Version of Record: [Link]

com/science/article/pii/S0925231215017439
Manuscript_bea7dae7e157a3a2755a445dc6ef0f45

Stochastic Gradient Based Extreme Learning Machines For Stable Online

Learning of Advanced Combustion Engines

Vijay Manikandan Janakiramana,∗, XuanLong Nguyenb , Dennis Assanisc

a Department of Mechanical Engineering, University of Michigan, Ann Arbor, MI, USA.
b Department of Statistics, University of Michigan, Ann Arbor, MI, USA.
c Stony Brook University, NY, USA.

Abstract

We propose and develop SG-ELM, a stable online learning algorithm based on stochastic gradients and Extreme
Learning Machines (ELM). We propose SG-ELM particularly for systems that are required to be stable during learn-
ing; i.e., the estimated model parameters remain bounded during learning. We use a Lyapunov approach to prove
both asymptotic stability of estimation error and stability in the estimated parameters suitable for identification of
nonlinear dynamic systems. Using the Lyapunov approach, we determine an upper bound for the learning rate of
SG-ELM. The SG-ELM algorithm not only guarantees a stable learning but also reduces the computational demand
compared to the recursive least squares based OS-ELM algorithm (Liang et al., 2006). In order to demonstrate the
working of SG-ELM on a real-world problem, an advanced combustion engine identification is considered. The algo-
rithm is applied to two case studies: An online regression learning for system identification of a Homogeneous Charge
Compression Ignition (HCCI) Engine and an online classification learning (with class imbalance) for identifying the
dynamic operating envelope. The case studies demonstrate that the accuracy of the proposed SG-ELM is comparable
to that of the OS-ELM approach but adds stability and a reduction in computational effort.
Keywords: Online Learning, Extreme Learning Machine, System Identification, Lyapunov Stability, Engine
Control, Operating Envelope.

1. Introduction

Homogeneous Charge Compression Ignition (HCCI) Engines are of significant interest to the automotive industry
owing to their ability to reduce emissions and fuel consumption significantly compared to existing methods such
as spark ignition (SI) and compression ignition (CI) engines [1, 2, 3]. Although HCCI engines tend to do well in
laboratory controlled tests, practical implementation is quite challenging because HCCI engines do not have a direct
trigger for ignition (such as spark in SI or fuel injection in CI). Further, HCCI requires some special engine designs
such as exhaust gas recirculation (EGR) [4], variable valve timings (VVT) [5], intake charge heating [6] among

∗ Corresponding author. Presently at NASA Ames Research Center, MS 269-1, Moffett Field, CA 94035-1000, USA. Phone: +1-734-358-6633.

Email: vijai@[Link].

Preprint submitted to Elsevier November 1, 2015

© 2015 published by Elsevier. This manuscript is made available under the Elsevier user license
[Link]
others. Such advanced designs also increase the complexity of the engine operation making it unstable and extremely
sensitive to operational disturbances [7, 8]. A model based control is typically opted to address the challenges involved
in controlling HCCI [9, 5, 10]. For model development, both physics based approaches [9, 5, 10] and data based
approaches [11, 12, 13] were shown to be effective. A key requirement for a model based control is the ability of the
models to accurately predict the engine state variables for several operating cycles ahead of time, so that a control
action with a known impact can be applied to the engine. Further, in order to be vigilant against the engine drifting
towards instabilities such as misfire, ringing, knock, etc [14, 15], the operating limits of the engine particularly in
transients, is required. In order to develop controllers and operate the engine in a stable manner, both models of the
engine state variables as well as the operating envelope are necessary.
Data based modeling approaches for the HCCI engine state variables and dynamic operating envelope were
demonstrated using neural networks [11], support vector machines [12], extreme learning machines [13] by the
authors. However, the previous research considered an offline approach where the data collected from engine ex-
periments were taken offline and models were developed using computer workstations that had high processing and
memory. However, a key requirement in advancing HCCI modeling is to perform online learning for the following
reasons. The models developed offline are valid only in the controlled experimental conditions. For instance, the
experiments are performed at a controlled ambient temperature, pressure and humidity conditions. As a result, the
models developed are valid for the specified conditions and when the models are implemented on a vehicle, the expec-
tation is that the model works on a wide range of climatic conditions that the vehicle is exposed to, possibly conditions
that were not experimented. Hence, an online adaptation to learn the behavior of the system at new/unfamiliar situ-
ations is required. Also, since the offline models are developed directly from experimental data, they may perform
poorly in certain operating regions where the density of experimental data is low. As more data becomes available in
such regions, an online mechanism can be used to adapt to such data. In addition, the engine produces high velocity
streaming data; operating at about 2500 revolutions per minute, an in-cylinder pressure sensor can produce about 1.8
million data observations per day. It becomes infeasible to store this volume of data for offline model development.
Thus, an online learning framework that processes every data observation, updates the model and throws away the
data is required for advanced engines like HCCI.
Online learning, as the name suggests, refers to a model that learns online; i.e., learning happens when the system is
in operation and as data is streaming in. Typically, the learning is sequential; i.e., the data from the system is processed
one-by-one or batch-by-batch and the model parameters are updated. A data processor on-board a combustion engine
usually is low on computation power and memory. Thus, simple linear models used to be the natural choice for
combustion engines. However, for a system like the HCCI engine, linear models may be insufficient to capture the
complex dynamics, particularly for predicting several steps ahead in time [11]. While numerous nonlinear methods for
online learning do exist in machine learning literature, a complete survey is beyond the scope of this article. The recent
paper on online sequential extreme learning machines (OS-ELM) [16] surveys popular online learning algorithms in
the context of classification and regression and develops an efficient algorithm based on recursive least squares. The
2
OS-ELM algorithm seems to be the present state of the art (although some variants have been proposed such as
[17, 18, 19]) for classification/regression problems achieving a global optimal solution, high generalization accuracies
and most importantly, in quick time. Also, based on observations from our previous work [20], we choose ELM over
other popular methods like neural networks and support vector machines for the HCCI engine problem. It has been
shown that both polynomial and linear methods were inferior in terms of prediction accuracy [12, 11] although they
have simple algorithms suitable for online applications. The online variants of SVM usually work by approximating
the batch (offline) loss function so that data can be processed sequentially [21, 22] and achieve accuracies similar to
that of the offline learning counterparts. However, SVMs comes with a high computation and memory requirement to
be used efficiently on a memory limited system such as the engine control unit [13]. Thus we prefer ELM over SVM
and other state of the art nonlinear models.
In spite of its known advantages, an over-parameterized ELM may suffer from ill-conditioning problem when a re-
cursive least squares type update is performed (as in OS-ELM). This sometimes results in poor regularization behavior
[23, 24, 19, 25], which leads to an unbounded growth of the model parameters and unbounded model predictions. This
may not be a serious problem for many applications as the model usually improves as more data becomes available.
However, for control problems in particular, if decisions are made simultaneously based on the online learned model
(as in adaptive control [26]), it is critical that the parameter estimation algorithm behaves in a stable manner so that
control actions can be trusted at all times. Hence a guarantee of stability and parameter boundedness is of extreme
importance. To address this issue, we propose the SG-ELM, a stable online learning algorithm based on stochastic
gradient descents and extreme learning machines. By extending ELM to include a notion of stable learning, we hope
that the simplicity and generalization power of ELM can be retained to learn stable models for real-time control ap-
plications. We use a Lyapunov approach to prove both asymptotic stability of estimation error and boundedness in the
estimated parameters suitable for identification of nonlinear dynamic systems. Using the Lyapunov approach, we de-
termine an upper bound for the learning rate of SG-ELM that seems to avoid bad regularization that may arise during
online learning. These are the main contributions of this article. Further, we also apply the SG-ELM algorithm to two
real-world HCCI identification problems including online state estimation and online operating boundary estimation
which is a novel application of extreme learning machines.
The remainder of the article is organized as follows. The ELM modeling approach is described in Section 2
along with algorithm details on batch (offline) learning as well as the present state of the art; the OS-ELM algorithm.
In Section 3, the stochastic gradient based ELM algorithm is derived along with a stability proof. In Section 4,
the background on HCCI engine and experimentation are discussed. Sections 5 and 6 cover the discussions on the
application of the SG-ELM algorithm on the two applications, followed by conclusions in Section 7.

3
2. Extreme Learning Machines

Extreme Learning Machine (ELM) is an emerging learning paradigm for multi-class classification and regression
problems [27, 28]. An advantage of the ELM method is that the training speed is extremely fast, thanks to the random
assignment of input layer parameters which do not require adaptation to the data. In such a setup, the output layer
parameters can be analytically determined using a least squares approach. Some of the attractive features of ELM
[27] include the universal approximation capability of ELM, the convex optimization problem of ELM resulting in
the smallest training error without getting trapped in local minima, closed form solution of ELM eliminating iterative
training and better generalization capability of ELM [28]. In comparison, a backpropagation neural network has the
same objective function as that of ELM but they often get trapped in local minima whereas ELM don’t. Support
vector machines on the other hand, solves a convex optimization problem but the computation involved is quite high
and running times are slow for large datasets. Thus, ELM appear to be very efficient both in terms of accuracy and
running times compared to several state-of-the-art algorithms.
Consider the following data set
{(x1 , y1 ), ..., (xN , yN )} ∈ X, Y ,


(1)

where N denotes the number of training samples, X denotes the space of the input features and Y denotes labels
whose nature differentiate the learning problem in hand. For instance, if Y takes integer values {1,2,3,..} then the
problem is referred to as classification and if Y takes real values, it becomes a regression problem. ELMs are well
suited for solving both regression and classification problems faster than state of the art algorithms [28]. A further
distinction could be made depending on the availability of training data during the learning process, as offline learning
(or batch learning) and online learning (or sequential learning). Offline learning could make use of all training data
simultaneously as all data is available to the algorithm and time is generally not a limiting factor. So it is possible to
have the model see the data several times (iterations) so that the best accuracy can be achieved. On the other hand,
there may be situations where offline learning becomes infeasible and one has to resort to online learning, such as
those involving high velocity steaming data where time taken for learning becomes a bottleneck. In an online learning
setting, data is available one-by-one or batch-by-batch and needs to be processed with limited computational effort
and storage. Further, inference is required to be made with each new available data along with the ones recorded in
the past. In this work, the online setting is considered as the data streaming from the engine is at a rate of 21 Hz 1 and
it becomes important to learn from the streaming data.

2.1. Batch (Offline) ELM

When the entire training data is available and a model is required to be learned using all the available data, batch
learning is adopted. In this case, the ELM algorithm involves solving the following optimization problem similar to

1 at a typical engine speed of 2500 rotations per minute

4
that of a ridge regression
n o
min kHW − Yk2 + λkWk2 (2)
W

H T = ψ(WrT x(k) + br ) ∈ Rnh ×1 , (3)

where λ represents the regularization coefficient determined using cross-validation, Y represents the vector of labels,
ψ represents the hidden layer activation function (sigmoidal, sinusoidal, radial basis etc [28]) and Wr ∈ Rn×nh , W ∈
Rnh ×yd represents the input and output layer parameters respectively. Here, n represents the dimension of inputs x(k),
nh represents the number of hidden neurons of the ELM model, H represents the hidden layer output matrix and yd
represents the dimension of outputs Y. The matrix Wr consists of randomly assigned elements that maps the input
vector to a high dimensional feature space while br ∈ Rnh is a bias component assigned in a random manner similar to
Wr . The number of hidden neurons determines the expressive power of the transformed feature space. The elements
can be assigned based on any continuous random distribution [28] and remains fixed during the learning process.
Hence the training reduces to a single step calculation given by equation (4). The ELM decision hypothesis can be
expressed as in equation (5) for classification and as in equation (6) for regression. It should be noted that the hidden
layer and the corresponding activation functions give a nonlinear mapping of the data, which if eliminated, the ELM
model becomes a linear least squares (Linear LS) model and is considered as one of the baseline models in this study.
 −1
W ∗ = H T H + λI H T Y (4)
 
f (x) = sgn W T [ψ(WrT x + br )] . (5)

f (x) = W T [ψ(WrT x + br )] (6)

Since training involves a linear least squares solution with a convex objective function, the solution obtained by
ELM is extremely fast and is a global optimum for the chosen nh , Wr and br . The above formulation for classification
(5), is not designed to handle imbalanced or skewed data sets. As a modification to weigh the minority class data
more, a simple weighting method can be incorporated in the ELM objective function (2) as
n o
min (HW − Y)T Γ(HW − Y) + λW T W (7)
W

 
 γ1 0 . .0 
 
γ2 . . 0 

 0 
Γ =  
 . . . . 0 

 
. . γN 

0 0


1 majority class data



γi = 

 (8)

r × f s

 minority class data

5
where Γ represents the weight matrix, r represents the ratio of number of majority class data to number minority class
data and f s represents a scaling factor to be tuned for a given data set [13]. This results in the training step given by
equation (9) and the decision hypothesis takes the same form as in equation (5):
 −1
W ∗ = H T ΓH + λI H T ΓY. (9)

2.2. Online Sequential ELM (OS-ELM)

The OS-ELM [16] is a recursive version of the batch ELM algorithm. This version of the algorithm is used for
online learning purposes where data is processed one-by-one or batch-by-batch and the model parameters are updated
after which the used data is not required to be stored. In this process, training involves two steps - initialization step
and sequential learning step. During the initialization step, a set of data observations (N0 ) are required to initialize the
H0 and W0 by solving the following optimization problem
n o
min kH0 W0 − Y0 k2 + λkW0 k2 (10)
W0

H0 = [g(WrT x0 + br )]T ∈ RN0 ×nh . (11)

The solution W0 is given by

W0 = K0−1 H0T Y0 (12)

where K0 = H0T H0 + λI. Suppose given another new data x1 , the problem becomes
    2
 H   Y 
0 0
min   W1 −  . (13)
 
W1    
H1  Y1 

The solution can be derived as

W1 = W0 + K1−1 H1T (Y1 − H1 W0 )

K1 = K0 + H1T H1 .

Based on the above, a generalized recursive algorithm for updating the least-squares solution can be computed as
follows
Mk+1 = Mk − Mk Hk+1
T
(I + Hk+1 Mk HK+1
T
)−1 Hk+1 Mk (14)

Wk+1 = Wk + Mk+1 Hk+1

T
(Yk+1 − Hk+1 Wk ) (15)

where M represents the covariance of the parameter estimate.

6
3. Stochastic Gradient Based ELM Algorithm

In this section, we propose an extension of the ELM algorithm for online learning using stochastic gradient descent
(SGD). Stochastic gradient descent methods have been popular for several decades for online learning but practically
limited because of poor optimization characteristics (failure to converge to an absolute minimum, for instance) and
slow convergence rates. However, only recently, the asymptotic behavior of SGD methods has been analyzed indicat-
ing that SGD methods can be very powerful for learning large data sets [29, 30]. SGD based algorithms have been
developed successfully for perceptron models, K-means, SVM and Lasso [29]. In this work, we propose to use the
simple and scalable SGD algorithm to extreme learning machines and derive stability properties, so that an online
learning algorithm useful for control purposes can be developed.
The justification of SGD based algorithms in machine learning can be briefly discussed as follows. In any learning
problem, three types of errors are encountered, namely the approximation error, the estimation error and the optimiza-
tion error [29], and the expected risk Eexp ( f ) and the empirical risk Eemp for a supervised learning problem can be
given by
Z
Eexp ( f ) = l( f (x), y)dP(x, y)
N
1 X
Eemp ( f ) = l( f (xi ), yi )
N i=1

where l( f (x), y) denote the loss function between the prediction f (x) and label y, P(x, y) denote the joint probability
density of x and y. Let f ∗ = argmin f Eexp ( f ) be the best possible prediction function. In practice, the prediction
function is chosen from a family of parametric functions denoted by F . Let fF∗ = argmin f ∈F Eexp ( f ) be the best pre-
diction function chosen from a parameterized family of functions F . When a training data set becomes available, the
empirical risk becomes a proxy for the expected risk for the learning problem [31]. Let f¯F∗ = argmin f ∈F Eemp ( f ) be
the solution that minimizes the empirical risk. However, the global solution is not typically obtained because of com-
putational limitations and hence the solution of the learning problem is reduced to finding f¯F = argmin f ∈F Eemp ( f ).
Using the above setup, the approximation error (Eapp ) is the error introduced in approximating the true function
space with a family of functions F , the estimation error (Eest ) is the error introduced in optimizing over Eemp ( f )
instead of Eexp ( f ), the optimization error (Eopt ) is the error induced as a result of stopping the optimization to f¯F . The
total error Etot can be expressed as

Eapp = Eexp ( f ∗ ) − Eexp ( fF∗ )

Eest = Eexp ( fF∗ ) − Eemp ( f¯F∗ )

Eopt = Eemp ( f¯F∗ ) − Eemp ( f¯F )

Etot = Eapp + Eest + Eopt

The following observations are taken from the asymptotic analysis of SGD algorithms [29, 32].
7
 1. The empirical risk Eemp ( f ) is only a surrogate for the expected risk Eexp ( f ) and hence an increased effort to
minimize Eopt may not translate to better learning. In fact, if Eopt is very low, there is a good chance that the
prediction function will over-fit the training data.
2. SGD are worst optimization algorithms (in terms of reducing Eopt ) but they minimize the expected risk relatively
quickly. Therefore, in the large scale setup, when the limiting factor is computational time rather than the
number of examples, SGD algorithms perform asymptotically better.
3. SGD results in a faster convergence when the loss function has strong convexity properties.

The last observation is key in developing our algorithm based on ELM models. The ELM models have a squared
loss function and when the hidden neurons are randomly assigned and fixed, the training translates to solving a convex
optimization problem. This motivates the use of ELM models for performing SGD based learning. The SGD based
algorithm can be derived for the ELM models as follows.

3.1. SG-ELM Parameter Update

Let (xi , yi ) where i = 1, 2, ..N be the streaming data in consideration. The data can be considered to be available to
the algorithm from a one-by-one continuous stream or artificially sampled one-by-one from a very large data set. Let
the ELM empirical risk be defined as follows
N
1X
J(W) = min kyi − φTi Wk2
W 2
i=1
( )
1 1
= min ky1 − φ1 Wk + .. + kyN − φN Wk
T 2 T 2
W 2 2
= min {J1 (W) + J2 (W) + .. + JN (W)} . (16)
W

where W ∈ Rnh ×yd , yi ∈ R1×yd φ ∈ Rnh ×yd is the hidden layer output (see H T in equation (3)). If an error ei ∈ R1×yd can
be defined as (yi − φTi W), the learning objective for a data observation i can be given by

1 T
Ji (W) = e ei
2 i
1
= (yi − φTi W)T (yi − φTi W)
2
1 T 1
= y yi + W T φi φTi W − yTi φTi W
2 i 2
∂Ji
= φi φTi W − φi yi = φi (φTi W − yi )
∂W
= −φi ei . (17)

8
In a regular gradient descent (GD) algorithm, the gradient of J(W) is used to update the model parameters as follows.

∂J ∂J1 ∂J2 ∂JN

= + + .. +
∂W ∂W ∂W ∂W
∂J
⇒ = −φ1 e1 − φ2 e2 − .. − φN eN
∂W
∂J
Wk+1 = Wk − ΓS G
∂W
= Wk + ΓS G (φ1 e1 ) + .. + ΓS G (φN eN ) (18)

where k is the iteration count, ΓS G ∈ Rnh ×nh represents the step size or update gain matrix for the GD algorithm.
It can be seen from equation (18) that the parameter matrix W is updated based on gradients calculated from
all the available examples. If the number of data observations is large, the gradient calculation can take enormous
computational effort. The stochastic gradient descent algorithm considers one example at a time and updates W based
on gradients calculated from (xi , yi ) as shown in

Wi+1 = Wi + ΓS G (φi ei ). (19)

From equation (18), it is clear that the optimal W is a function of gradients calculated from all the examples. As a
result, as more data becomes available, W converges close to its optimal value in SGD algorithm. Processing data
one-by-one significantly reduces the computational requirement making the algorithm scale well to large data sets.
In order to handle class imbalance, the algorithm in (19) can be modified by weighting the minority class data
more. The modified algorithm can be expressed as

Wi+1 = Wi + Γimb ΓS G (φi ei ) (20)

where Γimb = r × f s , r and f s represent the imbalance ratio (a running count of majority class data to minority class
data until that instant) and the scaling factor that needs to be tuned to obtain tradeoffs between high false positives and
missed detections for a given application.

3.2. Stability Analysis

The stability analysis of the SG-ELM algorithm can be derived as follows. The ELM structure makes the analysis
simple and similar to that of a linear gradient based algorithm [33].
The instantaneous prediction error ei (Here the error e and output y are transposed as opposed to their previous
definition in Section 3.1 for ease of derivations) can be expressed in terms of the parametric error (W̃ = W∗ − W) as

ei = yi − W T φi

= W∗T φi − W T φi

= W̃ T φi (21)

9
where W∗ represents true model parameters. Further, the parametric error dynamics can be obtained as follows.

W̃i+1 = W∗ − Wi+1

= W∗ − Wi − ΓS G φi eTi

= W̃i − ΓS G φi eTi (22)

Consider the following positive definite, decrescent and radially unbounded [33] Lyapunov function V

V(W̃) = tr(W̃ T Γ−1

S G W̃) (23)

where tr represents the trace of a matrix.

∆V(W̃i ) = V(W̃i+1 ) − V(W̃i )

= T
tr(W̃i+1 Γ−1
S G W̃i+1 ) − tr(W̃i ΓS G W̃i )
T −1

= tr((W̃i − ΓS G φi eTi )T Γ−1

S G (W̃i − ΓS G φi ei ))
T

−tr(W̃iT Γ−1
S G W̃i )

= tr(−2W̃iT φi eTi + ei φTi ΓS G φi eTi )

= tr(−2ei eTi + ei φTi ΓS G φi eTi )

= −2eTi ei + eTi ei φTi ΓS G φi

= −2eTi ei + eTi φTi ΓS G φi ei

= −eTi MS G ei (24)

where MS G = 2 − φTi ΓS G φi . It can be seen that Vi+1 − Vi ≤ 0 if MS G > 0 or 2 − φTi ΓS G φi > 0 or

0 < λmax (ΓS G ) < 2 (25)

When (25) is satisfied, V(W̃) ≥ 0 is non-increasing in i and the limit

lim V(W̃) = V∞ (26)

k→∞

exists. From (24),

Vi+1 − Vi = −eTi MS G ei
∞
X X∞
(Vi+1 − Vi ) = − eTi MS G ei
i=0 i=0
X∞
⇒ eTi MS G ei = V(0) − V∞ < ∞ (27)
i=0
(28)

10
Also,
∞
X ∞
X
eTi Iei ≤ eTi MS G ei < ∞ (29)
i=0 i=0
when MS G > I or when
λmax (ΓS G ) < 1. (30)

Hence, when (30) is satisfied, ei ∈ L2 . From (19), (Wi+1 − Wi ) ∈ L2 ∩ L∞ . Using discrete time Barbalat’s lemma [34],

lim ei = 0 (31)
i→∞

lim Wi+1 = Wi (32)

i→∞

Hence, the SGD learning law in (19) guarantees that the estimated output ŷi converges to the actual output yi and
the model parameters W converge to some constant values. The parameters converge to the true parameters W∗ only
under conditions of persistence of excitation [33] in input signals of the system (amplitude and frequency richness of
x). Further, using boundedness of Vi , ei ∈ L∞ which guarantees that the online model predictions are bounded as long
as the system output is bounded. As the error between the true model and the estimation model converges to zero, the
estimation model becomes a one-step ahead predictive model of the nonlinear system. In the next few sections, we
evaluate our SG-ELM algorithm on a HCCI engine identification problem.

4. Homogeneous Charge Compression Ignition Engine

This section gives an overview of the homogeneous charge compression ignition engine system and experimenta-
tion. This section is intended for completeness and some of the content is reproduced from [11, 12, 13]. The engine
specifications are listed in Table 1 [11]. HCCI is achieved by auto-ignition of the gas mixture in the cylinder. The
fuel is injected early in the intake stroke and given sufficient time to mix with air forming a homogeneous mixture. A
large fraction of exhaust gas from the previous cycle is retained to elevate the temperature and hence the reaction rates
of the fuel and air mixture. The variable valve timing capability of the engine enables trapping suitable quantities of
exhaust gas in the cylinder.
The engine can be controlled using precalculated inputs such as injected fuel mass (FM in mg/cyc), crank angle at
intake valve opening (IVO), crank angle at exhaust valve closing (EVC), crank angle at start of fuel injection (SOI).
The valve events are measured in degrees after exhaust top dead center (deg eTDC) while SOI is measured in degrees
after combustion top dead center (deg cTDC). Other important physical variables that influence the performance of
HCCI combustion include intake manifold temperature T in , intake manifold pressure Pin , mass flow rate of air at
intake ṁin , exhaust gas temperature T ex , exhaust manifold pressure Pex , coolant temperature T c , fuel to air ratio (FA)
etc. The engine performance metrics are given by combustion phasing indicated by the crank angle at 50% mass
fraction burned (CA50), combustion work output given by net indicated mean effective pressure (IMEP 2 ). For further
reading on HCCI combustion and related variables, please refer [35].

2 IMEP and NMEP has been interchangeably used in this paper although both refers to the net quantity

11
 Table 1: Specifications of the experimental HCCI engine

Engine Type 4-stroke In-line

Fuel Gasoline
Displacement 2.0 L
Bore/Stroke 86/86 mm
Compression Ratio 11.25:1
Injection Type Direct Injection
Variable Valve Timing with
hydraulic cam phaser having
Valvetrain 119 degree constant duration
defined at 0.25mm lift, 3.5mm peak
lift and 50 degree crank angle
phasing authority
HCCI strategy Exhaust recompression
using negative valve overlap

4.1. Experiment Design

In order to identify both models for HCCI state variables as well as models for dynamic operating boundary in
transient operation, we need to design appropriate experiments to obtain dynamic data from the engine. The modeled
variables such as engine states and operating envelope are dynamic variables and in order to capture both transient and
steady state behavior, a set of dynamic experiments is conducted at constant rotational speeds and naturally aspirated
conditions (no supercharging/turbocharging) by varying FM, IVO, EVC and SOI in a uniformly random manner. At
every step, the engine makes a transition between two set points and the transition is recorded as temporal data. In
order to capture several such transients, an amplitude modulated pseudo-random binary sequence (A-PRBS) has been
used to design the excitation signals for FM, IVO, EVC and SOI. A-PRBS excites the engine at different amplitudes
and frequencies exploring the operating space of the engine for the identification problem considered in this work.

4.2. HCCI Instabilities

3
A subset of the data collected from the engine is shown in Fig. 1 where it can be observed that for some
combinations of the inputs (left figures), the HCCI engine misfires (seen in the right figures where IMEP drops below
0 bar). HCCI operation is limited by several phenomena that lead to undesirable engine behavior. As described in [36],
the HCCI operating range is conceptually constrained to a small region of permissible unburned (pre-combustion) and
burned (post-combustion) charge temperature states. A sufficiently high unburned gas temperatures are required to
achieve ignition in the HCCI operating range without which a complete misfire will occur. If the resulting combustion

3 This figure has been taken from [12] for completeness. We will obtain copyright permission from the journal upon acceptance.

12
 Fuel Mass (mg/cyc)
11

IMEP (bar)
10 4

9
2
8
0
7
5200 5400 5600 5800 6000 5200 5400 5600 5800 6000

CA50 (deg aTDC)

IVO (deg aCTDC)

80
120 60
40
100 20
0
80 −20
5200 5400 5600 5800 6000 5200 5400 5600 5800 6000
EVC (deg aCTDC)

Rmax (bar/deg)
−80 10
−90
5
−100
−110
0
5200 5400 5600 5800 6000 5200 5400 5600 5800 6000
SOI (deg aETDC)

2
360
340
λ (−)

320 1.5

300
280 1
5200 5400 5600 5800 6000 5200 5400 5600 5800 6000

Figure 1: A subset of the HCCI engine experimental data showing A-PRBS inputs and engine outputs. The misfire regions are shown in dotted
rectangles. The data is indexed by combustion cycles.

cannot achieve sufficiently high burned gas temperatures, commonly occurring in conditions with low fuel to diluent
ratios or late combustion phasing, various degrees of quenching can occur resulting in reduced work output and
increased hydrocarbon and carbon monoxide emissions. Under some conditions, this may lead to high cyclic variation
due to the positive feedback loop existing through the trapped residual gas [14, 15]. Operation with high burned gas
temperature, although stable and commonly reached at higher fueling rates where the fuel to diluent ratio is also
high, yields high heat release and thus pressure rise rates that may pose challenges for engine noise and durability
constraints. A discussion of the temperatures at which these phenomena occur may be found in [36].
HCCI operation is limited by a combination of the above instabilities and during a highly transient operation, it
may be challenging to reactively respond to instabilities. Thus, a proactive means by which such instabilities can be
determined is by developing a dynamic model for the operating envelope of the engine discussed in Section 6.

4.3. Learning The HCCI Engine Data

In the HCCI modeling problem, both the inputs and the outputs of the engine are available as sensor measurements
creating a setting for supervised learning. The HCCI engine is a nonlinear dynamic system and sensor measurements
represent discrete time sequences. The input-output behavior can be modeled using a nonlinear auto regressive model

13
with exogenous input (NARX) [37] as follows

y(k) = fNARX [u(k − 1), .., u(k − nu ), y(k − 1), .., y(k − ny )] (33)

where u(k) ∈ Rud and y(k) ∈ Ryd represent the inputs and outputs of the system respectively, k represents the discrete
time index, fNARX (.) represents the nonlinear function mapping specified by the model, nu , ny represent the number of
past input and output samples required (order of the system) while ud and yd represent the dimension of inputs and
outputs respectively. The NARX model appears to do a good job for learning arbitrary complex nonlinear mapping
of the engine [11]. Let x represent the augmented input vector obtained by time-embedding the input and output
measurements from the system.

x = [u(k − 1), .., u(k − nu ), y(k − 1), .., y(k − ny )]T (34)

The measurement sequence can be converted to the form of training data

{(x1 , y1 ), ..., (xN , yN )} ∈ X, Y

(35)

where N denotes the number of training samples, X denotes the space of the input features (Here X = Rud nu +yd ny and
Y = R for regression and Y = {+1, −1} for a binary classification). The above conversion of system measurements
to training data is a natural definition for a series-parallel model architecture and the models can be used to perform a
one-step ahead prediction (OSAP) i.e., given a set of measurements until time index k, the model predicts the output
at time k + 1 (see equation (36)). A parallel architecture on the other hand can be used to perform multiple step ahead
predictions (MSAP) 4 by feeding back the predictions of the OSAP model in a recurrent manner (see equation (37)).
The series-parallel and parallel architectures are well explained in [38].

ŷ(k + 1) = fˆNARX [u(k), .., u(k − nu + 1), y(k), .., y(k − ny + 1)] (36)

ŷ(k + n pred ) = fˆNARX [u(k + n pred − 1), .., u(k − nu + n pred ), ŷ(k + n pred − 1), .., ŷ(k − ny + n pred )] (37)

The OSAP model is used for training as existing simple 5 training algorithms can be used and once the model becomes
accurate for OSAP, it can be converted to a MSAP model in a straightforward manner. The MSAP model can be used
for making long term predictions useful for predictive control [5, 39].

5. Application Case Study 1: Online regression learning for system identification of an HCCI Engine.

The problem considered in this case study is to develop a predictive model of the state variables of the HCCI
engine. The state variables of an engine are the fundamental quantities that represent the engine’s state of operation.

4 MSAP predictions are necessary for planning trajectories for a given engine operation, such as in the case of optimal control or trajectory
optimization
5 training a parallel model usually involves calculating gradients with respect to time which is quite challenging.

14
As a consequence, these variables also influence the performance of the engine such as fuel efficiency, emissions and
stability, and are required to be monitored/regulated carefully. The significance of the state variables for control and
a data based modeling approach were recently analyzed by the authors [11] where an offline model was developed
using archived data. In this paper, an online learning framework for modeling the state variables of HCCI engine such
as the IMEP and CA50, is developed and is shown to be comparable to that of the offline models.
Online learning algorithms using OS-ELM, SG-ELM are compared for regression performance evaluated using
one-step ahead and multi-step ahead predictions. A linear model and an offline trained nonlinear ELM model similar
to the one in [11] are included as baselines. The linear baseline model is included to justify the benefits of adopting a
nonlinear model while the offline trained model is included to show the effectiveness of online algorithms in capturing
the underlying behavior fully in spite of processing data sequentially.

5.1. Model Setup and Evaluation Metric

For the purpose of demonstration, the variables IMEP and CA50 are considered as outputs whereas the control
variables such as fueling (FM), exhaust valve closing (EVC) and fuel injection timing (SOI) are considered inputs.
Transient data from the HCCI engine at a constant speed of 1800 RPM and naturally aspirated conditions is used. A
NARX model as shown in section 4.3 is considered where u = [F M EVC S OI]T and y = [I MEP CA50]T , nu
and ny chosen as 1 (tuned by trial and error). The nonlinear model approximating fNARX is initialized to an extreme
learning machine model with random input layer weights and random values for the covariance matrices and output
layer weights. Four different models are considered including the state of the art OS-ELM algorithm, the proposed
SG-ELM algorithm, a baseline offline (batch) ELM (O-ELM) and a baseline linear system identification model. The
purpose of the baseline offline ELM algorithm is to evaluate the efficiency of the online learning models in learning
the HCCI behavior completely as an offline ELM model would do. The offline ELM model is expected to produce
an accurate model as it has sufficient time, computation and utilization of all training data simultaneously to learn the
HCCI behavior sufficiently well. The purpose of the linear baseline model is to justify the use of a nonlinear model
for HCCI dynamics.
All the nonlinear models consist of 100 hidden units with fixed randomized input layer parameters. About 11000
cycles of data is considered one-by-one as it is sampled by the engine ECU and model parameters updated in a
sequential manner. After the training phase, the parameter update is switched off and the models are evaluated for
the next 5100 cycles of data for one step ahead predictions. Further, to evaluate if the learned models represent the
actual HCCI dynamics, the multi-step ahead prediction of the models are compared using about 600 cycles of data.
It should be noted that both the one-step ahead and multi-step ahead evaluations were done using data unseen during
the training phase.
The parameters of each of the models are tuned for the given dataset. As recommended by OS-ELM [16], about
800 cycles of data was used for initializing the output layer parameters W0 and covariance matrix M0 (see equations
(14) and (15)). The initialization was performed using the batch ELM algorithm [28]. In order to have a fair compar-

15
 Parameter convergence of OS−ELM zoomed−in
5
1
0.5

0 0
−0.5
−1
−5
0 2000 4000 6000 8000 10000 12000 8000 9000 10000

Parameter convergence of SG−ELM zoomed−in

1
0.18
0 0.17
0.16
0.15
−1
0.14
0.13
−2
0 2000 4000 6000 8000 10000 12000 0 5000 10000
Engine Cycles Engine Cycles

Figure 2: Comparison of parameter evolution for the OS-ELM and SG-ELM algorithms during online learning (each engine cycle corresponds
to a sample of engine data processed by the models). A zoomed-in plot (figures to the right) shows that the parameter update for OS-ELM is
more aggressive compared to SG-ELM. Although both OS-ELM and SG-ELM parameters are initialized to the same values, the parameters of
OS-ELM continue to grow in magnitudes compared to the SG-ELM. Note that small parameter values indicate good regularization. This plot gives
a qualitative visualization of this behavior.

ison, the W0 is used as an initial condition for both OS-ELM and SG-ELM. The only parameter of SG-ELM, namely
the gradient step size was tuned to be ΓS G = 0.0008 I100 for best accuracy.
The performance of the models are measured using normalized root mean squared error (RMSE) given by
v
u
t X n X yd
1
RMS E = (yi − ŷij )2 (38)
n i=1 j=1 j

where both yij and ŷij are normalized to lie between -1 and +1.

5.2. Results and Discussion

On performing online learning, it can be observed from Fig. 2 that the parameters of OS-ELM grow more aggres-
sively as compared to the SG-ELM. In spite of both models having the same initial conditions, the step size parameter
ΓS G for SG-ELM gives additional control over the parameter growth and keeps them bounded as shown in section 3.2.

16
On the other hand, OS-ELM doesn’t have any control over the parameter evolution. It is governed by the evolution
of the co-variance matrix M (see equation (14)). It is expected that the co-variance matrix M would add stability to
the parameter evolution but in practice, it tends to be more aggressive especially when having correlated data and
over-parameterized models, leading to potential instabilities as reported by [23, 24, 19, 25]. As a consequence, the pa-
rameter values for SG-ELM remain small compared to the OS-ELM (the norm of estimated parameters for OS-ELM
is 16.64 and SG-ELM is 3.71). This has a significant implication in the statistical learning theory [31]. A small norm
of model parameters implies a simpler model which results in good generalization. Although this effect is slightly re-
flected in the results summarized in prediction results summarized in Table 2 (see SG-ELM having the lowest MSAP
RMSE), it is not significantly better for this problem possibly because of insufficient data for convergence. The value
of ΓS G has to be tuned correctly along with sufficient training data in order to ensure parameter convergence. Ulti-
mately, the online learning mechanism is aimed to run along with the engine and hence the slow convergence may not
be an issue in a real application.

Table 2: Performance comparison of OS-ELM and SG-ELM for the HCCI online regression learning problem. A baseline linear model and an
offline trained ELM model (O-ELM) are also included for comparison. The offline O-ELM algorithm has access to all the data and can use the
available memory and computational power, so its training time is not compared with the online algorithms. The RMSE values are averaged over
100 different trials.

Training OSAP MSAP

6
Time in s RMSE RMSE
Linear 1.0111 0.1277 0.1859
OS-ELM 9.6563 0.0952 0.1018
SG-ELM 1.5624 0.1050 0.0955
O-ELM - 0.1018 0.1027

The prediction results as well as training time6 for the online models are compared in Table 2 where each algorithm
is run for 100 trials and the average RMSE is reported. It can be observed that the computational time for SG-ELM
is significantly less (about 6.2 times) compared to OS-ELM showing the time gain in eliminating the covariance
estimation step. The reduction in computation is expected to be more pronounced as the dimension and complexity
of the data increase. It could be seen from Table 2 that the one-step ahead prediction accuracies (OSAP RMSE)
of the nonlinear models are similar, and OS-ELM winning marginally. On the other hand, the multi-step prediction
accuracies (MSAP RMSE) are similar for the nonlinear models with SG-ELM performing marginally better. The
MSAP accuracy reflect the generalization performance of the model and is more crucial for the modeling problem as
the models ultimately feed its prediction to a predictive control framework that requires accurate and robust predictions
of the engine several steps ahead of time. From our understanding on model complexity and generalization error, a

6 The training time is the time taken for training without considering tuning of hyper-parameters such as number of hidden neurons etc.

17
model that is less complex (indicated by minimum norm of parameters [28, 31]) tend to generalize better, which is
again demonstrated by SG-ELM. The performance of the linear baseline model is significantly low compared to the
nonlinear models justifying the need for nonlinear models for the HCCI system. In order to show that the results
of SG-ELM are statistically significant with respect to OS-ELM, a pairwise t-test is performed [40, 41] using 100
bootstrapped sub-sample instances7 . The p-values of the pairwise t-test for the OSAP RMSE and MSAP RMSE are
2.9632E-96 and 3.8491E-76, indicating that the results of the SG-ELM is statistically significant from that of the
OS-ELM with a very low significance level (high probability that the two algorithms are statistically significant).
The MSAP predictions of the models are summarized in Figures 3a-3d where model predictions for NMEP and
CA50 are compared against real experimental data. Here the model is initialized using the experimental data at the
first instant and allowed to make predictions recursively for several steps ahead. It can be seen that the nonlinear
models outperform the linear model and at the same time the online learning models perform similar to the offline
trained models indicating that online learning can fully identify the engine behavior at the operating condition where
the data is collected. It should be noted that this task is a case of multi-input multi-output modeling which adds
some limitations to the SG-ELM methods. When the model complexity increases, the SG-ELM may require more
excitations for convergence, as opposed to OS-ELM which converges more aggressively (although possibly losing
stability). Further, tuning the learning rate ΓS G may be time-consuming for models predicting multiple outputs with
different noise characteristics and stability requirements.

6. Application Case Study 2: Online classification learning (with class imbalance) for identifying the dynamic
operating envelope of an HCCI Engine

The problem considered in this case study is to model the dynamic operating envelope of the HCCI engine. The
dynamic operating envelope of an engine can be defined as the stable operating space of the engine. The significance
of the operating envelope and data based modeling approaches were recently introduced by the authors [13] where
an offline model was developed using archived data. In this paper, an online learning framework for modeling the
operating envelope of HCCI engine is developed is shown to be as good as the offline models in determining the HCCI
operating envelope.
We consider the operating envelope defined by two common HCCI unstable modes - a complete misfire and a
high variability combustion (a more detailed description is given in section 4.2) is studied. The problem of identifying
the HCCI operating envelope using experimental data can be posed as a binary classification problem. The engine
sensor data can be labeled as being stable or unstable depending on some engine based heuristics [13]. Further, the

7 For the pairwise t-test, the proposed SG-ELM is compared with the competing OS-ELM algorithm and a statistical test is performed with the
null hypothesis being that the two algorithms are not statistically significant. About 100 trials were performed with a subset of data being sampled
with replacement from the training data set for both algorithms and the OSAP RMSE and MSAP RMSE were measured. Using the 100 values of
OSAP and MSAP RMSE for both SG-ELM and OS-ELM, the pairwise t-test was carried out.

18
 3.5 3.5
NMEP

NMEP
3 3

Predicted Predicted
Actual Actual
2.5 2.5
0 100 200 300 400 500 600 0 100 200 300 400 500 600
Engine Cycles Engine Cycles

0 0

−5 −5
CA50

CA50

−10 −10

−15 −15
0 100 200 300 400 500 600 0 100 200 300 400 500 600
Engine Cycles Engine Cycles

(a) OS-ELM MSAP Prediction (b) SG-ELM MSAP Prediction

3.5 3.5

3
NMEP

NMEP

3
2.5
Predicted Predicted
Actual Actual
2.5 2
0 100 200 300 400 500 600 0 100 200 300 400 500 600
Engine Cycles Engine Cycles

0 0

−5
−5
CA50

CA50

−10
−10
−15

−15 −20
0 100 200 300 400 500 600 0 100 200 300 400 500 600
Engine Cycles Engine Cycles

(c) O-ELM MSAP Prediction (d) Linear MSAP Prediction

Figure 3: 600-step ahead prediction results of OS-ELM, SG-ELM, O-ELM and the linear model are compared. The OS-ELM, SG-ELM and linear
models are learned online using 11000 cycles of data while the O-ELM model is trained using the same data but in an offline manner. The 600-step
ahead prediction is performed on an unseen dataset. For each of the models, NMEP and CA50 predictions are compared to the experimentally
recorded values. It has to be noted that the control inputs at the 600 cycles are used while NMEP and CA50 are recurrently fed back to the model
to perform multi-step ahead predictions.

19
engine dynamic data consists of a large number of stable class data compared to unstable class data, which introduces
a class imbalance. A cost-sensitive approach that modifies the objective function of the learning system to weigh the
minority class data more heavily, is preferred over under-sampling and over-sampling approaches [13] and is used in
this study.
Online learning algorithms using OS-ELM, SG-ELM are compared for classification performance. A linear clas-
sification model and an offline trained nonlinear ELM model similar to the one in [13] are included as baselines to
make similar justifications as in the previous case study. The linear baseline model is included to justify the benefits of
adopting a nonlinear model while the offline trained model is included to show the effectiveness of online algorithms
in capturing the underlying behavior fully in spite of processing data sequentially.

6.1. Model Setup and Evaluation Metric

The HCCI operating envelope is a function of the engine control inputs and engine physical variables such as
temperature, pressure, flow rate etc. Also, the envelope is a dynamic system and so a predictive model requires the
measurement history up to an order of Nh . The dynamic classifier model can be given by

ŷk+1 = sgn( f (xk )) (39)

where sgn(.) represents the sign function, ŷk+1 indicates model prediction for the future cycle k + 1, f (.) can take any
structure depending on the learning algorithm and xk is given by

xk = [IVO, EVC, F M, S OI, T in , Pin , ṁin , T ex , Pex , T c , FA, I MEP, CA50]T (40)

at cycle k up to cycle k − Nh + 1. In the following sections, the function f (.) is learned using the available engine ex-
perimental data using OS-ELM and SG-ELM algorithms. The engine measurements and their time histories (defined
by xk ) are considered inputs to the model while the stability labels are considered outputs. The feature vector is of
dimension n=39 includes sensor measurements such as FM, IVO, EVC, SOI, T c , T in , Pin , ṁin , T ex , Pex , IMEP, CA50
and FA along with Nh = 1 cycles of history (see (40)). The engine experimental data is split into training and testing
sets. The training set consists of about 14300 cycles of data processed one-by-one as sampled by the engine ECU.
After the training phase, the parameter update is switched off and the models are evaluated for the next 6200 cycles
of data for one step ahead classification. The ratio of number of majority class data to number minority class data (r)
for the training set is about 4.5:1 and for the testing set is 9:1. The nonlinear model approximating f (.) is initialized
to an extreme learning machine model with random input layer weights and random values for the covariance matri-
ces and output layer weights. All the nonlinear models consist of 10 hidden units with fixed randomized input layer
parameters. Similar to the previous case study, a small portion of the training data is used to initialize the ELM model
parameters as well as the covariance matrix. The SG-ELM parameter ΓS G is tuned to be 0.001 I10 using trial and
error. A weighted classification version of the algorithms is developed to handle the class imbalance problem. The
minority class data is weighted higher by r times f s where r is the imbalance ratio of the training data and is computed
online as the ratio of the number of majority class to number of minority class data until that instant.
20
 20 20

10 10
CA50

CA50
0 0

−10 −10
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
4 4
NMEP

NMEP
2 2

0 0

0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
30 30
Fuel input

Fuel input
20 20

10 10
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
Combustion Cycles Combustion Cycles

(a) OS-ELM (dataset 1) (b) SG-ELM (dataset 1)

Figure 4: Online classification results of OS-ELM and SG-ELM models showing CA50, NMEP and one input variable (fueling) for 2 different
unseen data sets. The color code indicates model prediction - green (and red) indicate stable (and unstable) prediction by the model. The horizontal
dotted line in the NMEP plot indicates misfire limit, dotted ellipse in CA50 plot indicates high variability instability mode while dotted rectangle
shows false alarms by model.

For the class imbalance problem considered here, a conventional classifier metric like the overall misclassification
rate cannot be used as it would find a biased classifier, i.e., it would find a classifier that ignores the minority class
data. For instance, a data set that has 95% of majority class data (with label +1) would achieve 95% classification
accuracy by predicting all the labels to be +1 which is obviously undesirable. Hence the following evaluation metric
used for skewed data sets is considered. Let T P and T N represent the total number of positive (stable operation) and
negative class (unstable modes) data classified correctly by the classifier. If N + and N − represent the total number
of positive and negative class data respectively, the true positive rate (TPR) and true negative rate (TNR), geometric
mean (GM) of TPR and TNR, and the total accuracy (TA) of the classifier can be defined as follows [42]. It should be
noted that the total accuracy and geometric mean weights the accuracy of majority and minority classes equally, i.e.,
they have high values only when both classes of data are classified correctly.

TP
T PR =
N+
TN
T NR =
N−
√
GM = T PR × T NR

TA = 0.5(T PR + T NR). (41)

6.2. Results and Discussion

The results of online imbalance classification can be summarized in Table 3 where computational time as well
as classification performance can be compared based on 100 different trials. It can be observed that for the HCCI
21
classification problem, all the models perform with similar average accuracies. Both OS-ELM and SG-ELM achieve
results similar to an offline model indicating completeness of learning. The SG-ELM has a slight advantage in terms
of training time because of the simplicity of SG-ELM compared to OS-ELM.

Table 3: Performance comparison of the nonlinear models (OS-ELM and SG-ELM) for the online class imbalance learning problem. A baseline
linear model and an offline trained ELM model (O-ELM) are also used for comparison. The O-ELM results are included for comparing classification
accuracies. The offline O-ELM algorithm has access to all the data and can use the available memory and computational power, so its training time
is not compared with the online algorithms. All values reported are averaged over 100 different trials.

Algorithms Training TPR TNR Total GM

Time6 in s Accuracy Accuracy
Linear 1.8648 0.9996 0.5665 0.7830 0.7524
OS-ELM 1.8443 0.7000 0.8713 0.7857 0.7791
SG-ELM 1.6011 0.8605 0.6923 0.7764 0.7714
O-ELM - 0.7530 0.8426 0.7978 0.7947

In the experiments above, it should be noted that for different initializations of the 100 trials, the model’s hyper-
parameters are not fine-tuned and so it may be possible to achieve better performance by fine-tuning. A further
experiment is conducted where the hyper-parameters of each algorithm are fine-tuned for one particular initialization
so that the best model is identified for further engine controls development. Ignoring the results of the linear model
(that had a large imbalance in TPR and TNR similar to the average results in Table 3), the results of the fine-tuned
nonlinear models are reported in Table 4. It can be seen that the accuracies of all the algorithms improved significantly
with SG-ELM slightly better and with a stability guarantee, indicates the suitability of SGD based online learning for
the HCCI problem. A subtle advantage observed for the OS-ELM is that, although the combined accuracy is slightly
inferior to that of the SG-ELM, the accuracies of the positive examples and negative examples are very close to each
other indicating that the model is well balanced to predict both majority class as well as minority class data well. The
SG-ELM on the other hand, in spite of fine-tuning the parameters, fails to achieve this. A further tuning can be done
to improve the accuracy of a particular class of data, typically sacrificing some accuracy predicting the other. In order
to show that the results of SG-ELM are statistically significant with respect to OS-ELM, a pairwise t-test is performed
[40, 41] using 100 bootstrapped sub-sample instances8 . The p-value of the pairwise t-test is 0.0208 which indicates
that the results of the SG-ELM is statistically significant from that of the OS-ELM with a significance level of 5%.
The models developed using OS-ELM and SG-ELM algorithms are used to make predictions on unseen engine
inputs and class predictions are summarized in Fig. 4, while quantitative results are included in Table 3. As mentioned

8 For the pairwise t-test, the proposed SG-ELM is compared with the competing OS-ELM algorithm and a statistical test is performed with the
null hypothesis being that the two algorithms are not statistically significant. About 100 trials were performed with a subset of data being sampled
with replacement from the full data set for both algorithms and the total accuracy was measured. Using the 100 values of total accuracy for both
SG-ELM and OS-ELM, the pairwise t-test was carried out.

22
Table 4: Comparison of best performance of the nonlinear models after fine tuning the parameters and with a good set of initial conditions. These
models are used for final prediction of the operating envelope of the HCCI engine.

Algorithms Training TPR TNR Total GM

Time6 in s Accuracy Accuracy
OS-ELM 1.8374 0.8328 0.8341 0.8335 0.8335
SG-ELM 0.9822 0.9876 0.7707 0.8792 0.8725
O-ELM - 0.8265 0.8569 0.8417 0.8416

earlier, the operating envelope is a decision boundary in the input space within which any input operates the HCCI
in a stable manner and any input outside the envelope might operate the engine in an unstable manner. The HCCI
state variables such as IMEP, CA50 and engine sensor observations such as T in , Pin , ṁin , T ex , Pex , T c at time instant k,
along with engine control inputs such as FM, EVC, SOI at time instant k + 1, are given as input to the models (see
(40)). The model predictions at time k + 1 are obtained. The engine’s actual response at time k + 1 is also recorded. A
data point is marked in red if the model predicts the engine operation to be unstable (labeled as -1) while it is marked
in green if the model predicts the data point to be stable (labeled as +1). In the figures, a dotted line in the NMEP
plot indicates the misfire limit, a dotted ellipse in CA50 plot indicates high variability instability mode while a dotted
rectangle indicates misclassified predictions by model. To understand the variation of NMEP and CA50 with changes
in control inputs, the fueling input (abbreviated as FM) is also included in the plots. It should be noted that FM is not
the only input for prediction and the signals are defined as in equation (40) but only the fueling input is shown in the
plots owing to space constraints.
It can be seen from the above plots that as a whole, both OS-ELM and SG-ELM models classify the HCCI engine
data fairly well in spite of the high amplitude noise inherent in the HCCI experimental data. The data consists of
step changes in FM, EVC and SOI and whenever a ‘bad’ combination of inputs is chosen, the engine either misfires
completely (see NMEP fall below misfire limit) or exhibits high variability combustion (see dotted ellipses). The
goal of this work as stated previously, is to predict if a future HCCI combustion event is stable or unstable based on
available measurements. The results summarized in Table 3 indicates that the developed models indeed accomplished
the goal with a reasonable accuracy. From Fig. 4, it is observed that the OS-ELM has some clear false alarms in
predicting stable class data (see dotted rectangles in the plots) while this is not observed for SG-ELM. This is not
surprising as the false alarm rate of SG-ELM (see Table 3) is pretty low9 . On the other hand, the SG-ELM has an
inferior TNR. By adjusting the weighting factor Γimb in equation (20), one can achieve a required tradeoff between
TPR and TNR as desired in the application.

9 The false alarm rate of SG-ELM is about 1.24%. In this study, the label of -1 corresponds to a ‘bad’ data and so false alarm rate corresponds
to false negative rate which is 1-TPR

23
7. Conclusion

A stochastic gradient descent based online learning algorithm for ELM has been developed, that guarantees stabil-
ity in parameter estimation suitable for control purposes. Further, the SG-ELM demands less computation compared
to the OS-ELM algorithm, as the covariance estimation step is eliminated. A stability proof is developed based on
Lyapunov approach. However, the SG-ELM algorithm might involve tedious tuning of step-size parameter as well
as suffer from slow convergence. The tuning of step-size parameter and convergence properties of SG-ELM will be
considered for future work.
The SG-ELM and OS-ELM algorithms are applied to develop models for state variables and dynamic operating
envelope of a HCCI engine to assist in model based control. The results from this article suggest that good gener-
alization performance can be achieved using both OS-ELM and SG-ELM methods but the SG-ELM might have an
advantage in terms of stability, crucial for designing robust control systems.
Although the SG-ELM appears to perform well in the HCCI identification problem, a comprehensive analysis
and evaluation on several benchmark data sets is required and will be considered for future. From an application
perspective, interesting areas for exploration include implementing the algorithm in real-time hardware, exploring a
wide operating range of HCCI operation and development of controllers.

Acknowledgment

This material is based upon work supported by the Department of Energy [National Energy Technology Labora-
tory] under Award Number(s) DE-EE0003533. This work is performed as a part of the ACCESS project consortium
(Robert Bosch LLC, AVL Inc., Emitec Inc.) under the direction of PI Hakan Yilmaz, Robert Bosch, LLC. Prof. X.
Nguyen is supported in part by NSF Grants CCF-1115769 and ACI-1047871.

Disclaimer

This report was prepared as an account of work sponsored by an agency of the United States Government. Neither
the United States Government nor any agency thereof, nor any of their employees, makes any warranty, express or
implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any infor-
mation, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights.
Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or
otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States
Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or
reflect those of the United States Government or any agency thereof.

24
References

[1] R. Thring, Homogeneous-charge compression-ignition engines, 1989, SAE paper 892068.

[2] M. Christensen, P. Einewall, B. Johansson, Homogeneous Charge Compression Ignition Using Iso-octane, Ethanol and Natural Gas - A
Comparison to Spark Ignition Operation, in: International Fuels & Lubricants Meeting & Exposition, Tulsa, OK, USA, 1997, SAE paper
972874.
[3] T. Aoyama, Y. Hattori, J. Mizuta, Y. Sato, An experimental study on premixed-charge compression ignition gasoline engine, in: International
Congress & Exposition, Detroit, MI, USA, 1996, SAE paper 960081.
[4] K. Chang, A. Babajimopoulos, G. A. Lavoie, Z. S. Filipi, D. N. Assanis, Analysis of Load and Speed Transitions in an HCCI Engine Using
1-D Cycle Simulation and Thermal Networks, SAE International, 2006.
[5] J. Bengtsson, P. Strandh, R. Johansson, P. Tunestal, B. Johansson, Model predictive Control of Homogeneous Charge Compression Ignition
(HCCI) engine dynamics, in: 2006 IEEE International Conference on Control Applications, 2006.
[6] Y. Wang, S. Makkapati, M. Jankovic, M. Zubeck, D. Lee, Control Oriented Model and Dynamometer Testing for a Single-Cylinder, Heated-
Air HCCI Engine, SAE International, 2009.
[7] Y. Urata, M. Awasaka, J. Takanashi, T. Kakinuma, T. Hakozaki, A. Umemoto, A study of gasoline-fuelled HCCI engine equipped with an
electromagnetic valve train, SAE International, 2004.
[8] R. Scaringe, C. Wildman, W. K. Cheng, On the high load limit of boosted gasoline HCCI engine operating in NVO mode, SAE Int. J. Engines
3 (2010) 35–45.
[9] C. Chiang, C. Chen, Constrained control of homogeneous charge compression ignition (HCCI) engines, in: 5th IEEE Conference on Industrial
Electronics and Applications (ICIEA), 2010.
[10] N. Ravi, M. J. Roelle, H. H. Liao, A. F. Jungkunz, C. F. Chang, S. Park, J. C. Gerdes, Model-based control of HCCI engines using exhaust
recompression, in: IEEE Transactions on Control Systems Technology, 2009.
[11] V. M. Janakiraman, X. Nguyen, D. Assanis, Nonlinear identification of a gasoline HCCI engine using neural networks coupled with principal
component analysis, Applied Soft Computing 13 (5) (2013) 2375 – 2389.
[12] V. M. Janakiraman, X. Nguyen, J. Sterniak, D. Assanis, A system identification framework for modeling complex combustion dynamics using
support vector machines, in: Informatics in Control, Automation and Robotics, Vol. 283 of Lecture Notes in Electrical Engineering, Springer
International Publishing, 2014, pp. 297–313.
[13] V. M. Janakiraman, X. Nguyen, J. Sterniak, D. Assanis, Identification of the Dynamic Operating Envelope of HCCI Engines Using Class
Imbalance Learning, Neural Networks and Learning Systems, IEEE Transactions on 26 (1) (2015) 98–112.
[14] G. T. Kalghatgi, R. A. Head, Combustion limits and efficiency in a homogeneous charge compression ignition engine, Int. J. Engine Res 7
(2006) 215–236.
[15] M. Shahbakhti, C. R. Koch, Characterizing the Cyclic Variability of Ignition Timing in a Homogenous Charge Compression Ignition Engine
Fueled With N-Heptane/Iso-Octane Blend Fuels, Int. J. Engine Res 9 (2008) 361–397.
[16] N. Liang, G. Huang, P. Saratchandran, N. Sundararajan, A fast and accurate online sequential learning algorithm for feedforward networks,
Neural Networks, IEEE Transactions on 17 (6) (2006) 1411–1423.
[17] Y. Jun, M. J. Er, An enhanced online sequential extreme learning machine algorithm, in: Control and Decision Conference, [Link] 2008.
Chinese, 2008, pp. 2902–2907.
[18] B. Mirza, Z. Lin, K.-A. Toh, Weighted online sequential extreme learning machine for class imbalance learning, Neural Processing Letters
38 (3) (2013) 465–486.
[19] M. T. Hoang, H. Huynh, N. Vo, Y. Won, A robust online sequential extreme learning machine, in: Advances in Neural Networks, Vol. 4491
of Lecture Notes in Computer Science, Springer Berlin / Heidelberg, 2007, pp. 1077–1086.
[20] V. M. Janakiraman, Machine learning for identification and optimal control of advanced automotive engines, Dissertation, University of
Michigan Ann Arbor, [Link]
[21] J. Kivinen, A. Smola, R. Williamson, Online learning with kernels, Signal Processing, IEEE Transactions on 52 (8) (2004) 2165–2176.

25
[22] A. Bordes, L. Bottou, The huller: a simple and efficient online svm, in: In Machine Learning: ECML 2005, Lecture Notes in Artificial
Intelligence, LNAI 3720, Springer Verlag, 2005, pp. 505–512.
[23] G. Zhao, Z. Shen, C. Miao, Z. Man, On improving the conditioning of extreme learning machine: A linear case, in: Information, Communi-
cations and Signal Processing, 2009. ICICS 2009. 7th International Conference on, 2009, pp. 1 –5.
[24] F. Han, H.-F. Yao, Q.-H. Ling, An improved extreme learning machine based on particle swarm optimization, in: Bio-Inspired Computing
and Applications, Lecture Notes in Computer Science.
[25] H. T. Huynh, Y. Won, Regularized online sequential learning algorithm for single-hidden layer feedforward neural networks, Pattern Recog-
nition Letters 32 (14) (2011) 1930 – 1935.
[26] V. Akpan, G. Hassapis, Adaptive predictive control using recurrent neural network identification, in: Control and Automation, 2009. MED
’09. 17th Mediterranean Conference on, 2009, pp. 61 –66.
[27] G.-B. Huang, Q.-Y. Zhu, C.-K. Siew, Extreme learning machine: Theory and applications, Neurocomputing 70 (2006) 489–501.
[28] G.-B. Huang, H. Zhou, X. Ding, R. Zhang, Extreme learning machine for regression and multiclass classification., IEEE Transactions on
Systems, Man, and Cybernetics, Part B 42 (2) (2012) 513–529.
[29] L. Bottou, Large-scale machine learning with stochastic gradient descent, in: Y. Lechevallier, G. Saporta (Eds.), Proceedings of COMP-
STAT’2010, Physica-Verlag HD, 2010, pp. 177–186.
[30] N. Le Roux, M. Schmidt, F. Bach, A Stochastic Gradient Method with an Exponential Convergence Rate for Strongly-Convex Optimization
with Finite Training Sets, Tech. Rep. arXiv:1202.6258v1, INRIA (2012).
[31] V. Vapnik, The Nature of Statistical Learning Theory, Springer, New York, 1995.
[32] S. Shalev-Shwartz, N. Srebro, SVM optimization: inverse dependence on training set size, in: Proceedings of the 25th international conference
on Machine learning, ICML ’08, ACM, New York, NY, USA, 2008, pp. 928–935.
[33] P. Ioannou, J. Sun, Robust adaptive control, PTR Prentice-Hall, 1996.
[34] J. Spooner, M. Maggiore, R. Ordóñez, K. Passino, Stable Adaptive Control and Estimation for Nonlinear Systems: Neural and Fuzzy
Approximator Techniques, Adaptive and Learning Systems for Signal Processing, Communications and Control Series, Wiley, 2004.
[35] F. Zhao, T. N. Asmus, D. N. Assanis, J. E. Dec, J. A. Eng, P. M. Najt, Homogeneous Charge Compression Ignition (HCCI) Engines, SAE
International, March 2003.
[36] G. A. Lavoie, J. Martz, M. Wooldridge, D. Assanis, A multi-mode combustion diagram for spark assisted compression ignition, Combustion
and Flame 157 (6) (2010) 1106 – 1110.
[37] O. Nelles, Nonlinear System Identification: From Classical Approaches to Neural Networks and Fuzzy Models, Springer, 2001.
[38] K. S. Narendra, K. Parthasarathy, Identification and control of dynamical systems using neural networks 1 (1) (1990) 4–27.
[39] L. Re, F. Allgöwer, L. Glielmo, C. Guardiola, I. Kolmanovsky, Automotive Model Predictive Control: Models, Methods and Applications,
Lecture Notes in Control and Information Sciences, Springer, 2010.
[40] J.-B. Yang, C.-J. Ong, Feature selection using probabilistic prediction of support vector regression, Neural Networks, IEEE Transactions on
22 (6) (2011) 954–962.
[41] X. Liu, L. Wang, G.-B. Huang, J. Zhang, J. Yin, Multiple kernel extreme learning machine, Neurocomputing 149, Part A (0) (2015) 253 –
264.
[42] K.-A. Toh, Deterministic neural classification, Neural Comput. 20 (6) (2008) 1565–1595.

26
Author Biography

Vijay Manikandan Janakiraman received the bachelor’s degree in Mechanical Engineering

(2007) from Sri Venkateswara College of Engineering, Chennai, India. He received Master’s degrees in Mechanical
Engineering (2008) and Electrical Engineering - Systems (2013) and the Ph.D. degree in Mechanical Engineering
(2013) from the University of Michigan at Ann Arbor, MI, USA. Since August 2013, he has been working as a
research scientist with the Data Sciences Group, Intelligent Systems Division at the NASA Ames Research Center,
Moffett Field, CA, USA. His current research interests include machine learning, data mining in high dimensional
time series and decision making under uncertainty.

XuanLong Nguyen received the Ph.D. degree in computer science and the M.S. degree in
statistics, both from the University of California, Berkeley. He is currently an Assistant Professor of Statistics at the
University of Michigan. His research interests lie in distributed and variational inference, nonparametric Bayesian
statistics, and applications to detection/estimation problems in distributed and adaptive systems. Dr. Nguyen is a
recipient of the CAREER award from the NSF Division of Mathematical Sciences, the Leon O. Chua Award from the
UC Berkeley, the IEEE Signal Processing Societys Young Author Best Paper award, and an Outstanding Paper award
from the International Conference on Machine Learning.

Dennis Assanis received the Ph.D. degree in Power and Propulsion and the M.S. degrees in
Naval Architecture and Marine Engineering and Mechanical Engineering from the Massachusetts Institute of Tech-
nology. Dr. Assanis is a Professor in the Department of Mechanical Engineering and is also the Provost, Senior Vice
President for Academic Affairs, and Vice President for Brookhaven Affairs at the Stonybrook University, NY, USA.

27
Assanis served as the Jon R. and Beverly S. Holt Professor of Engineering and Arthur F. Thurnau Professor at the
University of Michigan, as well as Director of the Michigan Memorial Phoenix Energy Institute, Founding Director
of the US-China Clean Energy Research Center for Clean Vehicles and Director of the Walter E. Lay Automotive
Laboratory. Dr. Assanis research interests lie in the thermal sciences and their applications to energy conversion,
power and propulsion, and automotive systems design. His research focuses on analytical and experimental studies
of the thermal, fluid and chemical phenomena that occur in internal combustion engines, after-treatment systems, and
fuel processors. His efforts to gain new understanding of the basic energy conversion processes have made significant
impact in the development of energy and power systems with significantly improved fuel economy and dramatically
reduced emissions. His groups research accomplishments have been published in over 250 articles in journals and
international conference proceedings. Dr. Assanis is a Member of the National Academy of Engineering and is an
ASME and SAE Fellow.

28