 𝑛 𝑧,𝐶𝐿,max ,𝑆 Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 = 𝑛 𝑧,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆
Δ𝛼lim = 𝛼 − 𝛼𝐶𝐿,max Δ𝛼𝑡𝑟 𝑖𝑚 = 𝛼 − 𝛼𝑡𝑟 𝑖𝑚
Table 1 Definition of trim and lim domains
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

Derivation. In general, the transformation of variables between the lim and the trim domain can be carried out as
follows:
Δ𝑥𝑙𝑖𝑚 = Δ𝑥 𝑡𝑟 𝑖𝑚 − 𝑥 𝑚𝑎𝑥 + 𝑥 𝑡𝑟 𝑖𝑚 (17)
where Δ𝑥𝑙𝑖𝑚 = 𝑥 − 𝑥 𝑚𝑎𝑥 and Δ𝑥 𝑡𝑟 𝑖𝑚 = 𝑥 − 𝑥 𝑡𝑟 𝑖𝑚 . This holds for all variables. However, in the case of angle of attack,
the trim value 𝛼𝑡𝑟 𝑖𝑚 is unknown. Similarily, for the pitch-rate, the max pitch rate (i.e. 𝑞 𝐶𝐿,max ) is unknown. Therefore,
the trim and lim domain for 𝑞 and 𝛼 is connected over the load factor, where both the trim and lim reference values (i.e.
𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 and 𝑛 𝑧,𝐶𝐿,max ,𝑆 ) are known. Thereby the following relations between angle of attack and load factor increments
is used.
𝑔 1
(𝛼𝐶𝐿,𝑚𝑎𝑥 − 𝛼𝑡𝑟 𝑖𝑚 ) = − (𝑛 𝑧,𝐶𝐿,𝑚𝑎𝑥 ,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 ) (18)
𝑉 𝑍𝛼
Using the steady state relation (i.e. 𝛼¤ = 0) between load factor and pitch rate:
𝑔

(𝑞 − 𝑞 𝑡𝑟 𝑖𝑚 ) = 𝑛 𝑧,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 (19)
𝑉
The steady state pitch rate 𝑞 𝐶𝐿,max at 𝑛 𝑧,𝐶𝐿,𝑚𝑎𝑥 ,𝑆 , can be obtained by:
𝑔
(𝑞 𝐶𝐿,𝑚𝑎𝑥 − 𝑞 𝑡𝑟 𝑖𝑚 ) = (𝑛 𝑧,𝐶𝐿,𝑚𝑎𝑥 ,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 ) (20)
𝑉
Hence, using Eq. (17) , Eq. (18) and Eq. (20)
𝑔

Δ𝛼𝑡𝑟 𝑖𝑚 = Δ𝛼lim − 𝑛 𝑧,𝐶𝐿,max ,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆
𝑉 𝑍𝛼
𝑔
(21)
Δ𝑞 lim = Δ𝑞 𝑡𝑟 𝑖𝑚 − 𝑛 𝑧,𝐶𝐿,𝑚𝑎𝑥 ,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆
𝑉
𝑔

is obtained. Further, substitution of Δ𝛼𝑡𝑟 𝑖𝑚 and Δ𝑞 𝑡𝑟 𝑖𝑚 = Δ𝑞 lim + 𝑉 𝑛 𝑧,𝐶𝐿,𝑚𝑎𝑥 ,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 into the angle of attack
dynamics in Eq. (13) results in
𝑔
𝑔

Δ𝛼¤ 𝑡𝑟 𝑖𝑚 = 𝑍 𝛼 (Δ𝛼lim − 𝑛 𝑧,𝐶𝐿,max ,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 ) + (Δ𝑞 lim + 𝑛 𝑧,𝐶𝐿,𝑚𝑎𝑥 ,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 ) = 𝑍 𝛼 Δ𝛼lim + Δ𝑞 lim (22)
𝑉 𝑍𝛼 𝑉
Since 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 and 𝛼𝐶𝐿,𝑚𝑎𝑥 is independent of time, and additionally for the short period 𝑛 𝑧,𝐶𝐿,max ,𝑆 is independent of
time:  
𝑑 𝑔

Δ𝛼¤ 𝑡𝑟 𝑖𝑚 = Δ𝛼lim − 𝑛 𝑧,𝐶𝐿,max ,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 = Δ𝛼¤ 𝑙𝑖𝑚 (23)
𝑑𝑡 𝑉 𝑍𝛼
holds. By substituting Eq. (23) into Eq. (22) the first equation in Eq. (14) is obtained. Hence the result. 
The reference model is implemented as shown in Fig. 4, using Eq. (14) and Eq. (15).

6
 Pitch Acceleration Dynamics

Δ𝑞ሶ 𝐶,𝑟𝑒𝑓,𝐶,𝑠𝑎𝑡 + 𝜔𝜈 + 1 1 Δ𝑞ሶ 𝑟𝑒𝑓

2𝜁𝜈 𝜔𝜈
- 2𝜁𝜈 - 𝑠 𝑠

Fig. 5 Equivalent Pitch Acceleration Dynamics

B. Equivalent Pitch Acceleration Dynamics

The equivalent pitch acceleration dynamics reflects the dynamics of the actuators, control surface linkage and control
surface dynamics. These are included in the reference dynamics to slow down the reference model response in the same
manner as the aircraft response is slowed down. Note that these dynamics influence only the reference model variables,
and hence the variables used for the error controller. As seen in Fig. 3, these dynamics does not affect the feed forward
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

command Δ𝑞¤𝑟 𝑒 𝑓 , 𝑓 𝑓 .
The dynamics of the pitch acceleration is chosen as second order, as depicted in Fig. 5, to reflect electro-mechanical
actuators. The parameters of the pseudo control dynamics 𝜔 𝜈 and 𝜁 𝜈 are chosen such that the actual actuation response
is approximated.

C. Actuator Rate Constraint Propagation

As described in Section VII, the control allocation directly accounts for surface deflection limits and rate limits.
Due to surface position and rate constraints the commanded body angular accelerations (pseudo control commands)
might not be achievable. As described in Section VII, the control allocation predicts when commands exceed these
constraints, and modifies the commands accordingly. The resulting pseudo command deficiency can be calculated based
on the difference between the desired pseudo control command and the modified pseudo control command resulting
from the limited actuator command increment as described in Section VII.
The knowledge of this deficiency can be used to constrain the reference model, such that unachievable commands do not
progress into the reference model, as such unachievable commands would produce unachievable reference trajectories.
As shown in Fig. 6, the pseudo control command consists of several portions, namely the trim rate contribution 𝑞¤ 𝑡𝑟 𝑖𝑚 ,
Pseudo control 𝜈 Pseudo control 𝜈 Pseudo control 𝜈

Deficiency
Pilot cmd Pilot cmd
Considered
Deficiency Error Ctrl Pilot cmd
Achievable

Remaining Deficiency Remaining Deficiency

Error Ctrl Error Ctrl
Trim rate
Trim rate Trim rate Achievable Achievable

Fig. 6 Incremental pseudo control deficiency

the error controller contribution 𝑞¤ 𝑒𝑐 and the pilot command contribution Δ𝑞¤𝑟 𝑒 𝑓 , 𝑓 𝑓 as shown in Eq. (42).
Since the deficiency is used to constrain the reference model commands Δ𝑞¤𝑟 𝑒 𝑓 , 𝑓 𝑓 only, the following considerations
need to be taken into account when constructing the compensation scheme:

Deficiency considerations
1) The reference model command should not change sign due to the deficiency
2) The reference model command should not get larger in absolute value due to the deficiency
The above described behaviour will result in an intuitive aircraft response to the pilot in case of saturations. I.e. statement
1 ensures that the command will not reverse due to the saturations, and statement 2 ensures that the command will not
be amplified with no pilot input. Figure 6 visualizes three different situations:
• In the first situation the deficiency in the pseudo control command is smaller in amplitude than the pilot portion of the
pseudo control command. Consequently the pilot portion of the pseudo control command in the reference model can

7
 be reduced by the deficiency such that the resulting reference model pseudo control command will correspond to the
achievable.
• In the second situation the trim rate and error controller command are together larger than the achievable command.
Since the pilot command is not to be reversed, the considered deficiency in the reference model only covers the
pilot command, resulting in a remaining deficiency that is not accounted for by the compensation. This behaviour is
ensured by statement 1.
• In the third situation the trim rate and error controller command is again larger than the achievable command, however,
the pilot command is in the opposite direction of those. In that case the pilot command shall stay unchanged, to
prevent an un-intended amplification of the pilot command. This behaviour is ensured by statement 2.
The above described behaviour is realized by the double dynamic saturation depicted in Fig. 7.

𝜈-deficiency adjustment
|Δ𝑞ሶ 𝑟𝑒𝑓,𝑓𝑓 |
Δ𝑞ሶ 𝑟𝑒𝑓,𝑓𝑓 + Δ𝑞ሶ 𝑟𝑒𝑓,𝐶,𝑠𝑎𝑡
-
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

|Δ𝑞ሶ 𝑟𝑒𝑓,𝑓𝑓 | −|Δ𝑞ሶ 𝑟𝑒𝑓,𝑓𝑓 |

𝒆𝑗𝑇 𝝂𝐶,𝑑𝑓𝑐 = 𝑞ሶ 𝐶,𝑑𝑓𝑐

−|Δ𝑞ሶ 𝑟𝑒𝑓,𝑓𝑓 |

Fig. 7 𝝂 deficiency adjustment

D. Reference Model Controller

The purpose of the reference model controller is to shape the dynamics of the reference model, i.e. to shape the desired
dynamics of the pilot response such that the behaviour satisfies all relevant handling quality requirements. This could be
done in many ways. This paper proposes to apply a classical linear PI-type controller to the reference model, because
the blending between command variables can be incorporated in a straight forward manner. Since the reference model
parameters have no uncertainty, stability and robustness requirements for this controller are trivial. The gain design to
obtain a closed-loop behaviour that satisfies relevant handling quality criteria can be obtained from classical techniques,
since the reference model reflect the physics of the aircraft.
The reference model controller is depicted in Fig. 8. It is composed of three units:
• A command scaling which converts stick commands to physical quantities of load factor, pitch rate and angle of attack
expressed as an equivalent incremental load factor command.
• A reference control variable calculation which blends and transforms the reference model variables into a blended
equivalent load factor increment for steady state accuracy.
• A control law to shape the dynamic response.
All three subsystems are discussed in the following.

Reference Model Controller 𝑐𝑎𝑤

𝛿𝑠,𝑙𝑜𝑛 Δ𝑛𝑧,𝑡𝑟𝑖𝑚,𝑒𝑞,𝐶,𝑆 Δ𝑞ሶ 𝑟𝑒𝑓,𝐶

𝐶𝑜𝑚𝑚𝑎𝑛𝑑 𝑅𝑒𝑓𝑒𝑟𝑒𝑛𝑐𝑒 𝑀𝑜𝑑𝑒𝑙
𝑆𝑐𝑎𝑙𝑖𝑛𝑔 𝐶𝑜𝑛𝑡𝑟𝑜𝑙 𝐿𝑎𝑤
Δ𝑛𝑧,𝑡𝑟𝑖𝑚,𝑒𝑞,𝑆
𝒚𝑟𝑒𝑓
𝑅𝑒𝑓𝑒𝑟𝑒𝑛𝑐𝑒 𝐶𝑉
𝐶𝑎𝑙𝑐𝑢𝑙𝑎𝑡𝑖𝑜𝑛

𝑛𝑧,𝑡𝑟𝑖𝑚,𝑆 Δ𝑞𝑡𝑟𝑖𝑚,𝑟𝑒𝑓
Hard protections Δ𝛼𝑙𝑖𝑚,𝑟𝑒𝑓 , Δ𝛼ሶ 𝑙𝑖𝑚,𝑟𝑒𝑓
Upset recovery Δ𝑛𝑧,𝑡𝑟𝑖𝑚,𝑟𝑒𝑓,𝑆 , Δ𝑛ሶ 𝑧,𝑡𝑟𝑖𝑚,𝑟𝑒𝑓,𝑆
𝑛𝑧,𝑡𝑟𝑖𝑚,𝑆

Fig. 8 Reference Model Controller

1. Control Variables (CV) and Reference CV Calculation

The steady state control variable differs depending on stick deflection and airspeed. As Fig. 9 depicts, four regions
based on the stick deflection and the calibrated airspeed, can be distinguished. At high stick deflections, the limiting

8
 flight envelope variable is chosen as the control variable, i.e. angle of attack below the manuever speed 𝑉 𝐴 and load
factor above. At low stick deflections, a blend of pitch rate and load factor is chosen (𝐶 ∗ ).

Structural failure area Stick Deflection

Load Factor 𝛿𝑆
𝑛𝑧
Stall Area Structural damage area 𝑉𝐴
𝑛𝑧,𝑚𝑎𝑥 𝛿𝑆,𝑚𝑎𝑥

𝐶𝐿,𝑚𝑎𝑥
𝐶𝑉: 𝑛𝑧 𝐶𝑉: 𝛼 𝐶𝑉: 𝑛𝑧
𝐶𝑉: 𝛼

1 𝐶𝑉: 𝑞
𝐶𝑉: 𝑞 𝐶𝑉: 𝑛𝑧 𝛾ሶ
𝐶𝑉: 𝑛𝑧 (𝛾)
ሶ
0 𝑉𝑆1𝑔 𝑉𝐴 𝑉𝑁𝐸 𝑉𝐶𝐴𝑆 𝑉𝑆1𝑔 𝑉𝑁𝐸 𝑉𝐶𝐴𝑆
𝐶𝑉: 𝛼 𝐶𝑉: 𝑛𝑧
𝑛𝑧,𝑚𝑖𝑛
Structural damage area
𝐶𝐿,𝑚𝑖𝑛 𝛿𝑆,𝑚𝑖𝑛 𝐶𝑉: 𝛼 𝐶𝑉: 𝑛𝑧
Stall Area 𝑉𝐴,𝑛𝑒𝑔

Fig. 9 Control Variable Regions

Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

Both the load factor and angle of attack have a higher relative degree than the pitch rate. In order to obtain CV’s with
same relative degree, their relative degree is reduced by the following choice of the final control variables
Δ𝐶𝑉 𝛼,𝑙𝑖𝑚 = 𝑘 1 Δ𝛼𝑙𝑖𝑚,𝑟 𝑒 𝑓 + 𝑘 2 Δ𝛼¤ 𝑙𝑖𝑚,𝑟 𝑒 𝑓
Δ𝐶𝑉𝑛𝑧 ,𝑡𝑟 𝑖𝑚 = 𝑘 1 Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑟 𝑒 𝑓 ,𝑆 + 𝑘 2 Δ𝑛¤ 𝑧,𝑡𝑟 𝑖𝑚,𝑟 𝑒 𝑓 ,𝑆 (24)
Δ𝐶𝑉𝑞,𝑡𝑟 𝑖𝑚 = Δ𝑞 𝑡𝑟 𝑖𝑚,𝑟 𝑒 𝑓
Note, that Δ𝛼¤ 𝑙𝑖𝑚 is directly available in the reference model, however, Δ𝑛¤ 𝑧,𝑡𝑟 𝑖𝑚,𝑆 is not, and hence obtained by filtering
Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 . Since Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 is calculated as defined in Eq. (14) and Eq. (15) by
𝑉

Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 = − 𝑍 𝛼 Δ𝛼lim + 𝑛 𝑧,𝐶𝐿,max ,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 (25)
𝑔
it includes the phugoid dynamics over the dynamic pressure measurement used for calculation of 𝑛 𝑧,𝐶𝐿,max ,𝑆 . By filtering
Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 an estimate for Δ𝑛¤ 𝑧,𝑡𝑟 𝑖𝑚,𝑆 is obtained including the influence of the time derivative of the velocity which
otherwise in an model based Δ𝑛¤ 𝑧,𝑡𝑟 𝑖𝑚,𝑆 estimate, based on the short period dynamics would not have been included.
If the 𝐶𝑉 is tracked in steady state, then choosing 𝑘 1 = 1, will lead to steady state accuracy in either Δ𝛼𝑙𝑖𝑚,𝑟 𝑒 𝑓 ,
Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑟 𝑒 𝑓 ,𝑆 or Δ𝑞 𝑡𝑟 𝑖𝑚,𝑟 𝑒 𝑓 , because in steady state 𝛼¤ 𝑙𝑖𝑚 = 𝑛¤ 𝑧 = 0. The control variables (CV’s) are converted, in the
Reference CV Calculation Block, into equivalent load factor increments according to
𝜕𝑛 𝑧
Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝐶𝑉 = Δ𝐶𝑉𝑡𝑟 𝑖𝑚 (26)
𝜕𝐶𝑉∞
in case the respective control variable is in terms of an increment to the trim value, and by
𝜕𝑛 𝑧
Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝐶𝑉 = Δ𝑛 𝑧,𝑙𝑖𝑚,𝑒𝑞,𝐶𝑉 + (𝑛 𝑧,𝐶𝐿,𝑚𝑎𝑥 ,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 ) = Δ𝐶𝑉𝑙𝑖𝑚 + (𝑛 𝑧,𝐶𝐿,𝑚𝑎𝑥 ,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 ) (27)
𝜕𝐶𝑉∞
in case the respective control variable is in terms of an increment to the lim value, where 𝐶𝑉∞ is the steady state CV, i.e.
𝛼, 𝑛 𝑧 or 𝑞. By this consideration, it is ensured that the steady state is obtained with respect to the commanded variable
whether it being as an increment to the trim or the lim domain. Conversion of the different CVs into an equivalent load
factor allows to blend the different control variables. The final equivalent load factor increment is obtained as a blend of
the respective CV equivalent load factors according to
Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝑆 = 𝑏 𝛼 Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞, 𝛼 + 𝑏 𝑛 Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝑛𝑧 + 𝑏 𝑞 Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝑞
 
−𝑉 𝑍 𝛼 𝑉 (28)
= 𝑏𝛼 Δ𝐶𝑉 𝛼,lim + 𝑛 𝑧,𝐶𝐿,max ,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 + 𝑏 𝑛𝑧 Δ𝐶𝑉𝑛𝑧,𝑡𝑟 𝑖𝑚 + 𝑏 𝑞 Δ𝐶𝑉𝑞,𝑡𝑟 𝑖𝑚
𝑔 𝑔
The blend is created by blending variables 𝑏 𝛼 , 𝑏 𝑞 , 𝑏 𝑛𝑧 ∈ [0, 1] that describe the amount of each CV in the final blend,
with the additional constraint:
𝑏 𝛼 + 𝑏 𝑞 + 𝑏 𝑛𝑧 = 1 (29)
The blending variables are chosen using linear blending according to Fig. 9, together with a selection function which
engages when the measured angle of attack or measured load factor are close to their respective limit. The selection
function ensures that in case a measured control variable is close to its limit it will be prioritized in the blend.

9
 2. Command Scaling
The stick scaling should reflect the following desired characteristics of the pilot command path:

Objectives
1) The stick scaling shall reflect the maneuver margin up to the limiting value of the flight envelope at the current flown
point, given by either the maximum load factor defined by structural limitations or aerodynamic limit given by the
maximum allowed lift coefficient.
2) There shall be a minimum stick gradient in terms of incremental angle of attack per stick deflection.
3) The stick gradient shall be linear where possible, and only gradual changes in stick gradient should be present if
necessary.
4) There shall be a natural speed stable tendency at low speed. This will prevent inadvertently commanding high angles
of attack with no pilot command at decreasing speeds. Additionally it provides a natural flare characteristic, as at
low speed, gradual increasing pilot command will be required to maintain a constant load factor when the speed
reduces during the flare.
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

The following choice of stick scaling, will obtain the above mentioned objectives: The stick scaling will be linear for the
range above a 𝛿 𝑠,𝑏,𝑙𝑜𝑛 = −0.25, which is the break stick deflection where the stick scaling below this value is changed to
match the requirement of reaching the negative maneuver limit at full stick forward. For stick deflections larger than
𝛿 𝑠,𝑙𝑜𝑛 > 𝛿 𝑠,𝑏,𝑙𝑜𝑛 , the equivalent load factor command increment Δ𝑛 𝑧,𝑒𝑞,𝑡𝑟 𝑖𝑚,𝐶,𝑆 = 𝑛 𝑧,𝑒𝑞,𝐶,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 , with 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆
given by Eq. (3), is linear until the maximum load factor value 𝑛 𝑧,𝑚𝑎𝑥 :

𝜕𝑛 𝑧,𝑒𝑞,𝐶
Δ𝑛 𝑧,𝑒𝑞,𝑡𝑟 𝑖𝑚,𝐶,𝑆 = (𝛿 𝑠,𝑙𝑜𝑛 − 0) (30)
𝜕𝛿 𝑠,𝑙𝑜𝑛

where the stick gradient is defined by:

𝜕𝑛 𝑧,𝑒𝑞,𝐶,𝑆
= Δ𝑛 𝑧,max,𝑆 = 𝑛 𝑧,𝑚𝑎𝑥,𝑆 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 (31)
𝜕𝛿 𝑠,𝑙𝑜𝑛

and where

𝑛 𝑧,𝑚𝑎𝑥,𝑆 = min 𝑛 𝑧,𝐶𝐿,max ,𝑆 , 𝑛 𝑧,𝑠𝑡𝑟 𝑢𝑐𝑡 ,max,𝑆 (32)
with 𝑛 𝑧,𝐶𝐿,max ,𝑆 given by Eq. (16) and the structural load factor limit 𝑛 𝑧,𝑠𝑡𝑟 𝑢𝑐𝑡 ,max,𝑆 .

Objective 1 is satisfied because the stick is scaled from 𝛿 𝑠,𝑙𝑜𝑛 ∈ [−1, 1], such that the commanded incremental
load factor from the trim load factor, Δ𝑛 𝑧,𝑒𝑞,𝑡𝑟 𝑖𝑚,𝐶,𝑆 will equal Δ𝑛 𝑧,max,𝑆 for a full stick deflection, i.e. 𝛿 𝑠 = 1.
Objective 2 is satisfied at the speed range 𝐿𝑆 = [𝑉𝑡𝑟 𝑖𝑚, 𝛼𝐶,max , 𝑉𝑡𝑟 𝑖𝑚, 𝛼𝐶,min ]. At this speed range the resulting minimum
stick gradient in terms of incremental angle of attack per stick deflection is, according to Appendix 7, given by the
difference between the angle of attack 𝛼𝑚𝑎𝑥 corresponding to 𝑛 𝑧,𝑚𝑎𝑥 , which at low speed will be 𝛼𝐶𝐿,𝑚𝑎𝑥 and the design
parameter 𝛼𝐶,min :
𝜕𝛼𝑒𝑞,𝐶

= 𝛼max − 𝛼𝐶,min (33)
𝜕𝛿 𝑠 𝑉 ∈𝐿𝑆
The velocities 𝑉𝑡𝑟 𝑖𝑚, 𝛼𝐶,min and 𝑉𝑡𝑟 𝑖𝑚, 𝛼max are the trim speed corresponding to a trim angle of attack of 𝛼𝐶,min and 𝛼max
respectively.
Objective 4 is satisfied at speeds 𝑉 ∈ 𝐿𝑆 as defined above, because at 0 stick deflection, the equivalent angle of attack
command will be 𝛼𝐶,min , because

𝜕𝛼𝑒𝑞,𝐶
𝛼𝐶 = 𝛼max − (1 − 0) = 𝛼𝐶,𝑚𝑖𝑛 (34)
𝜕𝛿 𝑠 𝑉 ∈𝐿𝑆

using the relation for angle of attack command in Appendix IX.A. Since 𝛼𝐶,min is lower than the trim angle of attack in
this region this command will result in a pitch down, and hence a pseudo speed stable aircraft, with 𝛼𝐶,min as the new
pseudo trim angle of attack.

For stick deflections lower than 𝛿 𝑠,𝑏,𝑙𝑜𝑛 = −0.25, the command is connected with a smooth transition to a lin-
ear domain towards the lower boundary of the envelope.

10
 3. Reference Model Control Law
The reference model controller, depicted in Fig. 10, is designed as a PI type controller with feed forward. The reference
model variables Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑟 𝑒 𝑓 ,𝑆 and Δ𝑞 𝑡𝑟 𝑖𝑚,𝑟 𝑒 𝑓 are used for the feedback. Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑟 𝑒 𝑓 ,𝑆 is converted to an equivalent
Δ𝛼𝑡𝑟 𝑖𝑚,𝑟 𝑒 𝑓 using Eq. (14) and Eq. (15). Δ𝑞 𝑡𝑟 𝑖𝑚,𝑟 𝑒 𝑓 and Δ𝛼𝑡𝑟 𝑖𝑚,𝑟 𝑒 𝑓 are then fed back with proportional gains 𝐾𝑞
and 𝐾 𝛼 . The longitudinal stick input 𝛿 𝑠,𝑙𝑜𝑛 is converted by the command stick scaling into an equivalent load factor
command Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝐶,𝑆 , which is described in detail in Section V.D.2. The command is filtered by a shaping filter 𝐹,
to achieve additional handling quality criterion’s such as attitude drop back, before it is fed forward with feed-forward
gain 𝐻 to cancel the integrator closed-loop pole. An integral action leads to steady state accuracy of the control variable
(CV) expressed in terms of an equivalent load factor defined in Eq. (28) and Eq. (24). Using the control error 𝑒, given

Reference Model Controller 𝑐𝑎𝑤

𝐻
𝛿𝑠,𝑙𝑜𝑛 Δ𝑛𝑧,𝑡𝑟𝑖𝑚,𝑒𝑞,𝐶,𝑆 - Δ𝑞ሶ 𝑟𝑒𝑓,𝐶
𝐶𝑜𝑚𝑚𝑎𝑛𝑑 + + 1 + +
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

𝐹 𝐾𝐼
𝑆𝑐𝑎𝑙𝑖𝑛𝑔 - 𝑠 -
𝐾𝑞 Δ𝑞𝑡𝑟𝑖𝑚,𝑟𝑒𝑓
Δ𝑛𝑧,𝑡𝑟𝑖𝑚,𝑒𝑞,𝑆 +

𝑅𝑒𝑓𝑒𝑟𝑒𝑛𝑐𝑒 𝐶𝑉 𝒚𝑟𝑒𝑓
𝐾𝛼 Δ𝛼𝑡𝑟𝑖𝑚,𝑟𝑒𝑓
𝐶𝑎𝑙𝑐𝑢𝑙𝑎𝑡𝑖𝑜𝑛

𝑛𝑧,𝑡𝑟𝑖𝑚,𝑆 Δ𝑞𝑡𝑟𝑖𝑚,𝑟𝑒𝑓
Hard protections Δ𝛼𝑙𝑖𝑚,𝑟𝑒𝑓 , Δ𝛼ሶ 𝑙𝑖𝑚,𝑟𝑒𝑓
Upset recovery Δ𝑛𝑧,𝑡𝑟𝑖𝑚,𝑟𝑒𝑓,𝑆 , Δ𝑛ሶ 𝑧,𝑡𝑟𝑖𝑚,𝑟𝑒𝑓,𝑆
𝑛𝑧,𝑡𝑟𝑖𝑚,𝑆

Fig. 10 Reference Model Controller

by

𝑒 =𝐹 (𝑠)Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝐶,𝑆 − Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝑆 (35)

where Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝐶,𝑆 is calculated by Eq. (30) and Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝑆 is calculated by Eq. (28), the control law can be
expressed as:
1
Δ𝑞¤𝑟 𝑒 𝑓 ,𝐶 =𝐻𝐹 (𝑠)Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝐶,𝑆 + (𝐾 𝐼 𝑒 + 𝑐 𝑎𝑤 )
𝑠 (36)
− 𝐾 𝛼 Δ𝛼𝑡𝑟 𝑖𝑚,𝑟 𝑒 𝑓 − 𝐾𝑞 Δ𝑞 𝑡𝑟 𝑖𝑚,𝑟 𝑒 𝑓

where Δ𝛼𝑡𝑟 𝑖𝑚,𝑟 𝑒 𝑓 is obtained by Eq. (15) and 𝑐 𝑎𝑤 is an integrator anti wind protection calculated by the phase plane
based protections module depicted in Fig. 3. It is shown in Appendix IX.F that the following choice of 𝑘 1 , 𝑘 2 from Eq.
(24),
𝑘1 = 1
(37)
𝑘 2 = −1/𝑍 𝛼
will lead to identical feedback gains 𝐾 𝐼 , 𝐾𝑞 and 𝐾 𝛼 and feed forward gain 𝐻 for all three 𝐶𝑉 cases when performing a
pole placement on the short period approximation, resulting in identical closed-loop integrator and short period poles
for all CV’s. Therefore, the blending can be performed straight forward without consideration of adapting the reference
model control law gains in the different 𝐶𝑉 regions. As the feed-forward gain 𝐻 is chosen to cancel the integrator pole,
the pre-filter 𝐹 can also be designed independently of the 𝐶𝑉 region.

VI. Error Controller

The error controller is depicted in Fig. 11. It can be divided into two main parts: The error dynamics calculation and a
non-linear dynamic inversion type of conversion from the 𝑛¥ 𝑧 domain into the 𝑞¤ domain to generate a pitch acceleration
command for the INDI controller.

11
 Error Controller
Δ𝑞ሶ 𝑟𝑒𝑓,𝑓𝑓
Δ𝑞𝑡𝑟𝑖𝑚,𝑟𝑒𝑓 Δ𝑞𝑡𝑟𝑖𝑚 𝑒𝑛𝑧,𝑒𝑞,𝑏𝑙 Calculation 𝑘𝑝
Δ𝛼𝑙𝑖𝑚,𝑟𝑒𝑓 𝑒𝑛ሷ 𝑧,𝑒𝑞,𝑏𝑙 𝑞ሶ 𝑒𝑐 + + 𝑞ሶ 𝐶
Δ𝛼ሶ 𝑙𝑖𝑚 𝑛ሷ 𝑧 → 𝑞ሶ
Δ𝛼ሶ 𝑙𝑖𝑚,𝑟𝑒𝑓 Δ𝛼𝑙𝑖𝑚 +
Δ𝑛𝑧,𝑡𝑟𝑖𝑚,𝑟𝑒𝑓,𝑆 Δ𝑛𝑧,𝑡𝑟𝑖𝑚,𝑆 𝑒𝑛ሶ 𝑧,𝑒𝑞,𝑏𝑙 Calculation 𝑘𝑑
𝑒𝑛ሶ 𝑧,𝑒𝑞,𝑏𝑙
𝑞ሶ 𝑡𝑟𝑖𝑚

Fig. 11 Error controller

A. Error dynamics
The error dynamics are chosen as second order dynamics in terms of blended equivalent load factor errors according to:

𝑒 𝑛¥ 𝑧,𝑒𝑞,𝑏𝑙 = Δ𝑛¥ 𝑧,𝑒𝑞,𝑏𝑙,𝑟 𝑒 𝑓 − Δ𝑛¥ 𝑧,𝑒𝑞,𝑏𝑙 = −𝑘 𝑝 𝑒 𝑛𝑧,𝑒𝑞,𝑏𝑙 − 𝑘 𝑑 𝑒 𝑛𝑧,𝑒𝑞,𝑏𝑙

¤ (38)

where 𝑒 𝑛𝑧,𝑒𝑞,𝑏𝑙 and 𝑒 𝑛𝑧,𝑒𝑞,𝑏𝑙 are the blended errors of the load factor equivalents:
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

𝑒 𝑛𝑧,𝑒𝑞,𝑏𝑙 = 𝑏 𝛼 𝑒 𝑛𝑧,𝑒𝑞, 𝛼 + 𝑏 𝑞 𝑒 𝑛𝑧,𝑒𝑞,𝑞 + 𝑏 𝑛𝑧 𝑒 𝑛𝑧 ,𝑒𝑞,𝑛𝑧 (39)

𝑒 𝑛𝑧,𝑒𝑞,𝑏𝑙
¤ = 𝑏 𝛼 𝑒 𝑛𝑧,𝑒𝑞,
¤ 𝛼 + 𝑏 𝑞 𝑒 𝑛𝑧,𝑒𝑞,𝑞
¤ + 𝑏 𝑛𝑧 𝑒 𝑛𝑧,𝑒𝑞,𝑛𝑧
¤ (40)
with the errors of the load factor equivalents given in Table 2. Using the short period relations in Eq. (13), the following

Error Error derivative

𝜕𝑛𝑧 𝜕𝑛¤ 𝑧
𝑒 𝑛𝑧,𝑒𝑞, 𝛼 = 𝜕𝛼 (Δ𝛼𝑟 𝑒 𝑓 ,lim − Δ𝛼lim ) 𝑒 𝑛𝑧,𝑒𝑞,
¤ 𝛼 = 𝜕 𝛼¤ (Δ𝛼¤ 𝑟 𝑒 𝑓 ,lim − Δ𝛼¤ lim )

𝜕𝑛𝑧 𝜕𝑛¤ 𝑧
𝑒 𝑛𝑧,𝑒𝑞,𝑞 = 𝜕𝑞 (Δ𝑞 𝑟 𝑒 𝑓 ,𝑡𝑟 𝑖𝑚 − Δ𝑞 𝑡𝑟 𝑖𝑚 ) 𝑒 𝑛𝑧,𝑒𝑞,𝑞
¤ = 𝜕𝑞 (Δ𝑞 𝑟 𝑒 𝑓 ,𝑡𝑟 𝑖𝑚 − Δ𝑞 𝑡𝑟 𝑖𝑚 )

𝑛¤ 𝑧
𝑒 𝑛𝑧 ,𝑒𝑞,𝑛𝑧 = (Δ𝑛 𝑧,𝑟 𝑒 𝑓 ,𝑡𝑟 𝑖𝑚 − Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚 ) 𝑒 𝑛𝑧,𝑒𝑞,𝑛𝑧
¤ = 𝜕𝜕𝑞 (Δ𝑞 𝑟 𝑒 𝑓 ,𝑡𝑟 𝑖𝑚 − Δ𝑞 𝑡𝑟 𝑖𝑚 )
𝜕𝑛¤ 𝑧
+ 𝜕𝑛𝑧 (Δ𝑛 𝑧,𝑟 𝑒 𝑓 ,𝑡𝑟 𝑖𝑚 − Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚 )

Table 2 Equivalent load factor errors

expression for the load factor derivative can be derived:

−𝑉 −𝑉
Δ𝑛¤ 𝑧 = 𝑍 𝛼 Δ𝛼¤ = (𝑍 𝛼 Δ𝑛 𝑧 + 𝑍 𝛼 Δ𝑞) (41)
𝑔 𝑔
The partial derivatives in Table 2 are summarized in Table 3.

𝑛𝑧 𝑛¤ 𝑧

𝜕𝑛𝑧 −𝑉 𝜕𝑛¤ 𝑧 −𝑉
𝜕𝛼 = 𝑔 𝑍𝛼 𝜕 𝛼¤ = 𝑔 𝑍𝛼

𝜕𝑛𝑧 𝑉 𝜕𝑛¤ 𝑧 −𝑉
𝜕𝑞 = 𝑔 𝜕𝑞 = 𝑔 𝑍𝛼

𝜕𝑛¤ 𝑧
= 𝑍𝛼 𝜕𝑛𝑧
Table 3 Equivalent load factor errors

12
 B. Pitch rate acceleration command conversion
The purpose of the pitch rate acceleration command conversion block is to transform the error dynamics which are
according to Section VI.A formulated in terms of blended equivalent load factor errors into a pitch rate acceleration
command that is used in the INDI control law. The conversion is given by Statement 2.
Statement 2 The absolute pitch acceleration can be split in contributions of reference model feed forward, trim rate
and error controller contributions as follows:

𝑞¤𝐶 = Δ𝑞¤𝑟 𝑒 𝑓 , 𝑓 𝑓 + 𝑞¤ 𝑡𝑟 𝑖𝑚 + 𝑞¤ 𝑒𝑐 (42)

with  
𝑔 𝑒 𝑛¥ 𝑧,𝑒𝑞,𝑏𝑙
𝑞¤ 𝑒𝑐 = − 𝑒 𝑛¤ 𝑧,𝑒𝑞,𝑏𝑙 (43)
𝑉 𝑍𝛼
and with 𝑒 𝑛¤ 𝑧,𝑒𝑞,𝑏𝑙 and 𝑒 𝑛¥ 𝑧,𝑒𝑞,𝑏𝑙 given by Eq. (40) and Eq. (38). The nonlinear compensation term 𝑞¤ 𝑡𝑟 𝑖𝑚 is obtained by
filtering 𝑞 𝑡𝑟 𝑖𝑚 given by Eq. (10).
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

Derivation. Differentiation of Eq. (41) results in

 
𝑉
Δ𝑛¥ 𝑧 = 𝑍 𝛼 Δ𝑛¤ 𝑧 − Δ𝑞¤ (44)
𝑔

By solving Eq. (44) for Δ𝑞¤ the following relations are obtained
 
𝑔 Δ𝑛¥ 𝑧
Δ𝑞¤ = Δ𝑛¤ 𝑧 − (45)
𝑉 𝑍𝛼
 
𝑔 Δ𝑛¥ 𝑧,𝑟 𝑒 𝑓
Δ𝑞¤𝑟 𝑒 𝑓 = Δ𝑛¤ 𝑧,𝑟 𝑒 𝑓 − (46)
𝑉 𝑍𝛼
Using
Δ𝑛¥ 𝑧 = Δ𝑛¥ 𝑧,𝑟 𝑒 𝑓 − 𝑒 𝑛¥ 𝑧 (47)
which is obtained from Eq. (38), results in
   
𝑔 Δ𝑛¥ 𝑧 𝑔 𝑔 𝑔 Δ𝑛¥ 𝑧,𝑟 𝑒 𝑓 − 𝑒 𝑛¥ 𝑧 𝑔 𝑔
Δ𝑞¤ = Δ𝑛¤ 𝑧 − + Δ𝑛¤ 𝑧,𝑟 𝑒 𝑓 − Δ𝑛¤ 𝑧,𝑟 𝑒 𝑓 = Δ𝑛¤ 𝑧 − + Δ𝑛¤ 𝑧,𝑟 𝑒 𝑓 − Δ𝑛¤ 𝑧,𝑟 𝑒 𝑓 (48)
𝑉 𝑍𝛼 𝑉 𝑉 𝑉 𝑍𝛼 𝑉 𝑉

Equation (48) can be rearranged such that

 
𝑔 Δ𝑛¥ 𝑧,𝑟 𝑒 𝑓 𝑔
Δ𝑞¤ = Δ𝑛¤ 𝑧,𝑟 𝑒 𝑓 − + (𝑍 𝛼 (Δ𝑛¤ 𝑧 − Δ𝑛¤ 𝑧,𝑟 𝑒 𝑓 ) + 𝑒 𝑛¥ 𝑧 ) (49)
𝑉 𝑍𝛼 𝑍 𝛼𝑉

Using Eq. (46) and 𝑒 𝑛¤ 𝑧 = Δ𝑛¤ 𝑧,𝑟 𝑒 𝑓 − Δ𝑛¤ 𝑧 results in

𝑔 𝑒 𝑛¥ 𝑧
Δ𝑞¤ = Δ𝑞¤𝑟 𝑒 𝑓 + ( − 𝑒 𝑛¤ 𝑧 ) (50)
𝑉 𝑍𝛼
Finally with Δ𝑞¤ = 𝑞¤ − 𝑞¤ 𝑡𝑟 𝑖𝑚
𝑔 𝑒 𝑛¥ 𝑧
𝑞¤ = Δ𝑞¤𝑟 𝑒 𝑓 + 𝑞¤ 𝑡𝑟 𝑖𝑚 + ( − 𝑒 𝑛¤ 𝑧 ) (51)
𝑉 𝑍𝛼
is obtained. Hence the result. 

VII. Control Allocation and INDI

The control allocation seeks to distribute the pseudo control commands 𝝂𝐶 = ( 𝑝¤𝐶 , 𝑞¤𝐶 , 𝑟¤𝐶 )𝑇 to control surface
commands. An incremental pseudo control command is formed by subtracting the current estimated pseudo controls as
shown in Fig. 12.
Δ𝝂𝐶 = 𝝂𝐶 − 𝝂ˆ (52)

13
 A redistributed pseudo inverse, [19], distributes Δ𝜈𝐶 to an incremental actuator command Δ𝒖 𝑅𝑃𝐼 taking into account
actuator position and rate constraints. This constitutes an INDI type control law[5]. A null-space transition algorithm,
[20], imposes additional constraints on the actuator commands, to fulfil certain secondary objectives, by calculating an
addition Δ𝒖 𝑛𝑡 to the incremental command.
Δ𝒖𝐶 = Δ𝒖 𝑅𝑃𝐼 + Δ𝒖 𝑛𝑡 (53)
The incremental command Δ𝒖𝐶 is transformed to an absolute actuator command Δ𝒖 𝑐𝑚𝑑 using a model of the actuators.
Before the command is transmitted to the actuators, notch filters are applied in order to prevent an excitation of
aero-elastic modes of the aircraft and control surfaces. The resulting Δ𝒖𝐶 will hence correspond, in situations where

Control Allocation
Actuators
𝝂ො Model
𝝂ො 𝚫𝒖𝑹𝑷𝑰 𝚫𝒖𝑪 𝒖𝒄𝒎𝒅 𝒖𝒄𝒎𝒅
𝝂𝑪 - 𝚫𝝂𝑪 + +
𝝂𝐶 RPI Notch Filters 𝝂𝑪,𝒅𝒇𝒄
+ + +
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

𝚫𝒖𝒏𝒕
NT
𝚫𝝂𝑪,𝒔𝒂𝒕
𝑩 −
𝝂𝑪,𝒅𝒇𝒄
+

Fig. 12 Block diagram of control allocation

the desired pseudo commands Δ𝝂𝐶 are not achievable due to actuator constraints, to a modified pseudo command
Δ𝝂𝐶,𝑠𝑎𝑡 = 𝑩Δ𝒖𝐶 . In Fig. 12, it is shown that the deficiency in the pseudo controls command Δ𝝂𝐶,𝑑 𝑓 𝑐 can be calculated
based on the limited actuator command increment by
𝝂𝐶,𝑑 𝑓 𝑐 = Δ𝝂𝐶 − 𝑩Δ𝒖𝐶 (54)
where 𝑩 is the matrix that relates the incremental pseudo- control commands to the incremental actuator commands
and Δ𝝂𝐶 are the desired pseudo- commands and Δ𝝂𝐶,𝑠𝑎𝑡 = 𝑩Δ𝒖𝐶 are the modified, actual pseudo- commands. This
deficiency represents the unachievable body angular acceleration due to actuator and actuator rate constraints.

VIII. Conclusion
In this paper several new ideas were introduced related to the synthesis of a piloted flight control law using incremental
non-linear dynamic inversion. A physical incremental reference model was proposed which eases the design of good
handling qualities. A reference model control structure was proposed to realize the desired dynamics, using a blending
scheme, in order to adjust the control variables over the envelope and stick deflection to support the operation of the
aircraft with the objectives relevant in a given situation. A method that introduces speed stability in a transparent way at
low speeds that relieves a dedicated flare law, was proposed, together with a natural protection against inadvertent high
angle of attack commands with reducing speeds. An algorithm for adjusting reference model commands in case of
actuator saturations and rate saturations was proposed, showing intuitive behaviour. A relation for connecting the trim
contribution, reference model and error contribution was proposed. Future work will focus on the evaluation of the
concept within a piloted simulator test campaign. Furthermore, it is planned to investigate an optimal control based flight
control clearance using the methods described in [21]. To investigate the control design with regard to uncertainties,
subset simulation could be performed using the Tool Box presented in [22, 23]. In case the evaluation of each sample is
computional expensive, the use of surrogate models as presented in [24] could be considered. Further research will
focus on implementing the control law on a flying research platform to demonstrate the concept in practice.

IX. Appendix

A. Derivation of of stick scaling with respect to angle of attack command

Statement 3 (stick scaling with respect to angle of attack command) Let the stick scaling be given by Eq. (30), i.e.
𝜕𝑛 𝑧
Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝐶 = 𝛿 𝑠,𝑙𝑜𝑛 = Δ𝑛 𝑧,𝑚𝑎𝑥 𝛿 𝑠,𝑙𝑜𝑛 (55)
𝜕𝛿 𝑠,𝑙𝑜𝑛

14
 where
Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝐶 = 𝑛 𝑧,𝐶 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚 (56)
and
𝜕𝑛 𝑧
= Δ𝑛 𝑧,𝑚𝑎𝑥 = 𝑛 𝑧,𝑚𝑎𝑥 − 𝑛 𝑧,𝑡𝑟 𝑖𝑚 (57)
𝜕𝛿 𝑠,𝑙𝑜𝑛
with 𝑛 𝑧,𝑚𝑎𝑥 given by Eq. (32). Then the corresponding commanded angle of attack can be represented by

𝜕𝛼
𝛼𝐶 = 𝛼max − 1 − 𝛿 𝑠,𝑙𝑜𝑛 (58)
𝜕𝛿 𝑠,𝑙𝑜𝑛

with
𝜕𝛼 𝜕𝛼
= Δ𝑛 𝑧,𝑚𝑎𝑥 (59)
𝜕𝛿 𝑠,𝑙𝑜𝑛 𝜕𝑛 𝑧

Derivation. We define
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

Δ𝑛 𝑧,𝑙𝑖𝑚,𝑚𝑎𝑥,𝐶 = 𝑛 𝑧,𝐶 − 𝑛 𝑧,𝑚𝑎𝑥 = Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝐶 − Δ𝑛 𝑧,𝑚𝑎𝑥 (60)

We insert Eq. (55) into Eq. (60), resulting in

Δ𝑛 𝑧,𝑙𝑖𝑚,𝑚𝑎𝑥,𝐶 = Δ𝑛 𝑧,𝑚𝑎𝑥 𝛿 𝑠,𝑙𝑜𝑛 − Δ𝑛 𝑧,𝑚𝑎𝑥 = −(1 − 𝛿 𝑠,𝑙𝑜𝑛 )Δ𝑛 𝑧,𝑚𝑎𝑥 (61)

Incorporationg the relations

𝜕𝑛 𝑧
Δ𝑛 𝑧,𝑙𝑖𝑚,𝑚𝑎𝑥,𝐶 = Δ𝛼𝑙𝑖𝑚,𝑚𝑎𝑥,𝐶 (62)
𝜕𝛼
with
Δ𝛼𝑙𝑖𝑚,𝑚𝑎𝑥,𝐶 = 𝛼𝐶 − 𝛼𝑚𝑎𝑥 (63)
and
𝜕𝑛 𝑧 𝜕𝛼
Δ𝑛 𝑧,𝑚𝑎𝑥 = (64)
𝜕𝛼 𝜕𝛿 𝑠,𝑙𝑜𝑛
which was obtained from Eq. (57), results in

𝜕𝑛 𝑧 𝜕𝑛 𝑧 𝜕𝛼
Δ𝛼𝑙𝑖𝑚,𝑚𝑎𝑥,𝐶 = −(1 − 𝛿 𝑠,𝑙𝑜𝑛 ) (65)
𝜕𝛼 𝜕𝛼 𝜕𝛿 𝑠,𝑙𝑜𝑛

Finally with Eq. (63)

𝜕𝛼
𝛼𝐶 = 𝛼max − 1 − 𝛿 𝑠,𝑙𝑜𝑛 (66)
𝜕𝛿 𝑠,𝑙𝑜𝑛
is obtained. Hence, the result. 

B. Equations of motion in aerodynamic (A) frame

Statement 4 (EOM in 𝐴 frame with uniform, non-steady wind) The equations of motion for a rigid body aircraft
given by the aerodynamic angle of attack 𝛼 𝐴 and sideslip angle 𝛽 𝐴, the magnitude of the aerodynamic velocity 𝑉 𝐴𝐺 , the


𝐺 𝐸
aerodynamic frame forces (𝑭) 𝐴 and the wind velocity 𝑽𝑊 is given by:

𝐸𝐴
𝐸
𝑉¤ 𝐴𝐺 / 𝑉 𝐴𝐺 Í 0   𝐸𝑂
© ª (𝑭) 𝐴 ©­ ª
­ ® =
𝐸 −­ ® − 1
𝑴 𝐴𝐵 𝑴 𝐵𝑂 𝑽¤ 𝐺
­ 𝛽¤ 𝐴 ® 𝑟 𝐾 cos 𝛼 𝐴 − 𝑝 𝐾 sin 𝛼 𝐴 ® 𝐸 𝑊
𝑚 𝑉 𝐴𝐺 𝑉 𝐴𝐺 𝑂
« 𝛼¤ 𝐴 cos 𝛽 𝐴 ¬𝐴 « 𝐾
𝑝 cos 𝛼 𝐴 sin 𝛽 𝐴 − 𝑞 𝐾 cos 𝛽 𝐴 + 𝑟 𝐾 sin 𝛼 𝐴 sin 𝛽 𝐴 ¬𝐴
(67)
With the following assumptions:
• Non-rotating earth: 𝝎 𝐾
𝐼𝐸 = 0

• Flat earth: 𝝎 𝐾 = 0, hence 𝝎 𝐾

𝐸𝑂 𝐸 𝐵 = 𝝎𝑂𝐵
𝐾
• Fixed mass: 𝑑𝑡 = 0.
𝑑𝑚

15
 Derivation: The translational equations of motion for a rigid body aircraft with the assumptions above are given by:
 𝐸𝐵  𝐸 Í 𝑭
¤
𝑽𝐾𝐺
+ 𝝎 𝐾 × 𝑽𝐾𝐺 =
𝑂𝐵
(68)
𝑚
The relation between kinematic velocity, aerodynamic velocity and wind velocity is given by:
 𝐸  𝐸  𝐸
𝑽𝐾𝐺 = 𝑽 𝐴𝐺 + 𝑽𝑊 𝐺
(69)

The cross product in Eq. (68) can be expanded with the wind and aerodynamic velocity:
 𝐸    𝐸  𝐸  𝐸
𝐸
𝑂𝐵
𝝎 𝐾 × 𝑽𝐾 𝐺 𝑂𝐵
= 𝝎𝐾 × 𝑽𝐴 𝐺
+ 𝑽𝑊𝐺
= 𝝎𝑂𝐵
𝐾 × 𝑽 𝐺
𝐴 + 𝝎 𝑂𝐵
𝐾 × 𝑽𝑊
𝐺
(70)

The kinematic acceleration in Eq. (68) can likewise be expanded:

 𝐸𝐵  𝐸𝐵  𝐸𝐵
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

𝑽¤ 𝐾𝐺 = 𝑽¤ 𝐴𝐺 + 𝑽¤ 𝑊
𝐺
(71)

The wind term derivative can additionally be expressed by its derivative with respect to the NED:
 𝐸𝐵   𝐸𝑂  𝐸
𝑽¤ 𝑊
𝐺
= 𝑽¤ 𝑊
𝐺
+ 𝝎𝐾𝐵𝑂
× 𝑽𝑊 𝐺
(72)

Using these relations, the EOM can be formulated:

 𝐸𝐵 Í

𝐺 𝐸
𝑽¤ 𝐾𝐺 = 𝑚𝑭 − 𝝎𝑂𝐵
𝐾 × 𝑽𝐾 ⇒
 𝐸𝐵 Í

 𝐸𝐵
𝐺 𝐸 𝐺 𝐸
𝑽¤ 𝐴𝐺 = 𝑚𝑭 − 𝝎𝑂𝐵 × 𝑽 − 𝝎 𝑂𝐵
× 𝑽 − ¤𝐺
𝑽 ⇒
 𝐸𝐵 Í   𝐸𝑂
𝐾 𝐴 𝐾 𝑊 𝑊


𝐺 𝐸


𝐺 𝐸


𝐺 𝐸
(73)
𝑽¤ 𝐴𝐺 = 𝑚𝑭 − 𝝎𝑂𝐵
𝐾 × 𝑽𝐴 − 𝝎𝑂𝐵
𝐾 × 𝑽𝑊 − 𝑽¤ 𝑊𝐺 𝐵𝑂
− 𝝎𝐾 × 𝑽𝑊 ⇒
 𝐸𝐵 Í
  𝐸𝑂
𝐺 𝐸
𝑽¤ 𝐴𝐺 = 𝑚𝑭 − 𝝎𝑂𝐵
𝐾 × 𝑽𝐴 − 𝑽¤ 𝑊
𝐺

These equations are though expressed in the 𝐴 frame: Expanding the first term in equation 70 the A frame:
𝐸


©
cos 𝛼 𝐴 cos 𝛽 𝐴 sin 𝛽 𝐴 sin 𝛼 𝐴 cos 𝛽 𝐴
ª© 𝐾
𝑝
ª
𝑉𝐺
© 𝐴 ª
­
𝑴 𝐴𝐵 𝝎 𝐾 𝐵 × 𝑽 𝐴 𝐴 = ­ − cos 𝛼 𝐴 sin 𝛽 𝐴 cos 𝛽 𝐴 − sin 𝛼 𝐴 sin 𝛽 𝐴 ® ­ 𝑞𝐾 ® ×­ 0 ® =
®­ ® ­ ®
𝑂𝐵 𝐺
𝐴
« − sin 𝛼 𝐴 0 cos 𝛼 𝐴 ¬ « 𝑟𝐾 ¬𝐵 « 0 ¬ 𝐴
𝐸
𝑝 cos 𝛼 𝐴 cos 𝛽 𝐴 + 𝑞 𝐾 sin 𝛽 𝐴 + 𝑟 𝐾 sin 𝛼 𝐴 cos 𝛽 𝐴 𝑉𝐺
© 𝐾 ª © 𝐴 ª
­ −𝑝 𝐾 cos 𝛼 𝐴 sin 𝛽 𝐴 + 𝑞 𝐾 cos 𝛽 𝐴 − 𝑟 𝐾 sin 𝛼 𝐴 sin 𝛽 𝐴 ® × ­ 0 ® = (74)
­ ® ­ ®
« −𝑝 𝐾 sin 𝛼 𝐴 + 𝑟 𝐾 cos 𝛼 𝐴 ¬𝐴 « 0 ¬𝐴
0
©
ª
­ 𝐺 𝐸 (𝑟 ®
­

𝑉𝐴 𝐾 cos 𝛼 𝐴 − 𝑝 𝐾 sin 𝛼 𝐴) ®
« 𝐴 𝐴 ¬𝐴
𝐺 𝐸 (
𝑉 𝑝 𝐾 cos 𝛼 𝐴 sin 𝛽 𝐴 − 𝑞 𝐾 cos 𝛽 𝐴 + 𝑟 𝐾 sin 𝛼 𝐴 sin 𝛽 )

The first term of the result in Eq. (71) can be expanded using the rotation of the aerodynamic frame with respect to the
body frame:
𝐸𝐴 𝐸
𝐸𝐴
 𝐸𝐵  𝐸𝐴     𝐸 © 𝑉𝐴 −𝛼¤ 𝐴 sin 𝛽 𝐴 𝑉¤ 𝐴𝐺
¤𝐺 𝑉 𝐴𝐺
ª © ª © ª ©
𝐸 ª
𝑽¤ 𝐴𝐺 = 𝑽¤ 𝐴𝐺 + 𝝎 𝐵 𝐴 × 𝑽 𝐴𝐺 = ­­ 0 ® +­
® ­ −𝛼¤ 𝐴 cos 𝛽 𝐴 ®® ×­­ 0 ®® = ­­ 𝑉 𝐴𝐺 𝛽¤ 𝐴 ® (75)
®
𝐴 𝐴 𝐴 𝐴

« 0 ¬𝐴 « ¬ 𝐴 « 0 ¬ 𝐴 « 𝑉 𝐴 𝛼¤ 𝐴 cos 𝛽 𝐴 ¬𝐴
𝛽¤ 𝐴 𝐺 𝐸

When inserting the above quantities in the derived equation for the A frame:
  𝐸 𝐵 Í (𝑭)      𝐸   𝐸𝑂
𝑽¤ 𝐴𝐺 = 𝐴
− 𝑴 𝐴𝐵 𝝎𝑂𝐵 𝐾 × 𝑽 𝐺
𝐴 − ¤𝐺
𝑽𝑊 (76)
𝐴 𝑚 𝐵 𝐴 𝐴 𝐴

16
 The result is:

𝐸𝐴
𝐸
𝑉¤ 𝐺 / 𝑉 𝐴𝐺 Í 0   𝐸𝑂
© 𝐴 ª (𝑭) 𝐴 ©­ ª
­ ® =
𝐸 ­ ® − 1
𝑴 𝐴𝐵 𝑴 𝐵𝑂 𝑽¤ 𝐺
­ 𝛽¤ 𝐴 ® − 𝑟 𝐾 cos 𝛼 𝐴 − 𝑝 𝐾 sin 𝛼 𝐴 ® 𝐸 𝑊
𝑚 𝑉 𝐴𝐺 𝑉 𝐴𝐺 𝑂
« 𝛼¤ 𝐴 cos 𝛽 𝐴 ¬𝐴 « 𝑝 𝐾 cos 𝛼 𝐴 sin 𝛽 𝐴 − 𝑞 𝐾 cos 𝛽 𝐴 + 𝑟 𝐾 sin 𝛼 𝐴 sin 𝛽 𝐴 ¬𝐴
(77)


C. Angle of attack dynamics

The relation given in Statement 5 for the angle of attack dynamics satisfies the following objectives:

Objectives
1) Obtain a relation for the aerodynamic angle of attack dynamics applicable for a steady (with respect to NED) wind
field.
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

2) Obtain this relation not using 𝛾 𝐴 and 𝜇 𝐴, rather using 𝛼 𝐴, 𝛽 𝐴, Φ and Θ, since estimates or measurements of these
variables are most convenient to obtain.
3) Obtain this relation in terms of specific force in the stability frame (note 𝑓 𝑧,𝑆 = 𝑓 𝑧, 𝐴).


𝐺 𝐸𝑂
Statement 5 (Angle of attack dynamics in S frame) Assuming that 𝑽¤ 𝑊 𝑂
= 0, the angle of attack dynamics can
be written as:
 
𝑔 𝑓 𝑧,𝑆
𝛼¤ 𝐴 =
𝐸 𝑔
+ sin 𝛼 𝐴 sin Θ + cos 𝛼 𝐴 cos Θ sin Φ + 𝑞 𝐾 − tan 𝛽 𝐴 ( 𝑝 𝐾 cos 𝛼 𝐴 + 𝑟 𝐾 sin 𝛼 𝐴) (78)
cos 𝛽 𝐴 𝑉 𝐴𝐺
Derivation. The first term in Eq. (67) can be split up into a gravity term and the specific forces ( 𝒇 ) 𝐴 in the aerodynamic
frame 𝐴 as follows
Í 𝑋𝐺
𝑔 ©­ ( 𝒇 ) 𝐴 1 ©­ 𝐺 ª ª

= 𝐺
𝐸 ­ 𝑔 + 𝑴 𝐴𝐵 𝑚𝑔 ­ 𝑌𝐺 ®® ®®
(𝑭) 𝐴 𝐺 (79)
𝐺 𝐸
𝑚 𝑉𝐴 𝑉𝐴
« « 𝐺 ¬𝐵 ¬
𝐺
𝑍
The gravity term, expressed in the body fixed frame 𝐵 can be obtained by using the euler angles:

𝑋𝐺 − sin Θ
1 ©­ 𝐺 ª ©
® = 𝑴 𝐵𝑂 ( 𝒈) 𝑂 = ­ cos Θ sin Φ
ª
®
𝑚𝑔 ­ 𝐺𝐺 ® ­ ®
𝑌𝐺 (80)
𝑔
« 𝑍𝐺 ¬𝐵 « cos Θ cos Φ ¬𝐵
with ( 𝒈) 𝑂 = (0, 0, 1)𝑇 and 𝑴 𝐵𝑂 given in Eq. (103). The 𝑧 component of Eq. (80) in the aerodynamic frame 𝐴 is given
by
𝐺
© ª
𝑋𝐺
1 ­ 𝐺 ® =
𝑴 𝐴𝐵 𝑚𝑔 ­ 𝑌𝐺 ®
« 𝑍𝐺 ¬𝐵
𝐺
(81)
cos 𝛼 𝐴 cos 𝛽 𝐴 sin 𝛽 𝐴 sin 𝛼 𝐴 cos 𝛽 𝐴 − sin Θ
© ª© ª © ª
...
­ ® ­ cos Θ sin Φ ® =­ ®
­ − cos 𝛼 𝐴 sin 𝛽 𝐴 cos 𝛽 𝐴 − sin 𝛼 𝐴 sin 𝛽 𝐴 ®­ ® ­ ... ®
« − sin 𝛼 𝐴 0 cos 𝛼 𝐴 ¬ « cos Θ cos Φ ¬ 𝐵 « sin 𝛼 𝐴 sin Θ + cos 𝛼 𝐴 cos Θ cos Φ ¬𝐴
Substituting Eq. (81) into Eq. (79), and then the resulting z-component back into the z component of Eq. (67) results in
Eq. (78), using that 𝑓 𝑧,𝑆 = 𝑓 𝑧, 𝐴 because the z axis of 𝐴 and 𝑆 frame is co-located. Hence the result. 

D. Trim contributions in angle of attack dynamics

Statement 6 (Trim contributions in angle of attack dynamics) The pitch-rate 𝑞 𝐾 can be split up in contributions
due to pilot command (maneuver) 𝑞 𝐾 ,𝑚𝑎𝑛 , trim compensation of gravity 𝑞 𝐾 ,𝑡𝑟 𝑖𝑚,𝑔 and kinematic coupling compensation
𝑞 𝐾 ,𝑘𝑐 .
𝑞 𝐾 = 𝑞 𝐾 ,𝑚𝑎𝑛 + 𝑞 𝐾 ,𝑡𝑟 𝑖𝑚,𝑔 + 𝑞 𝐾 ,𝑘𝑐 (82)

17
 where the kinematic coupling compensation pitch rate and trim compensation of gravity are given by
   
𝑞 𝐾 ,𝑘𝑐 = tan 𝛽 𝐴 𝒆 𝑖 𝑇 𝑴𝑆𝐵 𝝎𝑂𝐵
𝐾 = tan 𝛽 𝐴 ( 𝑝 𝐾 cos 𝛼 𝐴 + 𝑟 𝐾 sin 𝛼 𝐴) (83)
𝐵
and
𝑔  


𝑇
𝑞 𝐾 ,𝑡𝑟 𝑖𝑚,𝑔 = 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 − 𝒆 𝑘 𝑴 𝑆𝐵 𝑴 𝐵𝑂 ( 𝒈) 𝑂 /𝑔 (84)
𝐺 𝐸
cos 𝛽 𝐴 𝑉 𝐴
where 𝒆 𝑖 and 𝒆 𝑘 are the unit vectors along the 𝑥-axis and 𝑧-axis, and 𝑀𝑆𝐵 and 𝑀 𝐵𝑂 are given by Eq. (102) and Eq.
(103). Similar, the specific force can be split up into the reference value 𝑓 𝑧,𝑡𝑟 𝑖𝑚,𝑆 and an increment from that reference
value Δ 𝑓 𝑧,𝑡𝑟 𝑖𝑚,𝑆 :
𝑓 𝑧,𝑆 = 𝑓 𝑧,𝑡𝑟 𝑖𝑚,𝑆 + Δ 𝑓 𝑧,𝑡𝑟 𝑖𝑚,𝑆 (85)
Derivation. At trim condition and with no pilot command, i.e. 𝛼¤ 𝐴 = 0, Δ 𝑓 𝑧,𝑡𝑟 𝑖𝑚,𝑆 = 0 and 𝑞 𝐾 ,𝑚𝑎𝑛 = 0 the angle of
attack dynamics, Eq. (78), becomes:
 
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

𝑔 𝑓 𝑧,𝑡𝑟 𝑖𝑚,𝑆
𝛼¤ 𝐴 = 0 =
𝐸 𝑔
+ sin 𝛼 𝐴 sin Θ + cos 𝛼 𝐴 cos Θ sin Φ +𝑞 𝐾 ,𝑡𝑟 𝑖𝑚,𝑔 +𝑞 𝐾 ,𝑘𝑐 −tan 𝛽 𝐴 ( 𝑝 𝐾 cos 𝛼 𝐴 + 𝑟 𝐾 sin 𝛼 𝐴)
cos 𝛽 𝐴 𝑉 𝐴𝐺
(86)
where Eq. (82) and Eq. (85) were inserted. The different pitch rate contribution terms can now be identified as:
𝑞 𝐾 ,𝑘𝑐 = tan 𝛽 𝐴 ( 𝑝 𝐾 cos 𝛼 𝐴 + 𝑟 𝐾 sin 𝛼 𝐴) = tan 𝛽 𝐴 𝑝 𝑆
    (87)
= tan 𝛽 𝐴 𝒆 𝑖 𝑇 𝑴𝑆𝐵 𝝎𝑂𝐵 𝐾
𝐵

where 𝒆 𝑖 is the unit vector along the 𝑥-axis and

 
−𝑔 𝑓 𝑧,𝑡𝑟 𝑖𝑚,𝑆
𝑞 𝐾 ,𝑡𝑟 𝑖𝑚,𝑔 =
𝐸 𝑔
+ sin 𝛼 𝐴 sin Θ + cos 𝛼 𝐴 cos Θ sin Φ
cos 𝛽 𝐴 𝑉 𝐴𝐺
𝑔

=
𝐸 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 − sin 𝛼 𝐴 sin Θ − cos 𝛼 𝐴 cos Θ sin Φ (88)
cos 𝛽 𝐴 𝑉 𝐴𝐺
𝑔  
=
𝐸
𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 − 𝒆 𝑘 𝑇 𝑴𝑆𝐵 𝑴 𝐵𝑂 ( 𝒈) 𝑂 /𝑔
cos 𝛽 𝐴 𝑉 𝐴 𝐺

where 𝒆 𝑘 is the unit vector along the 𝑧-axis. 

E. Derivation of minimum stick gradient at low speed

Statement 7 (Minimum stick gradient at low speed) At the speed range 𝐿𝑆 = [𝑉𝑡𝑟 𝑖𝑚, 𝛼𝐶,max , 𝑉𝑡𝑟 𝑖𝑚, 𝛼𝐶,min ], the result-
ing minimum stick gradient in terms of angle of attack per stick deflection for the stick scaling given by Eq. (30) to Eq.
(32) is
𝜕𝛼𝑒𝑞,𝐶

= 𝛼max − 𝛼𝐶,min (89)
𝜕𝛿 𝑠 𝑉 ∈𝐿𝑆
where 𝑉𝑡𝑟 𝑖𝑚, 𝛼𝐶,min and 𝑉𝑡𝑟 𝑖𝑚, 𝛼max is the trim speed corresponding to a trim angle of attack of 𝛼𝐶,min and 𝛼max respectively
when pilot relief functions and soft protections are not active.
Derivation.
At low speed the stick gradient Eq. (31) will be
𝜕𝑛 𝑧,𝑒𝑞,𝐶
= Δ𝑛 𝑧,max,𝑆 𝑉 ∈𝐿𝑆
= 𝑛 𝑧,𝑚𝑎𝑥,𝑆 − 𝑛 𝑧, 𝛼𝐶,min ,𝑆 (90)
𝜕𝛿 𝑠,𝑙𝑜𝑛 𝑉 ∈𝐿𝑆

because of Eq. (4), in the case that pilot relief function and protections were not active. Note that at low speed
𝑛 𝑧,𝑚𝑎𝑥,𝑆 = 𝑛 𝑧,𝐶𝐿,𝑚𝑎𝑥 ,𝑆 because of Eq. (32) and because at low speed the structural load factor is larger than 𝑛 𝑧,𝐶𝐿,𝑚𝑎𝑥 ,𝑆 .
Equation (90) can be expressed in terms of an angle of attack increment by
𝜕𝑛 𝑧,𝑒𝑞,𝐶 𝜕𝑛 𝑧,𝑒𝑞,𝐶

= 𝑛 𝑧,𝑚𝑎𝑥,𝑆 − 𝑛 𝑧, 𝛼𝐶,min ,𝑆 = 𝛼max − 𝛼𝐶,min (91)
𝜕𝛿 𝑠,𝑙𝑜𝑛 𝑉 ∈𝐿𝑆 𝜕𝛼𝑒𝑞,𝐶

18
 where 𝛼max is the angle of attack corresponding to a load factor of 𝑛 𝑧,𝑚𝑎𝑥,𝑆 . Using the chain rule and relation (91), the
stick gradient in terms of angle of attack per stick deflection is expressed as
𝜕𝛼𝑒𝑞,𝐶 𝜕𝛼𝑒𝑞,𝐶 𝜕𝑛 𝑧,𝑒𝑞,𝐶

= = 𝛼max − 𝛼𝐶,min (92)
𝜕𝛿 𝑠 𝑉 ∈𝐿𝑆 𝜕𝑛 𝑧,𝑒𝑞,𝐶 𝜕𝛿 𝑠 𝑉 ∈𝐿𝑆

Hence the result. 

F. Steady state values of reference model CV’s

Statement 8 (Steady state values of reference model CV’s) Let the short period dynamics be given by Eq. (13) and
the control law by Eq. (36). With the following choice of the parameters 𝑘 1 , 𝑘 2

𝑘1 = 1
(93)
𝑘 2 = −1/𝑍 𝛼
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

the controller gain design will be independent of the blend 𝑏 𝛼 , 𝑏 𝑞 , 𝑏 𝑛𝑧 such that the controller gains 𝑘 𝛼 , 𝑘 𝑞 , 𝑘 𝐼 are
independent of the current CV.
Derivation. The control law is given by Eq. (36) and can be written as
 Δ𝛼𝑡𝑟 𝑖𝑚 
 
 
Δ𝑞¤𝑟 𝑒 𝑓 ,𝐶 = 𝐻𝐹Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝐶,𝑆 − [ 𝐾 𝛼 −𝐾 𝐼 ]  Δ𝑞 𝑡𝑟 𝑖𝑚  (94)
 
𝐾𝑞
 𝑒𝐼 
 
Only the integrator state 𝑒 𝐼 depends on 𝑘 1 , 𝑘 2 . The integrator ste is determined by the differential equation:
𝑒¤𝐼 = 𝐹Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝐶,𝑆 − Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝑆
= 𝐹Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝐶,𝑆 − (𝑏 𝛼 Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞, 𝛼,𝑆 + 𝑏 𝑛 Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝑛𝑧 ,𝑆 + 𝑏 𝑞 Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝑞,𝑆 )
  (95)
−𝑉 𝑍 𝛼 𝑉
= 𝐹Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝐶,𝑆 − (𝑏 𝛼 Δ𝐶𝑉 𝛼,lim + 𝑛 𝑧,𝐶𝐿,max − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 + 𝑏 𝑛𝑧 Δ𝐶𝑉𝑛𝑧,𝑡𝑟 𝑖𝑚,𝑆 + 𝑏 𝑞 Δ𝐶𝑉𝑞,𝑡𝑟 𝑖𝑚,𝑆 )
𝑔 𝑔
where Eq. (28) was inserted for Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝑆 . Inserting Δ𝐶𝑉 𝛼,lim , Δ𝐶𝑉𝑛𝑧,𝑡𝑟 𝑖𝑚,𝑆 given by Eq. (24) results in
 
−𝑉 𝑍 𝛼 𝑉
𝑒¤𝐼 = 𝐹Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝐶,𝑆 −𝑏 𝛼 (𝑘 1 Δ𝛼𝑙𝑖𝑚 + 𝑘 2 Δ𝛼¤ 𝑙𝑖𝑚 ) + 𝑛 𝑧,𝐶𝐿,max − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 −𝑏 𝑛𝑧 (𝑘 1 Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 +𝑘 2 𝑛¤ 𝑧,𝑡𝑟 𝑖𝑚,𝑆 )−𝑏 𝑞 Δ𝑞 𝑡𝑟 𝑖𝑚
𝑔 𝑔
(96)
We use the relations in Eq. (13) and Eq. (15), which are repeated here for convenience:
−𝑉 𝑍 𝛼
Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 = Δ𝛼𝑡𝑟 𝑖𝑚 (97)
𝑔
𝑔

Δ𝛼𝑙𝑖𝑚 = Δ𝛼𝑡𝑟 𝑖𝑚 + 𝑛 𝑧,𝐶𝐿,max − 𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑆 (98)
𝑉 𝑍𝛼
We express the derivatives Δ𝛼¤ 𝑙𝑖𝑚 and Δ𝑛¤ 𝑧,𝑡𝑟 𝑖𝑚 by
−𝑉 𝑍 𝛼
Δ𝑛¤ 𝑧,𝑡𝑟 𝑖𝑚,𝑆 = Δ𝛼¤ 𝑡𝑟 𝑖𝑚
𝑔 (99)
Δ𝛼¤ 𝑙𝑖𝑚 = Δ𝛼¤ 𝑡𝑟 𝑖𝑚 = 𝑍 𝛼 Δ𝛼𝑡𝑟 𝑖𝑚 + Δ𝑞 𝑡𝑟 𝑖𝑚
where Δ𝛼¤ 𝑙𝑖𝑚 = Δ𝛼¤ 𝑡𝑟 𝑖𝑚 was shown in Eq. (23). Choosing 𝑘 1 = 1 Eq. (96) can then be reduced to:
𝑉 𝑍𝛼 𝑉
𝑒¤𝐼 = 𝐹Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝐶,𝑆 + Δ𝛼𝑡𝑟 𝑖𝑚 (𝑏 𝛼 (1 + 𝑘 2 𝑍 𝛼 ) + 𝑏 𝑛𝑧 (1 + 𝑘 2 𝑍 𝛼 )) − Δ𝑞 𝑡𝑟 𝑖𝑚 (−𝑍 𝛼 𝑘 2 (𝑏 𝛼 + 𝑏 𝑛𝑧 ) + 𝑏 𝑞 ) (100)
𝑔 𝑔
It can be directly seen that by choosing 𝑘 2 = − 𝑍1𝛼 , 𝑒¤𝐼 becomes independent of the blend because according to Eq. (29)
𝑏 𝛼 + 𝑏 𝑛𝑧 + 𝑏 𝑞 = 1.
𝑉 𝑉
𝑒¤𝐼 = 𝐹Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝐶,𝑆 − Δ𝑞 𝑡𝑟 𝑖𝑚 (𝑏 𝛼 + 𝑏 𝑛𝑧 + 𝑏 𝑞 ) = 𝐹Δ𝑛 𝑧,𝑡𝑟 𝑖𝑚,𝑒𝑞,𝐶,𝑆 − Δ𝑞 𝑡𝑟 𝑖𝑚 (101)
𝑔 𝑔

19
 G. Transformations
Statement 9 (Rotation from body frame to stability frame) The coordinate rotation matrix from body frame 𝐵 to
stability frame 𝑆 is given by:
cos 𝛼 𝐴 0 sin 𝛼 𝐴
© ª
M𝑆𝐵 = ­ ­ ®
0 1 0 ® (102)
« − sin 𝛼𝐴 0 cos 𝛼 𝐴 ¬
Statement 10 (Rotation from NED to body frame ) The coordinate rotation matrix from NED frame 𝑂 to body frame
𝐵 is given by:

cos Ψ cos Θ sin Ψ cos Θ − sin Θ

© ª
M 𝐵𝑂 ­
= ­ cos Ψ sin Θ sin Φ − sin Ψ cos Φ sin Ψ sin Θ sin Φ + cos Ψ cos Φ cos Θ sin Φ ®
® (103)
« cos Ψ sin Θ cos Φ + sin Ψ sin Φ sin Ψ sin Θ cos Φ − cos Ψ sin Φ cos Θ cos Φ ¬
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

References
[1] Holzapfel, F., and Sachs, G., “Dynamic inversion based control concept with application to an unmanned aerial vehicle,” AIAA
Guidance, Navigation, and Control Conference and Exhibit, 2004, p. 4907.

[2] Harris, J. J., “F-35 flight control law design, development and verification,” 2018 Aviation Technology, Integration, and
Operations Conference, 2018, p. 3516.

[3] Delannoy, S., and Oudin, S., “Longitudinal control law for modern long range civil aircrafts,” Proceedings of the EuroGNC,
2013.

[4] Zhang, J., Wang, J., Zhang, F., and Holzapfel, F., “Modeling and Incremental Nonlinear Dynamic Inversion Control for a Highly
Redundant Flight System,” AIAA Scitech 2019 Forum, 2019, p. 1922.

[5] Sieberling, S., Chu, Q. P., and Mulder, J. A., “Robust Flight Control Using Incremental Nonlinear Dynamic Inversion and
Angular Acceleration Prediction,” Journal of Guidance, Control, and Dynamics, Vol. 33, No. 6, 2010, pp. 1732–1742.
[Link] URL [Link]

[6] Akkinapalli, V. S., and Holzapfel, F., “Incremental dynamic inversion based velocity tracking controller for a multicopter
system,” 2018 AIAA Guidance, Navigation, and Control Conference, 2018, p. 1345.

[7] Raab, S. A., Zhang, J., Bhardwaj, P., and Holzapfel, F., “Consideration of Control Effector Dynamics and Saturations in an
Extended INDI Approach,” AIAA Aviation 2019 Forum, 2019, p. 3267.

[8] Smeur, E., de Croon, G., and Chu, Q., “Cascaded incremental nonlinear dynamic inversion for MAV disturbance rejection,”
Control Engineering Practice, Vol. 73, 2018, pp. 79–90. [Link] URL
[Link]

[9] Raab, S. A., Zhang, J., Bhardwaj, P., and Holzapfel, F., Proposal of a Unified Control Strategy for Vertical Take-off
and Landing Transition Aircraft Configurations, chapter and pages. [Link] URL https:
//[Link]/doi/abs/10.2514/6.2018-3478.

[10] Bhardwaj, P., Raab, S. A., Zhang, J., and Holzapfel, F., Integrated Reference Model for a Tilt-rotor Vertical Take-off and Landing
Transition UAV, chapter and pages. [Link] URL [Link]

[11] Zhang, F., and Holzapfel, F., “Flight control using physical dynamic inversion,” AIAA Guidance, Navigation, and Control
Conference, 2015, p. 1758.

[12] Bhardwaj, P., Raab, S. A., and Holzapfel, F., “Higher Order Reference Model for Continuous Dynamic Inversion Control,”
AIAA Scitech 2021 Forum, 2021, p. 1130.

[13] Steffensen, R., Gabrys, A., and Holzapfel, F., “Flight envelope protections using phase plane limits and backstepping control,”
5th CEAS Conf. Guid. Navig. Control, 2019.

20
 [14] Gabrys, A. C., Steffensen, R., de Angelis Cordeiro, R., Azinheira, J. R., Moutinho, A., and Holzapfel, F., “Integration of Phase
Plane Flight Envelope Protections in Cascaded Incremental Flight Control *∗This research is co-funded by the European
Union in the scope of INCEPTION project, which has received funding from the EU’s Horizon2020 Research and Innovation
Programme under grant agreement No. 723515.” IFAC-PapersOnLine, Vol. 52, No. 12, 2019, pp. 429–435. [Link]
//[Link]/10.1016/[Link].2019.11.281, URL [Link] 21st IFAC
Symposium on Automatic Control in Aerospace ACA 2019.

[15] Gabrys, A., Holzapfel, F., Steffensen, R., and Merkl, C., “Flight Test Based Gain Tuning using non-parametric Frequency
Domain Methods,” AIAA Scitech 2021 Forum, 2021, p. 1424.

[16] Schatz, S. P., Gabrys, A. C., Gierszewski, D. M., and Holzapfel, F., “Inner Loop Command Interface in a Modular Flight Control
Architecture for Trajectory Flights of General Aviation Aircraft,” 2018 5th International Conference on Control, Decision and
Information Technologies (CoDIT), IEEE, 2018, pp. 86–91.

[17] Gossmann, F., Svaricek, F., and Gabrys, A., “Control of longitudinal aircraft motion with loadcase robustness using LPV-control
with partly-measurable parameters,” 2018 AIAA Guidance, Navigation, and Control Conference, 2018, p. 0854.
Downloaded by Bibliothek der TU Muenchen on April 20, 2024 | [Link] | DOI: 10.2514/6.2022-1230

[18] Goßmann, F., Gabrys, A., and Svaricek, F., “Longitudinal Short-Period Aircraft Motion Control Under Loadcase Variation,”
2020 American Control Conference (ACC), IEEE, 2020, pp. 4188–4194.

[19] Bodson, M., “Evaluation of optimization methods for control allocation,” Journal of Guidance, Control, and Dynamics, Vol. 25,
No. 4, 2002, pp. 703–711.

[20] Zhang, J., Bhardwaj, P., Raab, S. A., Saboo, S., and Holzapfel, F., “Control allocation framework for a tilt-rotor vertical take-off
and landing transition aircraft configuration,” 2018 Applied Aerodynamics Conference, 2018, p. 3480.

[21] Braun, D., Schweiger, F., Holzapfel, F., Diepolder, J., and Ben-Asher, J. Z., “Continuous Integration of Optimal Control Based
Flight Control Law Clearance,” AIAA Scitech 2022 Forum (submitted and accepted for publication), San Diego, CA, USA,
2022.

[22] Schwaiger, F., Mishra, C., and Holzapfel, F., “Acceleration Framework to Quantify the Influence of Uncertain Parameters on
Simulation Models using Matlab and Simulink,” AIAA Scitech 2021 Forum, 2021. [Link]

[23] Schwaiger, F., Shi, D., Mishra, C., Höhndorf, L., and Holzapfel, F., “A Modular Subset Simulation Toolbox for Matlab,” AIAA
Scitech 2022 Forum, 2022. Submitted for publication.

[24] Braun, D., Schweiger, F., Holzapfel, F., Diepolder, J., and Ben-Asher, J. Z., “Continuous Integration of Optimal Control Based
Flight Control Law Clearance,” AIAA Scitech 2022 Forum (submitted and accepted for publication), San Diego, CA, USA,
2022.

21