Jerk derivative feedforward control for motion systems

Matthijs Boerlage† Rob Tousain‡ Maarten Steinbuch†

−7
x 10
Abstract— This work discusses reference trajectory relevant 4
Measurement
model based feedforward design. For motion systems which Simulation
3 Acc (scaled)
contain at least one rigid body mode and which are subject
to reference trajectories with mostly low frequency energy, 2
the proposed feedforward controller improves tracking per-

Servo error [m]

1
formance significantly. The feedforward controller may be of
much lower order than the plant. The proposed feedforward 0
controller is introduced using a model of an industrial XY-
table as an application example. −1

−2
I. I NTRODUCTION
A
−3
In many today’s motion systems, performance require-
ments include short motion times and small settling times. −4
0 0.05 0.1 0.15 0.2
Typical examples are pick and place machines, hard disk Time [sec]

drives, XY-tables and many robots. To meet these re-

quirements typically, a combination of a feedback and Fig. 2. Servo errors during point-to-point motion of a XY-positioning
feedforward controller is used in a so called two degree of table using rigid body feedforward and PD controller. Measured servo error
(solid thick), simulated servo error (solid thin) and scaled acceleration of
freedom (2DOF) control architecture, see Figure 1, where reference trajectory (dashed). The area A is studied in later simulations.

F f
constant motion phases (velocity, acceleration, jerk, deriva-
r e u y
+ C + P tive of jerk, etc.), [1],[5]. A property of these profiles is that
- they contain mostly low frequency energy, as will be illus-
trated in section II. The profiles are generally designed such
that the resonance dynamics of the plant are not excited.
Fig. 1. 2DOF control architecture. This paper focuses on the case where this is done properly.
Would this not be the case, input shaping techniques may
P , C and F represent the plant, feedback controller and be applied, as demonstrated in e.g. [2],[4],[5],[7],[8]. In this
feedforward controller, respectively. Signals are denoted paper it is further assumed that the fourth derivative of
by lower case: the reference trajectory r, servo error e, the position trajectory exists. As will be shown later, this
plant input u, plant output y and feedforward function requirement is often satisfied in practice due to discrete time
f . The feedback controller guards stability and improves implementation aspects.
disturbance rejection [9], while the feedforward controller
is designed to improve tracking performance. This work will B. Plant dynamics
consider the design of a feedforward controller in order to In industrial systems, effort is put in designing mechanics
reduce tracking errors during motion (a typical industrial with very high resonance frequencies, with the goal to
example is displayed in Figure 2). extend the rigid body-like behavior over a frequency range
as large as possible. A typical motion system can therefore
A. Reference profiles be seen as a plant which contains one or more dominant
In industrial practice, commonly used reference pro- rigid body modes and several resonance modes at higher
files describe varying scanning and point-to-point motions. frequencies (due to limited mechanical stiffness), see also
These profiles are often designed as piecewise finite order Figure 3. A transfer function model of a SISO motion
polynomials. Finite order polynomials typically contain system can therefore be given as follows:
† Eindhoven University of Technology Faculty of Mechanical En- X N
1 ki
gineering, Control Systems Technology Group P.O. Box 513, 5600 P (s) = + , (1)
MB Eindhoven, The Netherlands Email: [Link]@[Link], mt s 2
i=1
mt (s + 2ζi ωi s + ωi2 )
2

[Link]@[Link]
‡ Philips CFT Mechatronics Research P.O. Box 218, SAQ-2115, 5600 where mt is the total mass of the system and N the
MD Eindhoven, The Netherlands Email: [Link]@[Link] number of resonance modes with ωi , ζi and ki the mode
resonance frequency, damping and mode gain ki ∈ {−1, 1} E. Outline
respectively. Note that by changing the sign of ki , the modes The next section starts with an analysis of the residual
can show up as resonances or anti-resonances in a bode servo errors when using acceleration feedforward control. In
diagram. When i = 1, a two-mass-spring system arises, as the third section the jerk derivative feedforward controller
presented in [13]. The model does not include damping or is described. In section IV, implementation aspects of the
friction to the world which is assumed to be negligible or proposed method will be discussed. Finally a short discus-
compensated for otherwise. sion wil follow and the work will close with conclusions.
Throughout this work, a model of an industrial XY-table is
C. Feedforward design used as a motivating example.
A common model based feedforward design approach is II. OPEN LOOP SERVO ERROR DURING MOTION
to make the feedforward controller F equal to the inverse of
the plant, thereby directly minimizing the transfer between To find the origin of residual servo errors during motion,
servo error e(s) and reference trajectory r(s), derived from the transfer between the reference trajectory (r(s)) and the
Figure 1; open loop servo error (eo (s)) is studied;
e(s) 1 − P (s)F (s) eo (s) = r(s) − P (s)F (s)r(s) (4)
= (2)
r(s) 1 + P (s)C(s)
If the plant is represented by Equation 1 and an acceleration
Model inversion is not always feasible due to non-minimum feedforward controller F (s) = mt s2 is used, the following
phase behavior of the plant. A popular method to over- relation between the reference acceleration and the open
come these difficulties is the application of a zero phase loop servo errors results;
error tracking controller (ZPETC) [10] or extensions to N
this scheme [3],[11],[12]. As many motion system contain eo (s) X ki
= 2 + 2ζ ω s + ω 2
. (5)
dominant rigid body behavior, a straightforward approach is r̈(s) i=1
s i i i
rigid body inversion by means of acceleration feedforward.
When it is assumed that the natural frequencies of the plant
In this case the feedforward controller equals a double
lay much higher then the spectral content of the reference
differentiator times the modelled mass m̂t of the plant;
profile (see Figure 3), Equation 5 may be approximated by;
F (s) = m̂t s2 (3) PN Q 2
eo (s) i=1 ki j∈{1,...,N |j6=i} ωj
|s→0 = QN . (6)
leading to a feedforward function which is equal to the r̈(s) i=1 ωi
2

scaled acceleration of the reference trajectory (hence justi- The open loop servo error during motion therefore has
fying the method’s name). Its simplicity and effectiveness the shape of the acceleration of the reference, scaled
made acceleration feedforward widely applicable in indus- by the right side of Equation 6. Hence, for a plant
try. with a high natural frequency (ωi ), low frequency
open loop servo errors remain during acceleration which
D. Problem statement can not be compensated for using acceleration feedforward.
Limitations of acceleration feedforward are experienced
in industrial applications. Servo errors during jerk phases The closed loop servo error e(s) during motion can
in a motion remain, which have typically low frequency be found by filtering the open loop servo error eo (s)
behavior and thereby imply an increase of settling time1 , with the sensitivity function S = (1 + P (s)C(s))−1 , see
deteriorating performance of motion systems, see Figure 2. Equation 2. The sensitivity function has at least a +2 slope
These residual servo errors are highly reproducible [6] and in lower frequencies due to the rigid body mode of the
have a strong dependence on the acceleration of the motion. plant. The closed loop servo error is therefore proportional
to the double derivative of the open loop servo error. This
In this work, a feedforward controller which extends results in peaks during non-zero jerk phases of a motion,
the commonly accepted acceleration feedforward with see Figure 2. In order to illustrate the insight presented in
an additional term, is proposed to reduce these residual this section, an example from industry is shown below.
servo errors. It will be shown that at least a fourth order
A. Motivating Example
feedforward controller is needed to compensate for low
frequency residual servo errors. The proposed feedforward We consider a motion system with frequency transfer
strategy is therefore named jerk derivative feedforward. function given in Figure 3. This model represents the
dynamics in one axis of an industrial XY-table. The
1 The ‘settling time’ is the time interval after acceleration or deceleration plant has a dominant rigid body behavior, its first natural
after which the servo error is required to be within certain bounds. frequency lays at approximately 600Hz. The system is
 −100 50

40
−150
30

20
−200

Magnitude [dB]
Magnitude [dB]

10

−250 0

−10
−300 −20

−30
−350
−40
Spec. r (scaled) F*
Plant −50 δ s2
−400 1 2 3
10 10 10 1 2 3
10 10 10
Frequency [Hz] Frequency [Hz]

Fig. 3. Spectrum of reference trajectory (scaled) and frequency response Fig. 4. Ideal additive feedforward controller F ∗ and approximation δs2 .
of the plant.

TABLE I
R EFERENCE TRAJECTORY PARAMETERS . new feedforward controller is assumed to be the original
acceleration feedforward controller with an additional term
F ∗;
Parameter xmax vmax amax jmax Ts F = m̂t s2 + F ∗ s2 . (7)
Value 8cm 0.5m/s 10m/s2 2000m/s3 0.25ms
A hint on how the term F ∗ can be designed can be found by
putting the open loop servo error during motion (4) equal
to zero;
subject to reference trajectories with parameters listed in 1
F ∗ = P −1 2 − mt . (8)
Table I. The (scaled) acceleration trajectory in time is s
depicted in Figure 2 by the dotted line. The spectrum For this example plant, the frequency response of this
of the reference position trajectory is shown in Figure ’ideal’ feedforward controller, is shown in Figure 4. Since
3. If acceleration feedforward control is used, residual we are interested in the low-frequency contribution of the
dynamics are described by Equation 5. The maximum additional feedforward only, F ∗ can be approximated by a
open loop servo error during motion may be estimated gain δ times a double differentiator, see Figure 4. So that
using Equation 6. As the maximum acceleration equals
10m/s2 , the maximum open loop error during motion is F ∗ |s→0 = δs2 , (9)
approximately 500nm. The open loop servo error has the then δ can be derived from Equation 1 and Equation 8;
shape of the acceleration of the reference trajectory, see PN Q
Figure 5. −mt i=1 ki j∈{1,...,N |j6=i} ωj2
δ= QN . (10)
2
i=1 ωi
Filtering the open loop servo error with the sensitivity
obtained by using a PD feedback controller (0dB crossing Hence, the proposed feedforward controller becomes;
at 80Hz, 6dB S peak) gives the closed loop servo error F = mt s2 + δs4 . (11)
shown in Figure 2 with the enlarged area A in Figure 5.
The second part of this controller is proportional to the
From this example, it appears that the assumptions fourth derivative or, in other words, the derivative of the
made earlier are justified in this application. Non-rigid jerk of the setpoint. Note that δ is a real scalar, which makes
body behavior causes servo errors during motion which online tuning feasible. The example below will illustrate an
cannot be compensated for using acceleration feedforward application of the jerk derivative feedforward controller.
control. The next section will propose an extension of
A. Example using jerk derivative feedforward
acceleration feedforward control in order to increase
tracking performance. The jerk derivative feedforward controller is applied to
the example discussed earlier. The constant δ is derived
III. J ERK DERIVATIVE FEEDFORWARD from Equation 10 and used in the new feedforward
In order to reduce the open loop servo error during controller. The reference profile with parameters given in
motion, a new feedforward controller is proposed. The Table I is applied. The resulting servo error during motion
 −7
x 10 eo/ddr
2.5
eo δ/mt
2 ecl,AccFF −120 Ftrunc
e
cl,djerkFF
1.5
−130

Magnitude [dB]
1
Servo error [m]

0.5 −140

−0.5 −150

−1
−160
−1.5

−2 −170
1 2 3
10 10 10
−2.5 Frequency [Hz]
0.04 0.05 0.06 0.07 0.08 0.09 0.1
Time [sec]

Fig. 7. Residual dynamics eo /r̈, δ and low frequency effect of F trunc .

Fig. 5. Servo error during point to point motion, reaching constant velocity
(Blok A, Figure 2). Dashed: (scaled) open loop servo error, Solid thin:
closed loop servo error using acceleration feedforward, Solid thick: closed
loop servo error using jerk derivative feedforward. residual dynamics of plants with orders much higher than
the order of the feedforward controller (= 4).
−80
eo,RBFF IV. I MPLEMENTATION ASPECTS
eo,djerkFF
−100 For practical implementation of the jerk derivative feed-
forward, a discrete time equivalent of Equation 11 needs
−120 to be derived. Neglecting measurement delay, a possible
implementation of the feedforward controller is presented
Magnitude [dB]

−140 in Figure 8, where q is the shift operator. The gray box

−160 I II

−180 acc Ts 2 q+T s2 1 r e

C u y
P
2(q-1) 2 q -
−200 f

1
10
2
10
3 Kacc
q
Frequency [Hz]
q-1 q-1
Kdjerk
T sq T sq
Fig. 6. Spectrum of open loop servo error using acceleration feedforward
(solid thin) and jerk derivative feedforward (solid thick).
Fig. 8. Possible implementation of jerk derivative feedforward.

is shown in Figure 5. Clearly, the servo error is reduced includes the blocks that are added to a conventional ac-
significantly. Comparison of the spectra of the open loop celeration feedforward solution. The conventional acceler-
servo errors of acceleration and jerk derivative control ation feedforward consists of an acceleration feedforward
shows that jerk derivative control leads to a significant gain Kacc (= mt ) and the discrete time equivalent of a
improvement of low frequency tracking, Figure 6. double integrator (block I) which computes the position
reference from the acceleration signal. The jerk-derivative
Notice that δ/mt is a low frequency approximation feedforward path basically consists of a series connection of
of Equation 5. For comparison, the frequency response two differentiating filters and a jerk derivative feedforward
of both terms is shown in Figure 7. Clearly, the jerk gain Kdjerk (= δ). The differentiating filters are, by means
derivative feedforward controller does not follow the of delay, implemented in a causal fashion. To compensate
residual dynamics at higher frequencies. However, the for the combined delay of the jerk derivative path, delay
low frequency contributions of all plant modes are operators have to be added to the position and acceleration
compensated for exactly. Jerk derivative feedforward feedforward path. This is done to match the phase of
therefore has the freedom to compensate for low frequency the setpoint position with the phase of the corresponding
terms in the feedforward controller. Finally, note that in the
discrete time implementation of the proposed controller the
−100
requirement for the reference trajectory to be at least of
fourth order can be relaxed.
−120

V. D ISCUSSION
−140
The proposed jerk derivative feedforward controller

Magnitude [dB]
proves to be capable of compensating low frequency −160
tracking errors which arise due to flexibility in a motion
system. The residual dynamics after using jerk derivative −180
feedforward can be calculated following the same strategy
as was used in section II. For the example used in this −200
work, these residual dynamics have a magnitude −280dB Plant
−1
djerkFF
in low frequency regions. And hence, higher order (higher −220 Ftrunc
than 4) feedforward controllers will not lead to a substantial 2
10
3
10
improvement in the low frequency region. In the high Frequency [Hz]
frequency region, higher order feedforward design may
lead to a better approximation of the residual dynamics. Fig. 9. Plant and inverse of feedforward controllers. Plant (solid thin),
This becomes particularly interesting when resonance Mode inversion based feedforward controller (dashed), derivative of jerk
feedforward controller (solid thick).
frequencies of a motion system are strongly excited by the
reference profile. However, as stated in the introduction
to this paper, this is in practice often prevented a priori VI. C ONCLUSIONS
through an appropriate design of the reference trajectory. For a plant with a rigid body mode and high frequency
dynamics, which is subject to reference profiles with mostly
The additive form of the jerk derivative feedforward low frequency energy, application of only rigid body feed-
controller has many practical advantages. After tuning forward may lead to performance limiting residual servo
the acceleration feedforward, δ may be tuned to cancel errors during motion. These residual servo errors result
residual servo errors during jerk phase. As δ is a constant from low frequency modal contributions which cannot be
parameter, the possibility arises to tune the jerk derivative compensated for using rigid body feedforward control. A
feedforward controller online, monitoring the servo error new, setpoint relevant, model based feedforward controller
in the time domain. This is similar to the way in which is proposed which uses the derivative of jerk as an additional
acceleration feedforward controllers are tuned in current parameter. Jerk derivative feedforward control compensates
industrial practice. for low frequency residual dynamics of plant with (possibly)
much higher order than the feedforward controller. Jerk
As was mentioned in the introduction, model based derivative feedforward can be tuned online, similar to com-
feedforward design is often based on obtaining an mon industrial acceleration feedforward tuning procedures.
approximate inverse of the plant. Often, from a practical
point of view, the order of the feedforward controller R EFERENCES
is constrained. A possible compromising design choice [1] B.H. Chang, Y. Hori, ”Trajectory Design considering Derivative of
could be to use the inverse of a truncated plant model Jerk for Head-positioning of Disk Drive System with Mechanical
VIbration”. Proc. of the American Control Conference, 2003, pp.
as a feedforward controller, F trunc . If the order is 4335-4340.
constrained to be 4, this feedforward controller will [2] B.G. Dijkstra, N.J. Rambaratsingh, C. Scherer, O.H. Bosgra, M.
have the same order as the jerk derivative feedforward Steinbuch, and S. Kerssemakers, ”Input design for optimal discrete-
time point-to-point motion of an industrial XY positioning table”,
controller. However, F trunc only compensates for the low in Proc. 39th IEEE Conference on Decision and Control, 2000, pp.
frequency contribution of this particular inverted mode, see 901-906.
Figure 7, while the jerk derivative feedforward controller [3] H. S. Park; P.H. Chang, D.Y. Lee, ”Continuous zero phase error
tracking controller with gain error compensation ”. Proc. of the
compensates for all low frequency modal contributions. American Control Conference, 1999, Vol. 5, pp. 3554-3558.
Jerk derivative feedforward will therefore have superior [4] P.H. Meckl, ”Discussion on : Comparison of Filtering Methods for
performance for references with mostly low frequency Reducing Residual Vibration”. European Journal of Control, 1999,
Vol. 5, pp. 219-221.
energy. In high frequency regions, F trunc approximates [5] P.H. Meckl, ”Optimized S-Curve Motion Profiles for Minimum
the resonance of the inverted mode, while jerk derivative Residual Vibration”. Proceedings of the American Control Confer-
feedforward shows a poor approximation of the resonance ence, June 1998.
[6] I. Rotariu, R.M.L. Ellenbroek, G.E. van Baars, M. Steinbuch, ”It-
behavior, Figure 9. Hence, it is expected that in this region, erative Learning Control for variable setpoints, applied to a motion
F trunc will outperform jerk derivative feedforward control. system”, European Control Conference 2003.
 [7] D. Roover, F. Sperling, 1997,”Point-to-point Control of a High
Accuracy Positioning Mechanism”, Proc. of the American Control
Conference,1997, pp.1350-1354.
[8] N. Singer, W. Singhose and [Link], ”Comparison of Filtering
Methods for Reducing Residual Vibration”. European Journal of
Control, 1999, Vol. 5, pp. 208-218.
[9] M. Steinbuch, M.L. Norg, ”Advanced Motion Control: an industrial
perspective”, European Journal of Control, 1998, pp.278-293.
[10] M. Tomizuka,”Zero Phase Error Tracking Algorithm for Digital
Control”. Trans. ASME J. Dynamic Systems, Measurement, and
Control, 1987, Vol.109, pp. 65-68.
[11] D.E. Torfs, R. Vuerinckx, J. Swevers, J. Schoukens, ”Comparison
of Two Feedforward Design Methods Aiming at Accurate Trajectory
Tracking of the End Point of a Flexible Robot Arm”, IEEE Trans-
actions on Control Systems Technology, 1998, Vol. 6(1). pp. 1-14.
[12] D.E. Torfs, J. Swevers, J. De Schutter,”Quasi-perfect tracking control
of non-minimal phase systems”. Proc. of the 30th Conference on
Decision and Control, 1991. pp. 241-244.
[13] B. Wie, D. Bernstein, ”A Benchmark Problem for Robust Control
Design,” Proc. of the American Control Conference, 1991.