This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3000152, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2020.Doi Number

A Study on Extreme Learning Machine for

Gasoline Engine Torque Prediction
Weiying Zeng1, (Student Member, IEEE), Mohammed A.S. Khalid1, (Senior Member, IEEE),
Xiaoye Han2 and Jimi Tjong2
1
Department of Electrical and Computer Engineering, University of Windsor, Windsor, ON, N9B 3P4, Canada
2
Department of Mechanical, Automotive and Materials Engineering, University of Windsor, Windsor, ON, N9B 3P4, Canada
Corresponding author: Xiaoye Han (e-mail: hanxy02@hotmail.com)

ABSTRACT This research presents an extreme learning machine (ELM) based neural network modeling
technique for gasoline engine torque prediction. The technique adopts a single-hidden layer feedforward
neural network (SLFN) structure which has the potential to approximate any continuous function with high
accuracy. To verify the robustness of this technique, over 3300 data points collected from a real-world
gasoline engine are used to train, validate, and test the model. These data points cover a wide spectrum of
normal engine operating conditions, with the engine speed from 1000 rpm to 4500 rpm, and the engine torque
from idle to full load. The experiment results demonstrate that the model can predict the gasoline engine
torque with high accuracy. Moreover, this research proposes a weight factor approach to further improve the
prediction accuracy of the model in the desired data regions without modifying the input data set. The
evaluation shows that the weight factor approach can reduce the overall prediction errors in the regions
significantly. This feature is particularly useful in tuning the performance of the model when the significance
of the individual data points varies, or when the distribution of the data points is imbalanced. In practice, the
modeling approaches presented in this research will help reduce the engine test and verification time.

INDEX TERMS Artificial neural network, Extreme learning machine, Gasoline engine model, Torque
prediction, Weight factor

I. INTRODUCTION tuning the Wiebe variables. Togun et al. have proposed a

Creating accurate gasoline engine torque models is a
challenging task. Firstly, the gasoline engine is a complex
system, whose operation often involves multidisciplinary
interactions and contains a lot of transient states. Thus, it could
be very difficult to create an accurate model purely based on
the physical/mathematical principles. Secondly, many
modeling approaches only focus on one engine combustion
cycle and assume the boundary conditions to be ideal. As a
result, the model often fails to deliver consistent results when
being applied to a more realistic engine operating
environment. Thirdly, as more and more sophisticated
components and control logics have been introduced to
modern engines, such a task has become increasingly difficult
and time-consuming [1], [2].
In the past, various engine or engine subsystem models have
been proposed by the researchers. For instance, Tolou et al.
have presented a semi-predictive model of a homogeneous
turbocharged gasoline direct injection engine [3]. The model
approximates the combustion heat release with a double-
Wiebe function and predicts the cylinder peak pressure by
nonlinear mathematical approach to model the engine torque
using a recursive least square method, in which the
nonlinearity is determined with Hammerstein structure and the
system is identified by studying multi-order linear dynamics
[4]. Tan and Reitz have presented a spark model with an
equation to calculate the propagation rate of the ignition flame
[5]. The formation and properties of the flame kernel are also
modeled in [6] by Boudier et al. Their model results match a
set of typical engine experiment data without parameter
adjustments. To investigate the cycle-to-cycle performance
variance of gasoline engines, Daw et al. have proposed a
model that shows the autocorrelation between the stochastic
fluctuation of the engine parameters and the nonlinear
deterministic coupling between successive engine cycles [7].
In addition, some computational simulation approaches are
also adopted to model the engine operation. For instance,
computational fluid dynamics (CFD) and MATLAB/Simulink
modeling technologies are demonstrated [8]–[11]. However,
to use the previously mentioned methods properly, a
comprehensive understanding of the internal combustion

VOLUME XX, 2020 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3000152, IEEE Access
engine or its subsystems is required, which has limited the
non-expert’s accessibility to such approaches. Moreover, as
pointed out in [12], such modeling strategies often fail to work
desirably in practice due to the complexity of multivariate
coupling and unexpected noise.
In contrast to the above-mentioned modeling approaches,
the data-driven modeling techniques can work well even
without a deep understanding of the fundamental physics of
the target system. Generally, these techniques can create
models directly from the data collected on the target system
instead of from the complicated physical/mathematical
equations. Moreover, the data-driven modeling techniques
could function well when multivariate correlation exists.
Among many data-driven techniques for the engine-related
applications, artificial neural network (ANN) is very popular
and has been increasingly studied. For instance, Wu et al. have
proposed a mass air flow (MAF) sensor model on a dual-cam
engine using a two hidden layer ANN structure [2]. The
simulated results have shown good agreement between the
dynamometer data and the vehicle test data. Togun and
Baysec have demonstrated the capability of predicting engine
torque and brake fuel specific consumption using neural
networks [13]. Cay has investigated the applicability of ANN
to predict engine performance and exhaust temperature values
[14]. Cycle-to-cycle combustion variation is studied by Di
Mauro et al. in [15], which helps identify the pre-ignition and
pre-combustion factors, as well as predict the variation of the
indicated mean effective pressure. A neural network based
nonlinear predictive control scheme has been investigated by
Hu et al., in which the coordinated control of throttle and
wastegate on a turbocharged gasoline engine is explored [16].
Li et al. have presented a model for nitrogen oxide and smoke
emissions on a diesel engine [17]. In their research, the authors
have not only discussed how to select the model inputs based
on the physical analysis, but also showed the insightful
comparisons of a range of neural network architectures in
terms of model complexity and accuracy. In addition to
characterizing engine performance and predicting engine
behaviors, neural network models have also been used in
engine fault diagnostics. Zheng et al. have revealed how
misfire conditions can be detected and categorized under
different engine conditions using Elman neural network [18].
Bearing knock fault features can be detected by proper ANNs
as demonstrated by Chen and Randall [19]. Vibration data are
investigated by Ahmed et al. with an ANN model that can
identify various engine faults and categorize the severity of the
faults [20]. Wen et al. have proposed a bearing fault detection
model using convolutional neural network, which can
autonomously learn the unique features of each type of bearing
fault and identify the defective bearings with its fault type [21].
However, the neural network approaches in the
aforementioned research papers still have some room for
improvement. Firstly, the neural networks all use a
backpropagation algorithm to train the weight of each neuron.
In practice, this algorithm may not be easy enough to
VOLUME XX, 2020
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
implement, especially for the non-experts, due to the
disadvantages such as difficulty in determining proper
learning rate, getting trapped in local minima, prone to over-
training, and very time-consuming for most of the applications
[22]. Secondly, these neural network models are typically
created and tested with about 80 to 130 data points. Thus, there
is still room to explore the robustness of a neural network if
the number and complexity of the data points have increased
significantly. Thirdly, as all the data points are treated equally
in the neural network optimization process, some practical
engineering considerations can be further investigated. For
instance, as the gasoline engines usually operate at certain
conditions much more often than other conditions, a neural
network model is desirable to have higher accuracy at the
more frequently operated engine conditions, even though the
accuracy at the less frequently operated engine conditions has
to be sacrificed slightly.
Among many ANN training algorithms, neural network
with random weights (NNRW) is appealing due to its
advantages in network simplicity, faster training speed, and
competitive accuracy [23]. Compared to the NNRWs, such as
Random Vector Functional Link (RVFL) [24] and standard
feed-forward neural network with random weights (the
Schmidt’s method) [25], ELM outshines in the aspects like
setting the bias of the output node to zero, transforming
different hidden nodes to one unified form and generating
random node parameters prior to knowing the training data
[26]. In addition, ELM can be extended to multiple hidden
layer architecture and has a strong potential for big data
analytics [23], [27].
In this paper, we present an approach to create a single-
hidden layer feedforward neural network based regression
model using ELM that can predict the output torque of a
gasoline engine. The contributions of this paper are threefold:
1) It establishes a procedure to create an engine torque
prediction model using ELM approach so that a non
engine expert can also benefit from it.
2) It trains and evaluates the neural network with a large
amount of experimental data, which not only
demonstrates the capability of the modeling approach, but
also validates its ability to generalize.
3) It proposes a weight factor approach to further enhance
the model accuracy in the desired data regions.
The rest of the paper is organized as follows. Section II
introduces the background of the linear regression problem
using an SLFN and presents an ELM approach to train the
SLFN. It also proposes a weight optimizing algorithm to
enhance the regression accuracy in the selected areas. Section
III discusses the experimental setup and data acquisition
approach. Section IV presents the model regression and
accuracy enhancement results. Section V concludes this paper.
2
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3000152, IEEE Access

II. MODELING METHODOLOGY where

𝑔(𝑤1 ∙ 𝑥1 + 𝑏1 ) ⋯ 𝑔(𝑤𝑁̃ ∙ 𝑥1 + 𝑏𝑁̃ )
A. BACKGROUND OF SINGLE-HIDDEN LAYER
FEEDFORWARD NEURAL NETWORK 𝐻=[ ⋮ ⋮ ⋮ ] (3)
APPROXIMATION 𝑔(𝑤1 ∙ 𝑥𝑁 + 𝑏1 ) ⋯ 𝑔(𝑤𝑁̃ ∙ 𝑥𝑁 + 𝑏𝑁̃ ) 𝑁×𝑁̃
In this research, we explore the approach to predicting engine
𝛽1𝑇 𝑦1𝑇
torque with a general SLFN approximation problem because
𝛽=[ ⋮ ] and 𝑌=[ ⋮ ] (4)
of two main reasons. Firstly, the SLFN has a very simple
𝛽𝑁𝑇̃ ̃ 𝑦𝑁𝑇 𝑁×𝑘
structure, which makes it much easier to deal with than many 𝑁×𝑘
other complicated ANN structures. Secondly, an SLFN with Let 𝛽̂ be the estimation of 𝛽, 𝑤
̂ be the estimation of 𝑤, and
at most 𝑁 ̃ hidden nodes and with almost any nonlinear ̂
𝑌 be the estimated output of 𝑌 . If the number of neurons
activation function can learn 𝑁 distinct observations with zero equals the number of samples (𝐻 is a full rank square matrix),
error, which implies that it has the potential for accurate −1
estimation [22], [28], [29]. The description of a generic SLFN there exists a perfect estimation of 𝛽̂ = 𝐻(𝑤̂, 𝑥, 𝑏̂) ∗ 𝑌 that
approximation problem is as follows. makes ‖𝑌̂ − 𝑌‖ = 0 . However, in practice, the number of
For 𝑁 arbitrary distinct samples (𝑥𝑖 , 𝑦𝑖 ), where 𝑥𝑖 = neurons is often much less than the number of samples (𝑁 ̃≪
[𝑥𝑖1 , 𝑥𝑖2 , … , 𝑥𝑖𝑚 ]𝑇 ∈ 𝑅𝑚 and 𝑦 𝑖 = [𝑦𝑖1 , 𝑦𝑖2 , … , 𝑦𝑖𝑘 ]𝑇 ∈ 𝑁). As a result, there is no perfect estimation of 𝛽̂ that makes
𝑅𝑘 , consider the standard model for SLFN approximation ‖𝑌̂ − 𝑌‖ = 0 . Therefore, the target of the SLFN
with 𝑁 ̃ nodes, approximation turns into finding out the appropriate
̃
𝑁 ̃
𝑁 estimations of 𝛽̂, 𝑤̂ and 𝑏̂, which can minimize the cost of the
𝑦𝑗 = ∑ 𝛽𝑖 𝑔𝑖 (𝑥𝑗 ) = ∑ 𝛽𝑖 𝑔𝑖 (𝑤𝑖 ∙ 𝑥𝑗 + 𝑏𝑖 ), estimation 𝐸, where
𝑖=1 𝑖=1
̂𝑖 , 𝑏̂𝑖 )𝛽̂𝑖 − 𝑌‖
𝐸 = ‖𝑌̂ − 𝑌‖ = 𝑚𝑖𝑛 ‖𝐻(𝑤 (5)
𝑗 = 1,2, … , 𝑁, (1) 𝑤𝑖 ,𝑏𝑖 ,𝛽𝑖

where 𝑚 is the dimension of the input data, 𝑘 is the
dimension of the output data, 𝑤𝑖 = [𝑤𝑖1 , 𝑤𝑖2 , … , 𝑤𝑖𝑚 ] is the
weight vector connecting the 𝑖𝑡ℎ hidden node and the input
neurons, 𝛽𝑖 = [𝛽𝑖1 , 𝛽𝑖2 , … , 𝛽𝑖𝑘 ]𝑇 is the weight vector
connecting the 𝑖𝑡ℎ hidden node and the output neurons, and 𝑏𝑖
is the bias of the 𝑖th hidden neuron. A structure of a general
SLFN is illustrated in Fig. 1. The term 𝑤 ∙ 𝑥 is the inner
product of 𝑤𝑖 and 𝑥𝑗 , and 𝑔 is the activation function of the 𝑖𝑡ℎ
hidden neuron. In practice, 𝑔 can be any nonlinear continuous
function, such as sine, sigmoid, hyperbolic tangent, or radial
basis function. The term 𝑔𝑖 (∙) represents the data feature
extracted by the 𝑖𝑡ℎ neuron. The compact version of (1) can be
written as:
𝑌 = 𝐻𝛽
B. EXTREME LEARNING MACHINE APPROXIMATION
FOR SLFN
Since the gradient-based approximation method, such as
backpropagation, has several prominent disadvantages as
pointed out in [22], a different algorithm is preferred in this
research. It is revealed that ELM has superior training speed
and comparable accuracy over many other popular SLFN
training algorithms, such as radial basis function (RBF),
support vector machine (SVM) and AdaBoost [30]. However,
its application on gasoline engine torque prediction has yet to
be explored. Therefore, in this paper, an ELM based SLFN
gasoline engine torque model is explored and assessed.
With prefixed 𝑤 ̂ and 𝑏̂ (which are usually randomly
(2) generated numbers between -1 and 1, the approximation
problem stated in (5) can be simplified as:

Hidden Layer
Input Layer
h1
Output Layer
x1
h2 y1
(𝑤11 , 𝑏11 ), (𝑤21 , 𝑏21 ), …, (𝑤𝑁̃ 1, 𝑏𝑁̃ 1)
(𝛽11 , …, 𝛽1𝑘 )
x2 (𝑤12 , 𝑏12 ), (𝑤22 , 𝑏22 ), …, (𝑤𝑁̃ 2, 𝑏𝑁̃ 2) ……
…… h3 (𝛽𝑁̃ 1, …, 𝛽𝑁̃ 𝑘)
(𝑤1𝑚 , 𝑏1𝑚 ), (𝑤2𝑚 , 𝑏2𝑚 ), …, (𝑤𝑁̃ 𝑚, 𝑏𝑁̃ 𝑚)
yk
xm
̃
𝑁

FIGURE 1. Basic structure of an SLFN

VOLUME XX, 2020 3

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3000152, IEEE Access

𝐸 = ‖𝑌̂ − 𝑌‖ = 𝑚𝑖𝑛‖𝐻𝛽̂𝑖 − 𝑌‖ (6) C. DETERMINE THE K VALUE FOR REGRESSION

𝛽𝑖 PROBLEM
The minimal norm least square estimation of 𝛽 is as the As pointed out in the previous section, the constant 𝑘
following [22]: mentioned in (9) needs to be determined for better regression
performance. It is proven that the best 𝑘 is a positive value
𝛽̂ = 𝐻 † 𝑌 (7) very close to zero, but it is impossible to calculate out
†
where 𝐻 is the Moore-Penrose inverse of matrix 𝐻. The mathematically, as it ultimately depends on the unknown
unbiased least square estimation of the Moore-Penrose inverse parameter being estimated [31], [33]. To discover a closer
is: estimation of the true 𝑘 , several approaches have been
proposed, such as ridge trace method [31], Akaike’s
(𝐻 𝑇 𝐻)−1 𝐻 𝑇 , if H T H is nonsingular information criterion (AIC) [34], and Adjusted 𝑅2 [35], [36].
𝐻† = { 𝑇 (8)
𝐻 (𝐻𝐻 𝑇 )−1 , if HHT is nonsingular However, each of these approaches alone has some unique
disadvantages. For instance, ridge trace may be too subjective
However, according to the ridge regression theory [31], the and lack generalization for the new data set. AIC may be good
variance of the above unbiased least square estimation may be in selecting the best model among all candidates, but it
large and far from the true value when multicollinearity exists. requires that the best model is well established among the
In addition, when 𝐻𝐻𝑇 ( or 𝐻 𝑇 𝐻 ) is not invertible, there candidate models, which is often impractical [37]. Adjusted
would be no unique solution of 𝛽̂ . In order to improve the 𝑅2 is useful for evaluating parameters, but it is incapable of
stability of the estimation, a small positive value 𝑘 is added to model selection.
the diagonal elements of the correlation matrix. Thus, we have In this research, a ten-fold cross-validation approach
the ridge regression estimation of 𝛽 as: incorporated with ridge regression is used to determine a
̂∗ = H † Y
𝛽 proper 𝑘 value that could minimize the estimation error. This
method has two main benefits: it delivers a deterministic
(𝑘𝐼 + 𝐻 𝑇 𝐻)−1 𝐻𝑇 𝑌, if H T H is nonsingular result, and the result is statistically representative of the
={ 𝑇 (9)
𝐻 (𝑘𝐼 + 𝐻𝐻 𝑇 )−1 𝑌, if HHT is nonsingular targeted data set. The key steps of the approach are as follows:
As a result, the estimated output of the network becomes: 1) Randomly divide the training data set into ten equal (or
almost equal) partitions, as 𝑋 = (𝑋1 , 𝑋2 , … , 𝑋10 ).
̂∗
𝑌̂ = 𝐻𝛽 (10) 2) For each 𝑛 = 1,2,3, … ,10 , set the 𝑛𝑡ℎ partition as the
The variance of the ridge regression estimator is: validation data set and take all the rest data sets as the
training data set.
σ2 Z(H T H)−1 ZT , if H T H is nonsingular 3) For each potential 𝑘 value, compute 𝛽 ̂∗ using the training
Var(β̂∗ ) = { 2 (11)
𝜎 𝑍(𝐻𝐻 𝑇 )−1 𝑍 𝑇 , if HHT is nonsingular data set.
4) Calculate the distance between the estimated 𝑦̂ and the
where expected 𝑦 using the validation data set.
(kI + H T H)−1 , if H T H is nonsingular 𝐸(𝑦̂ − 𝑦)𝑘 = 𝑠𝑞𝑟𝑡(∑(𝑦̂ − 𝑦)2𝑘 /𝑚) (14)
Z={ (12)
(kI + HH T )−1 , if HHT is nonsingular
where 𝑚 is the size of the validation data set.
The squared bias of the ridge regression estimator is 5) Repeat steps 3) and 4) for all the data partitions and
̂∗ ) = 𝑘 2 𝛽𝑇 𝑍 −1 𝛽
𝐵𝑖𝑎𝑠 2 (𝛽 (13) compute the cross-validation distance for the overall data
set.
From the above equations, it is demonstrated that the total (10)
̂∗ is a monotonic decreasing function with respect
variance of 𝛽 𝐶𝑉_𝐸(𝑦̂ − 𝑦)𝑘 = ∑10
𝑛 𝐸(𝑦
̂ − 𝑦)𝑘 (15)
to 𝑘 , while the total squared bias of 𝛽 ̂∗ is a monotonic
6) Choose the 𝑘 value which has the minimum overall cross-
increasing function with respect to 𝑘 . According to the validation distance.
existence theorem, these properties lead to a conclusion that
(10)
there always exists such a 𝑘 that makes the mean square error 𝑘 = 𝑎𝑟𝑔𝑚𝑖𝑛 𝐶𝑉_𝐸(𝑦̂ − 𝑦)𝑘 (16)
𝑘
(MSE) of 𝛽 ̂∗ smaller than the MSE of 𝛽̂ . Therefore, 𝛽 ̂∗ is a
closer estimation of the true model parameter 𝛽 . Since it D. WEIGHTED ELM REGRESSION FOR ENGINE MODEL
almost only requires a matrix inversion operation to determine Though the ELM method described in the above sections can
the unsolved model parameters, the training speed of ELM is optimize ordinary linear regression by reducing the overall
significantly faster than the traditional gradient-based distance between the estimated and expected values, it may
approaches [22], [32]. not be well suited for the applications where higher accuracy
is preferred at certain points. For instance, as pointed out in
[38], a typical light-duty vehicle engine actually spends most
of the time at the low to medium speed operating (LMSO)

VOLUME XX, 2020 4

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3000152, IEEE Access

Test Cell Control Room

Cooling Tower
Environment
Control System
P
Emission Sampling Dyno Control System

Intake and Exhaust

Management

Engine
AC Dyno

FIGURE 2. Test system diagram

conditions in the standard United States Environmental

Protection Agency (EPA) vehicle and fuel emission tests, such
as Urban Dynamometer Driving Schedule (UDDS) and
Highway Fuel Economy Test (HWFET). Therefore, it is
desirable that an engine regression model could have higher
accuracy at these operating points, even though the accuracy
at other operating points may have to be compromised
slightly. In the past, a few researchers have proposed a weight
matrix on ELM to solve imbalanced sampling and multiclass
classification problems [30], [39], [40]. However, to the best
of the authors’ knowledge, there is no exploration of using
weight regulated ELM to enhance a regression model only at
certain data areas without changing the input data sets. Hence,
this paper proposes the approach of implementing weight
regulated ELM on an SLFN linear regression model to serve
such a purpose.
By introducing an 𝑁x𝑁 diagonal matrix 𝑊, the weight of
each data point is adjusted by a factor 𝑤𝑖 , where 𝑖 is the
position of the input data point. According to the study by
Zong et al., the objective of the weighted function becomes
[39]:
1 1
Minimize: 𝐿𝑃𝐸𝐿𝑀 = ‖𝛽‖2 + 𝑘𝑊 ∑𝑁
𝑖=1‖𝜉𝑖 ‖
2
2 2
h(xi )β = yi − ξi , i = 1, … , N
Subject to: {
𝜉𝑖 = 𝑦𝑖 − 𝑦̂𝑖
According to Karush Kuhn Tucker (KKT) theorem [41], the
solution to (17) is:
𝑇 𝑇 −1
̂∗ = {𝐻 (𝑘𝐼 + 𝑊𝐻𝐻 ) 𝑊𝑌, when N is small
𝛽 (19)
(𝑘𝐼 + 𝐻 𝑇 𝑊𝐻)−1 𝐻𝑇 𝑊𝑌, when N is large
Once 𝛽̂∗ is determined, the model can be considered as
trained and ready for testing. Then the predicted data are
calculated according to (10).
III. EXPERIMENT SETUP AND DATA ACQUISITION
A. TEST CELL SETUP
The model is validated on an 8-cylinder, 4-stroke, dual-equal
variable camshaft timing (VCT) gasoline engine. The
experiment is performed at an industry-leading engine dyno
testing facility. The diagram of the key test cell setup is shown
in Fig. 2. A dyno control system controls the test cell operation
and records the test data. The test data from both the test cell
and the engine control unit (ECU) are recorded. The
experiment engine is connected to an alternative current
dynamometer, which controls the engine operating speed. A
heat exchange cooling tower is used to maintain the engine
coolant out temperature and pressure at 90.5±2.8 degC and
145±3.4 kPa, respectively. Engine oil sump temperature is
capped at 90.5±2.8 degC during the whole test. The
environment control system regulates the ambient temperature
and circulates air in the test cell. The combustion air is
supplied by the intake and exhaust management system with
the temperature maintained at 23.8±2.8 degC and the humidity
at 7.9±0.7 g_H2O/kg_Air. The emission gases are sampled by
(17) the emission bench for composition analysis. The ECU is
equipped with Vision 5.0 software, through which the ECU
(18) parameters are monitored and controlled in real-time. An
Ethernet-based communication protocol is used to exchange
information between the dyno controller and Vison 5.0
software, which updates the ECU parameters to the dyno
control system at a frequency of 20 Hz.
B. TEST POINTS AND DATA ACQUISITION
As the number of control variables increases, the size of the
engine map rises exponentially. Consequently, the complexity
and the number of points to characterize the engine also
increase, as shown in Fig. 3. In order to build a neural network
model that can represent the engine at a wide range of
operating conditions, and to test the generalization of the

VOLUME XX, 2020 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3000152, IEEE Access

Engine Speed (RPM)

500 1000 1500 2000 2500 3000 3500 4000 4500
0 23.9 26.3 29.4 37.6 39.5 31.8 33.4 34.5 31.4
Engine Speed (RPM)
10 27.7 30.4 33.8 42.9 45.0 36.5 38.3 39.5 36.0

Pedal Position (%)

500
20 1000 32.0
29.2 1500 35.6
2000 27.9
2500 39.03000 33.9
3500 26.6
4000 32.6
4500 29.4
0 23.9
30 26.3
28.8 29.4
31.6 37.6
35.2 39.5
25.8 31.8
36.4 33.4
31.7 34.5
18.1 31.4
25.5 22.7
10 27.7 30.4 Engine
33.8 Speed
42.9 (RPM)
45.0 36.5 38.3 39.5 36.0
40 27.9 30.7 34.1 25.5 35.4 30.8 16.5 18.4 16.0

Pedal Position (%)

20500 501000
29.2 1500
32.0
27.0 2000
35.6
29.7 2500
27.9
33.0 3000
39.0
25.3 3500
33.9
34.6 4000
26.6
29.2 4500
32.6
18.3 29.4
12.6 10.5
Engine Speed (RPM) 0 23.9
30 26.3
28.8
60
29.4 35.2
31.6
26.1
37.6 25.8
28.7
39.5 36.4
31.9
31.8 31.7
25.1
33.4 18.1
33.9
34.5 25.5
27.3
31.4 22.7
18.1 12.7 10.6
10 27.7
40 30.4
27.9 33.8
30.7 27.7 42.9
34.1 30.8 45.0
25.5 24.8 36.5
35.4 33.1 38.3
30.8 25.4 39.5
16.5 18.0 36.0
18.4 13.5
16.0
500 1000 1500 2000 2500 3000 3500 4000 4500 70 25.2 11.3

Pedal Position (%)

20 29.2
50 32.0
27.0 35.6 33.0
29.7 27.9 25.3
39.0 34.6
33.9 29.226.6 18.3
32.6 12.6
29.4 10.5
80 24.3 26.7 29.7 24.6 32.4 23.5 17.8 15.2 12.9
0 23.9 26.3 29.4 37.6 39.5 31.8 33.4 34.5 31.4 30 28.8
60 31.6
26.1 35.2 31.9
28.7 25.8 25.1
36.4 33.9
31.7 27.318.1 18.1
25.5 12.7
22.7 10.6
90 23.4 25.7 28.6 24.3 31.7 21.6 17.7 13.5 11.3
10 27.7 30.4 33.8 42.9 45.0 36.5 38.3 39.5 36.0 40 27.9
70 30.7
25.2 34.1 30.8
27.7 25.5 24.8
35.4 33.1
30.8 25.416.5 18.0
18.4 13.5
16.0 11.3
100 22.5 24.7 27.5 24.1 30.9 19.7 17.6 12.9 10.8
Pedal Position (%)

50 27.0
80 29.7
24.3 33.0 29.7
26.7 25.3 24.6
34.6 32.4
29.2 23.518.3 17.8
12.6 15.2
10.5 12.9
20 29.2 32.0
30 28.8 31.6
35.6
Map
MapSize
Size
35.2
27.9
25.8
39.0
36.4
33.9
31.7
26.6
18.1
32.6
25.5
29.4
22.7
60
70
26.1
90
25.2
100
28.7
23.4
27.7
22.5
31.9 28.6
25.7 25.1 24.3
30.8 27.5
24.7
33.9 31.7
24.8 24.1
27.3 21.6
33.1 30.9
25.4 19.7
18.1 17.7
12.7 13.5
18.0 17.6
10.6 11.3
13.5 12.9
11.3 10.8
40 27.9 30.7 34.1 25.5 35.4 30.8 16.5 18.4 16.0 80 24.3 26.7 29.7 24.6 32.4 23.5 17.8 15.2 12.9
50
60
27.0
26.1
29.7
28.7
m
m xxnn
33.0
31.9
25.3
25.1
34.6
33.9
29.2
27.3
18.3
18.1
12.6
12.7
10.5
10.6
90
100
23.4
22.5
25.7
24.7
28.6
27.5
24.3
24.1
31.7
30.9
21.6
19.7
17.7
17.6
13.5
12.9
11.3
10.8

70 25.2 27.7 30.8 24.8 33.1 25.4 18.0 13.5 11.3

80 24.3 26.7 29.7 24.6 32.4 23.5 17.8 15.2 12.9
Engine Speed (RPM)
90 23.4 25.7 28.6 24.3 31.7 21.6 17.7 13.5 11.3
500 1000 1500 2000 2500 3000 3500 4000 4500
100 22.5 24.7 27.5 24.1 30.9 19.7 17.6 12.9 10.8 0 23.9 26.3 29.4 37.6 39.5 31.8 33.4 34.5 31.4

Pedal Position (%)

Engine Speed (RPM)

ol
10 27.7 30.4 33.8 42.9 45.0 36.5 38.3 39.5 36.0

Pedal Position (%)

500 1000 32.0
1500 35.6
2000 27.9
2500 39.03000 33.9
3500 26.6
4000 32.6
4500

ntr
20 29.2 29.4
0 23.9
30 26.3
28.8 29.4
31.6 37.6
35.2 39.5
25.8 31.8
36.4 33.4
31.7 34.5
18.1 31.4
25.5 22.7
Engine Speed (RPM)
VCT Control o d
10 27.7
40 30.4 30.7
27.9 33.8 34.1
42.9 25.5
45.0 35.436.5 30.8
38.3 16.5
39.5 18.4
36.0 16.0

R C uce

Pedal Position (%)

20500 501000
29.2 1500 29.7
32.0
27.0 2000 33.0
35.6 2500 25.3
27.9 3000 34.6
39.0 3500 29.2
33.9 4000 18.3
26.6 4500 12.6
32.6 29.4 10.5
0 23.9
30 26.3
28.8 29.4 35.2
31.6 37.6 25.8
39.5 36.4
31.8 31.733.4 18.1
34.5 25.5
31.4 22.7
Introduced EG trod 10 27.7
40
60
30.4
27.9
26.1 28.7
33.8 34.1
30.7
31.9
42.9 25.5
25.1
45.0 35.4
33.9
36.5 30.8
27.3
38.3 16.5
18.1
39.5 18.4
12.7
36.0 16.0
10.6
70
Map Size
25.2 27.7 30.8 24.8 33.1 25.4 18.0 13.5 11.3

Pedal Position (%)

20 29.2
50 32.0
27.0 35.6 33.0
29.7 27.9 25.3
39.0 34.6
33.9 29.226.6 18.3
32.6 12.6
29.4 10.5
80 24.3 26.7 29.7 24.6 32.4 23.5 17.8 15.2 12.9
In 30
40
28.8
60
27.9
70
31.6
26.1
90
30.7
25.2
100
35.2 31.9
28.7
23.4
25.8 25.1
25.7
34.1 30.8
27.7
22.5
36.4 33.9
28.6
25.5 24.8
24.7
31.7 27.3
24.3
35.4 33.1
27.5
30.8 25.4
24.1
18.1 18.1
31.7
25.5 12.7
21.6
16.5 18.0
30.9
22.7 10.6
17.7
18.4 13.5
19.7
13.5
16.0 11.3
17.6 12.9
11.3
10.8

Engine Speed (RPM)

50
60
70
27.0
80
26.1
90
25.2
100
m x n x k xh
29.7
24.3
28.7
23.4
27.7
22.5
33.0 29.7
26.7
25.7
25.3 24.6
31.9 28.6
30.8 27.5
24.7
34.6 32.4
25.1 24.3
24.8 24.1
29.2 23.5
33.9 31.7
27.3 21.6
33.1 30.9
25.4 19.7
18.3 17.8
12.6 15.2
18.1 17.7
10.5 12.9
12.7 13.5
18.0 17.6
10.6 11.3
13.5 12.9
11.3 10.8 Speed (RPM)

°)
500 1000 1500 2000 2500 3000 3500 4000 4500
80 24.3 26.7 29.7 24.6 32.4 23.5 17.8 15.2 12.9

T(
0 23.9 26.3 29.4 37.6 39.5 31.8 33.4 34.5 31.4
90 23.4 25.7 28.6 24.3 31.7 21.6 17.7 13.5 11.3
30.4 Engine Speed
42.9 (RPM)
Pedal Position (%)

VC
10 27.7 33.8 45.0 36.5 38.3 39.5 36.0 100 22.5 24.7 27.5 24.1 30.9 19.7 17.6 12.9 10.8
Pedal Position (%)

500
20 1000 32.0
29.2 1500 35.6
2000 27.9
2500 39.0
3000 3500
33.9 4000
26.6 4500 29.4
32.6
0 23.9
30 26.3
28.8 29.4
31.6 37.6 25.8
35.2 39.5 36.4
31.8 33.4
31.7 34.5
18.1 31.4 22.7
25.5
10 27.7
40 30.4
27.9
Engine
33.8
30.7
Speed
34.1
(RPM)
42.9 25.5
45.0 35.4
36.5 38.3
30.8 39.5
16.5 36.0 16.0
18.4 Engine Speed (RPM)
Pedal Position (%)

500
20 1000
29.2
50 1500 29.7
32.0
27.0 2000 33.0
35.6 2500 25.3
27.9 3000 34.6
39.0 3500
33.9 4000
26.6
29.2 4500
32.6
18.3 29.4 10.5
12.6
500 1000 1500 2000 2500 3000 3500 4000 4500
0 23.9
30 26.3
28.8
60 29.4 28.7
31.6
26.1 37.6 31.9
35.2 39.5 25.1
25.8 31.8 33.9
36.4 33.4
31.7 34.5
18.1
27.3 31.4
25.5
18.1 22.7 10.6
12.7
0 23.9 26.3 29.4 37.6 39.5 31.8 33.4 34.5 31.4
10 27.7
40 30.4
70
27.9 33.8 27.7
25.2
30.7 42.9 30.8
34.1 45.0 24.8
25.5 36.5 33.1
35.4 38.3
30.8 39.5
25.4
16.5 36.0
18.0
18.4 13.5
16.0 11.3 Engine Speed (RPM)
Map Size 10 27.7 30.4 33.8 42.9 45.0 36.5 38.3 39.5 36.0
Pedal Position (%)

20 29.2 32.0
80 35.6 26.7
24.3 27.9 29.7
39.0 24.6
33.9 32.4
26.6 32.6
23.5 29.4
17.8 15.2
10.5 12.9

Pedal Position (%)

50 27.0 29.7 33.0 25.3 34.6 29.2 18.3 12.6 500
20 1000
29.2 1500
32.0 2000
35.6 2500
27.9 3000
39.0 3500
33.9 4000
26.6 4500
32.6 29.4
30 28.8
60 31.6
90
26.1 35.2
23.4
28.7 25.8
25.7
31.9 36.4
28.6
25.1 31.7
24.3
33.9 18.1
31.7
27.3 25.5
21.6
18.1 22.7
17.7
12.7 13.5
10.6 11.3 0 23.9
30 26.3
28.8 29.4
31.6 37.6
35.2 39.5
25.8 31.8
36.4 33.4
31.7 34.5
18.1 31.4
25.5 22.7
10 27.7 30.4
Engine
33.8
Speed
42.9
(RPM)
45.0 36.5 38.3 39.5 36.0
40 27.9
70 30.7
100
25.2 34.1 24.7
22.5
27.7 25.5 27.5
30.8 35.4 24.1
24.8 30.8 30.9
33.1 16.5
25.4 18.4
19.7
18.0 16.0
17.6
13.5 12.9
11.3 10.8
mxnxk 40 27.9 30.7 34.1 25.5 35.4 30.8 16.5 18.4 16.0

Pedal Position (%)

50 27.0 29.7 33.0 29.7
25.3 24.6
34.6 32.4
29.2 23.5
18.3 12.6 10.5 20500
500 501000
1000
29.2 1500
27.01500
32.0 2000
29.72000
35.6 2500
33.02500
27.9 3000
25.33000
39.0 3500
3500
33.9
34.6 4000
4000
26.6
29.2 4500
4500
32.6
18.3 29.4
12.6 10.5
80 24.3 26.7 17.8 15.2 12.9
00 23.9
23.9
30 26.3
26.3
28.8 26.1 29.4
29.4
31.6 28.7 37.6
37.6
35.2 31.9 39.5
39.5
25.8 25.1 31.8
31.8
36.4 33.9 33.4
33.4
31.7 27.3 34.5
34.5
18.1 18.1 31.4
31.4
25.5 12.722.7
60 26.1
90 28.7
23.4 31.9 28.6
25.7 25.1 24.3
33.9 31.7
27.3 21.6
18.1 12.7
17.7 10.6
13.5 11.3 60 10.6
Speed (RPM) 10
10 27.7
27.7
40 30.4 33.8
30.4 30.7
27.9 42.9
33.8 34.1 45.0
42.9 25.5 36.5
45.0 35.4 38.3
36.5 30.8 39.5
38.3 16.5 36.0
39.5 18.436.0 16.0
)

70 25.2
100 27.7
22.5 30.8 27.5
24.7 24.8 24.1
33.1 30.9
25.4 19.7
18.0 13.5
17.6 11.3
12.9 10.8 70 25.2 27.7 30.8 24.8 33.1 25.4 18.0 13.5 11.3

Pedal Position (%)

(°

20
20 29.2
29.2
50 32.0 35.6
32.0 29.7
27.0 27.9
35.6 33.0 39.0
27.9 25.3 33.9
39.0 34.6 26.6
33.9 29.2 32.6
26.6 18.3 29.4
32.6 12.629.4 10.5
80 24.3 26.7 29.7 24.6 32.4 23.5 17.8 15.2 12.9 80 24.3 26.7 29.7 24.6 32.4 23.5 17.8 15.2 12.9
30 28.8 31.6 35.2 25.8 36.4 31.7 18.1 25.5 22.7
T

30 28.8
60 31.6 28.7
26.1 35.2 31.925.8 25.136.4 33.931.7 27.318.1 18.125.5 12.722.7 10.6
90 23.4 25.7 28.6 24.3 31.7 21.6 17.7 13.5 11.3
VC

90 23.4 25.7 28.6 24.3 31.7 21.6 17.7 13.5 11.3 40

40 27.9
27.9
70 30.7 34.1
30.7 27.7
25.2 25.5
34.1 30.8 35.4
25.5 24.8 30.8
35.4 33.1 16.5
30.8 25.4 18.4
16.5 18.0 16.0
18.4 13.516.0 11.3
100 22.5 24.7 27.5 24.1 30.9 19.7 17.6 12.9 10.8
100 22.5 24.7 27.5 24.1 30.9 19.7 17.6 12.9 10.8 50
50 27.0
27.0
80 29.7 33.0
29.7 26.7
24.3 25.3
33.0 29.7 34.6
25.3 24.6 29.2
34.6 32.4 18.3
29.2 23.5 12.6
18.3 17.8 10.5
12.6 15.210.5 12.9
60
60 26.1
26.1
90 28.7 31.9
28.7 25.7
23.4 25.1
31.9 28.6 33.9
25.1 24.3 27.3
33.9 31.7 18.1
27.3 21.6 12.7
18.1 17.7 10.6
12.7 13.510.6 11.3
70
70 25.2
25.2
100 27.7 30.8
27.7 24.7
22.5 24.8
30.8 27.5 33.1
24.8 24.1 25.4
33.1 30.9 18.0
25.4 19.7 13.5
18.0 17.6 11.3
13.5 12.911.3 10.8
80
80 24.3
24.3 26.7
26.7 29.7
29.7 24.6
24.6 32.4
32.4 23.5
23.5 17.8
17.8 15.2
15.2 12.9
12.9
90
90 23.4
23.4 25.7
25.7 28.6
28.6 24.3
24.3 31.7
31.7 21.6
21.6 17.7
17.7 13.5
13.5 11.3
11.3
100
100 22.5
22.5 24.7
24.7 27.5
27.5 24.1
24.1 30.9
30.9 19.7
19.7 17.6
17.6 12.9
12.9 10.8
10.8

FIGURE 3. Illustration of the exponential rise of the map size and complexity as the control parameters increase (the table contents are blurred
for proprietary reasons)

created model, a combination of speed, load, VCT and spark

sweeps is designed for this experiment. The sweeps are IV. NEURAL NETWORK PARAMETER SETTING AND
comprehensive enough to cover all the normal operating PERFORMANCE EVALUATION
conditions of this engine.
The range of each control parameter is shown in Table I. A. MODEL PARAMETER SELECTION AND DATA
PREPARATION
The total number of the collected data samples is 3356, which
The focus of this study is to create a model that can predict
substantially surpasses the number of data samples in all the
gasoline engine output torque from a given engine input data
ANN models reported in Section I. All the dyno cell
set. The predicted values are compared with the measured
measurements and the ECU data are sampled at 10 Hz rate by
values to assess the model’s capability of prediction. The
the dyno control system. In order to log only the steady-state
evaluation metrics are root mean square error (RMSE) and
data, the test is scheduled to stay at each running condition for
correlation coefficient. The model input data are comprised of
3 minutes. At 10 seconds to the end of each condition, a data
the parameters that can directly affect the combustion process
point is logged, which is the average of the data sampled in the
(such as VCT and spark advance angle) and the ones that are
previous 30 seconds.
TABLE I less relevant (such as barometric pressure and exhaust gas
RANGE OF THE CONTROL PARAMETERS temperature). All the collected data points are randomly
ECU Parameter Sweep Range Step Length partitioned as the training data set, the validation data set and
the testing data set at a ratio of 8:1:1 (or very close to 8:1:1).
Engine Speed 1000 - 4500rpm 500 rpm
As a result, there are 2685 data points in the training data set,
Pedal 0 - 100% 10%
335 data points in the validation data set, and 336 data points
VCT 0 - 50 deg 5 deg in the test data set. A list of the model input/output parameters
borderline - 4 deg before and their statistics in the training and the testing data sets are
Spark 1 deg
borderline
shown in Table II.

VOLUME XX, 2020 6

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3000152, IEEE Access

TABLE II
MODEL INPUT/OUTPUT PARAMETERS AND STATISTICS
Training Data Set Validation Data Set Testing Data Set
Model Parameters I/O
Min Mean Max Min Mean Max Min Mean Max
Engine Torque (Nm) Output 55.3 323.7 671.6 54.6 321.3 635.6 55 326.7 657.4
Engine Speed (rpm) Input 999 2658 4504 999 2728 4504 999 2586 4504
Intake Manifold Pressure (kPa) Input 24.9 71.4 100.3 24.9 70.4 100.3 25 71 100.3
Barometric Pressure (kPa) Input 98.2 99.7 100.9 98.2 99.7 100.9 98.2 99.8 100.9
Intake Air Temperature (degC) Input 17.7 24.2 35.5 19.3 24.2 35.3 19.4 24.2 34.6
VCT (deg) Input 0 25 50 0 24 50 0 23 50
Spark Advance Angle (deg) Input 0 28 55 6.5 28 55 5.75 27.9 55
Lambda (-) Input 0.84 0.96 1 0.84 0.95 1 0.84 0.96 1
Exhaust Gas Temperature (degC) Input 362.5 712.2 938.7 363.1 716.4 929.3 398.3 707.3 934.2
Exhaust Gas Pressure (kPa) Input 1.7 5.6 13.6 1.9 5.7 13.5 2.1 5.5 13.6

Hidden Layer
Input Layer
h1
x1-Raw x1
Output Layer
Normalization

h2
(𝑤11 , 𝑏11 ), (𝑤21 , 𝑏21 ), …, (𝑤𝑁̃ 1, 𝑏𝑁̃ 1)
x2-Raw x2 (𝑤12 , 𝑏12 ), (𝑤22 , 𝑏22 ), …, (𝑤𝑁̃ 2, 𝑏𝑁̃ 2)
Normalization (β1,…, ββ𝑁̃k)T y Restoration y-Raw
…… h3
(𝑤1𝑚 , 𝑏1𝑚 ), (𝑤2𝑚 , 𝑏2𝑚 ), …, (𝑤𝑁̃ 𝑚, 𝑏𝑁̃ 𝑚)

xn-Raw xm
Normalization
̃
𝑁

FIGURE 4. Network structure with data normalization and restoration schemes

As pointed out [42], [43], the empirical initialization of the
weights and biases in the random weight networks with non-
iterative training algorithms, such as ELM, may not always
lead to the optimal performance. Therefore, tuning of the
randomization range is required to achieve desirable
performance. In addition, since many of the activation
functions (such as sine or sigmoid function) have periodic or
asymptotic output ranges, it is preferable that given the
randomized weights and biases, the output of the neuron
response should avoid such ranges, so the change of the input
value can be reflected sufficiently on the output. In this
research, the activation function is selected as sigmoid
function. All the input/output parameters (such as those listed
in Table II) are normalized to [0, 1], where 0 corresponds to
95% of their minimum values and 1 corresponds to 105% of
their maximum values, respectively. Such normalization can
create some room of redundancy in case the estimated unseen
values fall slightly out of the range of the training values. The
randomization range of the hidden neuron weights and biases
are both selected as [-1, 1] after parameter tuning. For this
application, such settings would make the overall hidden
neuron inputs reside in [-3.6, 2.9], which is a suitable input
range to avoid the asymptotic output range of the activation
function. After the neural network calculation, the output data
shall be restored to their original scales to get the raw
engineering values. The network structure with data
normalization and restoration is illustrated in Fig. 4. The
training data set are used to train the network in order to extract
the data feature. The validation data set are used to validate the
trained model and tune the appropriate 𝛽 estimator. The test
data set are used to provide unbiased evaluations of the model
once the model is determined.
To finalize the model structure, the number of hidden
neurons also needs to be determined. Traditionally the manual
trial-and-error approach is used to find an acceptable number
of the hidden nodes. However, this approach is time-
consuming and cannot always guarantee that the optimal
number of hidden nodes exists in the experimental range. In
this research, random orthogonal projection based enhanced
bidirectional ELM (OEB-ELM) [44] is adopted to find the
optimal number of hidden neurons. Compared to B-ELM [45]
and EB-ELM [46] algorithms, this approach leads to better
generalization and a more compact network structure. Five
sets of the orthogonalized initialization matrices are evaluated
at each odd-hidden node step to search for the appropriate
initialization set that can deliver the minimum residue error. In
order to explore how the increase of the number of hidden
neurons would affect the performance of the model, sweeps of

VOLUME XX, 2020 7

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3000152, IEEE Access

the number of hidden neurons versus the normalized RMSE
are conducted on the training data set and the validation data
set. The residue error exit condition is not considered in the
sweep so the full curve trend can be revealed. As shown in Fig.
5, when the number of hidden neurons reaches 32, the
normalized RMSEs of both data sets start to converge, with
both values smaller than 0.1. Based on this discovery, the
number of hidden neurons is determined as 32 for this model.
1.1
1.0
0.9
0.8
Normalized RMSE
the predicted and observed values of the training data set as
shown in Fig. 6 (a). The RMSE is 9.18 Nm and the correlation
coefficient is 0.9983 between the predicted and observed
values of the validation data set as shown in Fig. 6 (b). The
RMSE is 9.29 Nm and the correlation coefficient is 0.9981
between the predicted and observed values of the testing data
set as shown in Fig. 6 (c). The results have shown good
generalization across the training, the validation, and the
testing data sets. Considering the wide torque distribution in
Normalized RMSE Training Data
Normalized RMSE Validation Data the data sets and practical engineering sense, the results can be
considered as within acceptable range.

0.7 C. WEIGHTED REGRESSION APPROACH AND

RESULTS
0.6
0.5
1) DATA PREPARATION
The performance of the above-established model can be tuned
0.4
to deliver even higher estimation accuracy at certain data
0.3
points by accepting slightly bigger errors at other data points.
0.2 This feature is desirable when the significance of the data
0.1 points varies, or when the distribution of the samples is
0.0 imbalanced across the whole data set. For instance, as pointed
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 out in Section II-D, the LMSO points in the engine operating
Number of Hidden Nodes
map deserve higher accuracy as they are the most often run
FIGURE 5. Number of hidden neurons vs. normalized RMSE points during the standard vehicle and fuel economy test
schedules, such as UDDS and HWFET. To demonstrate how
B. MODEL TRAINING AND TESTING RESULTS this feature works, 5 representative LMSO points (𝐿𝑀𝑆𝑂𝑖 , 𝑖 =
The training data set are firstly fed into the neural network with 1,2, … ,5) are selected as the central points, around which 5
the above-determined model parameters to learn the features special areas of interest (AOI) are defined. In this research, the
of the training data set. The performance of the trained model representative LMSO points are 1000rpm@15%load,
is then assessed by checking the RMSE of the validation data 1500rpm@30%load, 1500rpm@50%load,
set with respect to the regulation factor 𝑘. As described in 2000rpm@40%load, and 2000rpm@60%load. The size of
Section II-C, a deterministic regulation factor 𝑘 can be derived each AOI is ±50 rpm and ±4% load. The 5 representative
from the ten-fold cross-validation approach. If the RMSE of points and their respective AOIs are plotted in Fig. 7. For
the validation data set is within 5% of the range of RMSE of convenience, the remainder of the paper will address all the
the training data, the regulation factor 𝑘 would be deemed as points in a single AOI by the representative point. Each
acceptable and the model would be evaluated with the test data representative 𝐿𝑀𝑆𝑂𝑖 can be assigned a unique weight factor
set. Otherwise, the factor 𝑘 would have to be re-calculated. In 𝑤𝑖 (𝑤𝑖 ≥ 1). A square diagonal weight matrix is defined as
this gasoline engine torque model, the optimal 𝑘 is determined 𝑊 = 𝑑𝑖𝑎𝑔(𝑤1 , 𝑤2 , … , 𝑤𝑁 ) , where 𝑁 is the number of
as 5.007e-6 through this approach. The RMSE is 8.75 Newton training data points. In such a way, each training data point
Meter (Nm) and the correlation coefficient is 0.9983 between 𝐷𝑃𝑡 has a corresponding adjustable weight 𝑤𝑡 at the 𝑡𝑡ℎ
700 700 700
Training (Observed) Validation (Observed) Testing (Observed)
600 Training (Predicted) 600 Validation (Predicted) 600 Testing (Predicted)
Data Count = 2685 Data Count = 335 Data Count = 336
500 RMSE = 8.75 Nm 500 RMSE = 9.18 Nm 500 RMSE = 9.29 Nm
Torque [Nm]

Torque [Nm]

Torque [Nm]

Corr.Coef. = 0.9983 Corr.Coef. = 0.9983 Corr.Coef. = 0.9981

400 400 400

300 300 300

200 200 200

100 100 100

0 0 0
0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700
Torque [Nm] Torque [Nm] Torque [Nm]
(a) (b) (c)
FIGURE 6. Torque model regression results of (a) the training data set, (b) the validation data set, and (c) the testing data set

VOLUME XX, 2020 8

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3000152, IEEE Access

diagonal position of matrix 𝑊. If a 𝐷𝑃𝑡 falls in the AOI of an TABLE III

THE COMPOSITION OF THE TRAINING AND THE TESTING DATA SETS
𝐿𝑀𝑆𝑂𝑖 , then the diagonal weight element of matrix 𝑊 is set
Training Validation Testing
as 𝑤𝑡 = 𝑤𝑖 . Otherwise, 𝑤𝑡 = 1, which means no accuracy
Data Set Data Set Data Set
enhancement is applied to that data point. In this manner, it is Total Number 2685 335 336
possible to customize the weight factor of each data point and LMSO Points 285 37 34
best adapt this approach to the specific needs of different Non-LMSO Points 2400 298 302
applications. To simplify the demonstration, the same weight
factor is applied to all the representative LMSO points. Once
the weight matrix 𝑊 is determined, the torque regression 2) WEIGHTED REGRESSION RESULT
model derived from the previous section can be re-trained with In order to grasp the trend of how the weight factor would
the weight matrix to deliver more accurate estimations for the affect RMSE and correlation coefficient, sweeps are
points in the AOIs. conducted on the training, the validation, and testing data sets.
As can be seen in Fig. 8. the increase of the weight factor can
100
LMSO1 = 1000rpm@15%Load reduce the RMSE of the LMSO points, which means the
90 LMSO2 = 1500rpm@30%Load prediction accuracy of these points is improved. On the other
80 LMSO3 = 1500rpm@50%Load hand, the bigger weight factor would also cause greater RMSE
LMSO4 = 2000rpm@40%Load on the non-LMSO points as well as the overall data set.
70
60 LMSO5 = 2000rpm@60%Load
60 However, the accuracy gain on the LMSO points outweighs
Load [%]

50 the accuracy loss on the non-LMSO points significantly in this

50
40 experiment. For example, the RMSE of the LMSO points can
40 be reduced by about 6 Nm, while the RMSE of the other points
30
30 only increases by less than 3 Nm when the weight factor
20 15 reaches 60.
10 Similar patterns can also be observed on the correlation
coefficient. As demonstrated in Fig. 9, significant increases of
0
1000 1500 2000 2500 3000 3500 4000 the correlation coefficients are exhibited on the LMSO points
Engine Speed [rpm] in all data sets as the weight factor increases. In the meantime,
the correlation coefficients of the non-LMSO points and the
FIGURE 7. Selected LMSO points in a torque and engine speed map
overall data set only suffer very limited drops. In this
In order to compare the performance between the weighted experiment, the gain of the correlation coefficient is about
regression and the non-weighted regression, the same training, 0.003, whereas the loss is less than 0.001 for both data sets
validation and testing data sets as in Section IV-A are reused when the weight factor reaches 60.
in this experiment. In addition, the same neuron initialization When 𝑤 = 1 (no weight applied), the overall RMSEs are
and regulation factor 𝑘 as in Section IV-B are also reused. The 8.75 Nm, 9.18 Nm, and 9.29 Nm for the training, the
evaluation metrics are the RMSE and correlation coefficient validation, and the testing data sets, respectively; the overall
of the LMSO points before and after applying the weight correlation coefficients are 0.9983, 0.9983, and 0.9981 for the
matrix. The number of LMSO points and the non-LMSO training, validation, and testing data sets, respectively. These
points in the training, the validation, and the testing data sets numbers match the results of the torque regression model
are listed in Table III. demonstrated in Section IV-B, which indicates that the
weighted and non-weighted approaches can essentially be

16 16 16
Training_Overall Validation_Overall Testing_Overall
14 Training_Non-LMSO_Points 14 Validation_Non-LMSO_Points 14 Testing_Non-LMSO_Points
Training_LMSO_Points Validation_LMSO_Points Testing_LMSO_Points

12 12 12
RMSE [Nm]

RMSE [Nm]

RMSE [Nm]

10 10 10

8 8 8

6 6 6

4 4 4

2 2 2
0 5 10 15 20 25 30 35 40 45 50 55 60 0 5 10 15 20 25 30 35 40 45 50 55 60 0 5 10 15 20 25 30 35 40 45 50 55 60
Weight Factor Weight Factor Weight Factor
(a) (b) (c)
FIGURE 8. Impact of the change of weight factor on the RMSEs on (a) the training data set, (b) the validation data set, and (c) the testing data set

VOLUME XX, 2020 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3000152, IEEE Access

1.001 1.001 1.001

Training_Overall Validation_Overall Testing_Overall
Training_Non-LMSO_Points Validation_Non-LMSO_Points Testing_Non-LMSO_Points
1.000 Training_LMSO_Points 1.000 Validation_LMSO_Points 1.000 Testing_LMSO_Points

0.999 0.999 0.999

Corr. Coef.

Corr. Coef.

Corr. Coef.
0.998 0.998 0.998

0.997 0.997 0.997

0.996 0.996 0.996

0.995 0.995 0.995

0 5 10 15 20 25 30 35 40 45 50 55 60 0 5 10 15 20 25 30 35 40 45 50 55 60 0 5 10 15 20 25 30 35 40 45 50 55 60
Weight Factor Weight Factor Weight Factor
(a) (b) (c)
FIGURE 9. Impact of the change of weight factor on the correlation coefficients on (a) the training data set, (b) the validation data set, and (c) the
testing data set

16 16 16
Training Non-Weighted Valid. Non-Weighted Testing Non-Weighted
14 Training Weighted 14 Validation Weighted 14 Testing Weighted
12.3
12.28
12 10.9
10.92 12 12 11.1
11.08 11.0
10.99

9.3 9.3
RMSE [Nm]

RMSE [Nm]

RMSE [Nm]
10 8.9
8.88 9.30 10 10 9.31
7.8
7.77 7.8
7.80 7.8
7.82
8 8 7.3
7.29 8
6.3
6.29
6 5.3
5.29 6 5.3
5.28 6
3.9 4.2
4.23 4.0
4.00
3.86
4 3.1
3.09 3.0
2.96 4 3.2
3.24 3.0 4 2.9
2.88
3.2
3.25
2.7
2.72 2.6
2.58 2.6
2.65 2.96
2.5
2.45 2.8
2.75
2.1
2.10
2 2 1.3
1.34 2

0 0 0
LMSO1 LMSO2 LMSO3 LMSO4 LMSO5 LMSO1 LMSO2 LMSO3 LMSO4 LMSO5 LMSO1 LMSO2 LMSO3 LMSO4 LMSO5
(a) (b) (c)

FIGURE 10. RMSE comparison of the individual LMSO areas on (a) the training data set, (b) the validation data set, and (c) the testing data set
when weight factor = 30

unified as one general approach. Another interesting validation, and the testing data sets. These observations can be
observation is that when 𝑤 continues to grow, the marginal very helpful in improving the regression model. Firstly, the
benefits such as the decreasing RMSE and the increasing inconsistency of enhancement among different LMSO areas
correlation coefficient on the LMSO points start to diminish, indicates that a unified weight factor does not suit all the data
but the side effects like rising RMSE and decreasing regions. The quantitative difference can help determine better-
correlation coefficient on the other points keep expanding. customized weight factors for the individual data regions.
This trait reveals that the optimal 𝑤 should be determined Secondly, the inconsistent behavior among the training, the
based on the trade-off between the two aspects according to validation, and the testing data sets implies the data partition
the specific applications. may be imbalanced across the data sets. Therefore, the feature
In addition to considering all the LMSO points as a whole, learned in the one data set may not suit other data sets in the
it is worth exploring how much improvement each LMSO area same way. The quantitative difference can indicate the
has with the introduction of the weight factor. In order to do distribution discrepancy, which can help adjust the use of the
so, comparisons of the same LMSO areas before and after available data to deliver better generalization of the model.
applying the weight factor are also conducted in this research. In order to visualize the enhancement brought by the weight
As demonstrated in Fig. 10, the introduction of the weight factor on the individual data points, the comparison of the
factor (e.g., 𝑤 = 30) can reduce the RMSE by approximately observed data, non-weighted regression data, and the weighted
2~8 Nm to the LMSO areas in all data sets. Similar patterns regression data are illustrated in Fig. 12. For convenience, only
can be also observed on the correlation coefficients as shown LMSO1 points are presented. As can be seen in the figure, the
in Fig. 11. Generally, a higher correlation coefficient occurs weighted regression points are closer to the observed points
where a smaller RMSE exists, which means that the than the non-weighted regression points (e.g., the weighted
introduction of weight factor does enhance the regression regression data are closer to the 45-degree line than the non-
accuracy of that LMSO area. However, the extent of weighted regression points). This matches the finding in the
enhancement could be different among the training, the above section that the weighted regression data have smaller

VOLUME XX, 2020 10

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3000152, IEEE Access

1.10 1.10 1.10

Training Non-Weighted Validation Non-Weighted Testing Non-Weighted
1.05 Training Weighted 1.05 Validation Weighted 1.05 Testing Weighted
0.998
0.997 0.991
0.994 0.998 0.991 0.991
0.994 0.9970.991 0.984 0.991 0.991
1.00 0.983
0.983 0.981
0.981 1.00 1.00 0.984 0.971
0.964 0.971
0.964 0.943 0.942
0.934 0.95 0.943 0.940 0.95 0.935
0.935 0.934

Corr. Coef.

Corr. Coef.
0.95
Corr. Coef.

0.925 0.934 0.934

0.908 0.925 0.904 0.908
0.908 0.895 0.908 0.892 0.908
0.895 0.892
0.90 0.90 0.90
0.863
0.863
0.85 0.85 0.85
0.799 0.798
0.799 0.780 0.789
0.80 0.80 0.780 0.776
0.776
0.80 0.772
0.772
0.741
0.75 0.741 0.75 0.722 0.75
0.863
0.70 0.70 0.70
LMSO1 LMSO2 LMSO3 LMSO4 LMSO5 LMSO1 LMSO2 LMSO3 LMSO4 LMSO5 LMSO1 LMSO2 LMSO3 LMSO4 LMSO5

(a) (b) (c)

FIGURE 11. Correlation coefficient comparison of the individual LMSO areas on (a) the training data set, (b) the validation data set, and (c) the
testing data set when weight factor = 30

LMSO1_Training_Observed LMSO1_Validation_Observed LMSO1_Testing_Observed

LMSO1_Training_Non-Weighted LMSO1_Validation_Non-Weighted LMSO1_Testing_Non-Weighted
LMSO1_Training_Weighted LMSO1_Validation_Weighted LMSO1_Testing_Weighted
105 105 105

100 100 100

Torque_LMSO1 [Nm]
Torque_LMSO1 [Nm]

Torque_LMSO1 [Nm]

95 95 95

90 90 90

85 85 85

80 80 80
80 85 90 95 100 105 80 85 90 95 100 105 80 85 90 95 100 105
Torque_LMSO1 [Nm] Torque_LMSO1 [Nm] Torque_LMSO1 [Nm]
(a) (b) (c)

FIGURE 12. Comparison of regression results of the LMSO1 points in (a) the training data set, (b) the validation data set, and (c) the testing data
set when weight factor = 30

RMSE and higher correlation coefficient than the non- improvement in the desired regions individually, this research
weighted regression data. has also revealed that optimal overall improvement can be
achieved by customizing the weight factor for each desired
V. SUMMARY AND FUTURE WORK region accordingly.
This paper has presented an ANN approach to predict gasoline The modeling and tuning approaches presented in this paper
engine output torque with high accuracy using an ELM based are particularly helpful for rapid engine model creation and
single-hidden layer feedforward neural network. Ten-fold cross verification. In the future, the authors plan to study the
cross-validation is adopted to stabilize the model. Totally 3356 improvement of the ELM algorithm to achieve even higher
real-world engine experiment points have been collected for accuracy.
this model, in which 2685 and 335 experiment points have
been used to train and validate the model, respectively. The REFERENCES
remaining 336 collected data points have been kept unseen by [1] S. Pischinger, “Current and Future Challenges for Automotive
the model and are used to test the model unbiasedly. The Catalysis: Engine Technology Trends and Their Impact,” Top.
Catal., vol. 59, no. 10–12, pp. 834–844, Jul. 2016, doi:
results have demonstrated the model has the capability in 10.1007/s11244-016-0557-3.
predicting the engine torque with high accuracy. The RMSE [2] B. Wu, Z. Filipi, D.N. Assanis, D. M. Kramer, G. L. Ohl, M. J.
between the observed torque and the predicted torque is about Prucka, E. DiValentin, “Using Artificial Neural Networks for
Representing the Air Flow Rate through a 2.4 Liter VVT Engine,”
9 Nm, which is about 2.7% of the mean torque value. In presented at the 2004 Powertrain & Fluid Systems Conference &
addition, this research has also proposed a weight factor Exhibition, Tampa, USA, Oct. 2004, doi: 10.4271/2004-01-3054.
approach to further enhance the model accuracy in the [3] S. Tolou, R. T. Vedula, H. Schock, G. Zhu, Y. Sun, and A. Kotrba,
“Combustion Model for a Homogeneous Turbocharged Gasoline
designated data regions. The experiment has shown that the Direct-Injection Engine,” J. Eng. Gas Turbines Power, vol. 140, no.
RMSE in these regions can be further reduced by 10, p. 102804, Oct. 2018, doi: 10.1115/1.4039813.
approximately 6 Nm. Moreover, by exploring the quantitative

VOLUME XX, 2020 11

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3000152, IEEE Access
[4] N. Togun, S. Baysec, and T. Kara, “Nonlinear modeling and
identification of a spark ignition engine torque,” Mech. Syst. Signal
Process., vol. 26, pp. 294–304, Jan. 2012, doi:
10.1016/j.ymssp.2011.06.010.
[5] Z. Tan and R. D. Reitz, “An ignition and combustion model based
on the level-set method for spark ignition engine multidimensional
modeling,” Combust. Flame, vol. 145, no. 1–2, pp. 1–15, Apr. 2006,
doi: 10.1016/j.combustflame.2005.12.007.
[6] P. Boudier, S. Henriot, T. Poinsot, and T. Baritaud, “A model for
turbulent flame ignition and propagation in spark ignition engines,”
Symp. Int. Combust., vol. 24, no. 1, pp. 503–510, 1992.
[7] C. S. Daw, C. E. A. Finney, J. B. Green, M. B. Kennel, J. F. Thomas,
and F. T. Connolly, “A Simple Model for Cyclic Variations in a
Spark-Ignition Engine,” presented at the 1996 SAE International
Fall Fuels and Lubricants Meeting and Exhibition, Toronto,
Canada, Oct. 1996, p. 962086, doi: 10.4271/962086.
[8] S. G. Poulos and J. B. Heywood, “The effect of chamber geometry
on spark-ignition engine combustion,” SAE Trans., pp. 1106–1129,
1983.
[9] G. Fontana and E. Galloni, “Variable valve timing for fuel economy
improvement in a small spark-ignition engine,” Appl. Energy, vol.
86, no. 1, pp. 96–105, Jan. 2009, doi:
10.1016/j.apenergy.2008.04.009.
[10] J. M. Luján, C. Guardiola, B. Pla, and P. Bares, “Estimation of
trapped mass by in-cylinder pressure resonance in HCCI engines,”
Mech. Syst. Signal Process., vol. 66–67, pp. 862–874, Jan. 2016,
doi: 10.1016/j.ymssp.2015.05.016.
[11] R. W. Weeks and J. J. Moskwa, “Automotive Engine Modeling for
Real-Time Control Using MATLAB/SIMULINK,”SAE
Trans.Mater Manuf Process,Vol. 104, p. 950417, Feb. 1995,doi:
10.4271/950417.
[12] G. Jiang, P. Xie, H. He, and J. Yan, “Wind Turbine Fault Detection
Using a Denoising Autoencoder With Temporal Information,”
IEEEASME Trans. Mechatron., vol. 23, no. 1, pp. 89–100, Feb.
2018, doi: 10.1109/TMECH.2017.2759301.
[13] N. K. Togun and S. Baysec, “Prediction of torque and specific fuel
consumption of a gasoline engine by using artificial neural
networks,” Appl. Energy, vol. 87, no. 1, pp. 349–355, Jan. 2010, doi:
10.1016/j.apenergy.2009.08.016.
[14] Y. Cay, “Prediction of a gasoline engine performance with artificial
neural network,” Fuel, vol. 111, pp. 324–331, Sep. 2013, doi:
10.1016/j.fuel.2012.12.040.
[15] A. Di Mauro, H. Chen, and V. Sick, “Neural network prediction of
cycle-to-cycle power variability in a spark-ignited internal
combustion engine,” Proc. Combust. Inst., vol. 37, no. 4, pp. 4937–
4944, 2019, doi: 10.1016/j.proci.2018.08.058.
[16] Y. Hu, H. Chen, P. Wang, H. Chen, and L. Ren, “Nonlinear model
predictive controller design based on learning model for
turbocharged gasoline engine of passenger vehicle,” Mech. Syst.
Signal Process., vol. 109, pp. 74–88, Sep. 2018, doi:
10.1016/j.ymssp.2018.02.012.
[17] H. Li, K. Butts, K. Zaseck, D. Liao-McPherson, and I.
Kolmanovsky, “Emissions Modeling of a Light-Duty Diesel Engine
for Model-Based Control Design Using Multi-Layer Perceptron
Neural Networks,” presented at the WCXTM 17: SAE World
Congress Experience, Detroit, USA ,Mar. 2017, pp. 2017-01–0601,
doi: 10.4271/2017-01-0601.
[18] T. Zheng, Y. Zhang, Y. Li, and L. Shi, “Real-time combustion
torque estimation and dynamic misfire fault diagnosis in gasoline
engine,” Mech. Syst. Signal Process., vol. 126, pp. 521–535, Jul.
2019, doi: 10.1016/j.ymssp.2019.02.048.
[19] J. Chen and R. B. Randall, “Intelligent diagnosis of bearing knock
faults in internal combustion engines using vibration simulation,”
Mech. Mach. Theory, vol. 104, pp. 161–176, Oct. 2016, doi:
10.1016/j.mechmachtheory.2016.05.022.
[20] R. Ahmed, M. El Sayed, S. A. Gadsden, J. Tjong, and S. Habibi,
“Automotive Internal-Combustion-Engine Fault Detection and
Classification Using Artificial Neural Network Techniques,” IEEE
Trans. Veh. Technol., vol. 64, no. 1, pp. 21–33, Jan. 2015, doi:
10.1109/TVT.2014.2317736.
[21] L. Wen, X. Li, L. Gao, and Y. Zhang, “A New Convolutional Neural
Network-Based Data-Driven Fault Diagnosis Method,” IEEE
VOLUME XX, 2020
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
Trans. Ind. Electron., vol. 65, no. 7, pp. 5990–5998, Jul. 2018, doi:
10.1109/TIE.2017.2774777.
[22] Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew, “Extreme
learning machine: a new learning scheme of feedforward neural
networks,” in 2004 IEEE International Joint Conference on Neural
Networks (IEEE Cat. No.04CH37541), Budapest, Hungary, 2004,
vol. 2, pp. 985–990, doi: 10.1109/IJCNN.2004.1380068.
[23] W. Cao, X. Wang, Z. Ming, and J. Gao, “A review on neural
networks with random weights,” Neurocomputing, vol. 275, pp.
278–287, Jan. 2018, doi: 10.1016/j.neucom.2017.08.040.
[24] B. Igelnik and Yoh-Han Pao, “Stochastic choice of basis functions
in adaptive function approximation and the functional-link net,”
IEEE Trans. Neural Netw., vol. 6, no. 6, pp. 1320–1329, Nov. 1995,
doi: 10.1109/72.471375.
[25] W. F. Schmidt, M. A. Kraaijveld, and R. P. W. Duin, “Feedforward
neural networks with random weights.pdf,” in Proceedings., 11th
IAPR International Conference on Pattern Recognition., The Hague,
Netherlands, 1992, vol. Vol II, pp. 1–4.
[26] G.-B. Huang, “An Insight into Extreme Learning Machines:
Random Neurons, Random Features and Kernels,” Cogn. Comput.,
vol. 6, no. 3, pp. 376–390, Sep. 2014, doi: 10.1007/s12559-014-
9255-2.
[27] X.-Z. Wang, R. Wang, and C. Xu, “Discovering the Relationship
Between Generalization and Uncertainty by Incorporating
Complexity of Classification,” IEEE Trans. Cybern., vol. 48, no. 2,
pp. 703–715, Feb. 2018, doi: 10.1109/TCYB.2017.2653223.
[28] Guang-Bin Huang and H. A. Babri, “Upper bounds on the number
of hidden neurons in feedforward networks with arbitrary bounded
nonlinear activation functions,” IEEE Trans. Neural Netw., vol. 9,
no. 1, pp. 224–229, Jan. 1998, doi: 10.1109/72.655045.
[29] N. J. Guliyev and V. E. Ismailov, “A single hidden layer
feedforward network with only one neuron in the hidden layer can
approximate any univariate function,” Neural Comput., vol. 28, no.
7, pp. 1289–1304, Jul. 2016, doi: 10.1162/NECO_a_00849.
[30] Guang-Bin Huang, Hongming Zhou, Xiaojian Ding, and Rui Zhang,
“Extreme Learning Machine for Regression and Multiclass
Classification,” IEEE Trans. Syst. Man Cybern. Part B Cybern., vol.
42, no. 2, pp. 513–529, Apr. 2012, doi:
10.1109/TSMCB.2011.2168604.
[31] A. E. Hoerl and R. W. Kennard, “Ridge Regression: Biased
Estimation for Nonorthogonal Problems,” Technometrics, vol. 12,
no. 1, p. 55, Feb. 1970, doi: 10.2307/1267351.
[32] G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, “Extreme learning
machine: Theory and applications,” Neurocomputing, vol. 70, no.
1–3, pp. 489–501, Dec. 2006, doi: 10.1016/j.neucom.2005.12.126.
[33] G. H. Golub, M. Heath, and G. Wahba, “Generalized Cross-
Validation as a Method for Choosing a Good Ridge Parameter,”
Technometrics, no. 21(2), pp. 215–223, 1979.
[34] W. Pan, “Akaike’s Information Criterion in Generalized Estimating
Equations,” Biometrics, vol. 57, no. 1, pp. 120–125, Mar. 2001, doi:
10.1111/j.0006-341X.2001.00120.x.
[35] K. O. Kvålseth, “Cautionary Note About R 2,” Am. Stat., vol. 39,
no. 4, pp. 279–285, 1985.
[36] O. Harel, “The estimation of R 2 and adjusted R 2 in incomplete
data sets using multiple imputation,” J. Appl. Stat., vol. 36, no. 10,
pp. 1109–1118, Oct. 2009, doi: 10.1080/02664760802553000.
[37] K. P. Burnham, D. R. Anderson, and K. P. Burnham, Model
selection and multimodel inference: a practical information-
theoretic approach, 2nd ed. New York: Springer, 2002.
[38] Z. Gao et al., “Drive cycle simulation of high efficiency
combustions on fuel economy and exhaust properties in light-duty
vehicles,” Appl. Energy, vol. 157, pp. 762–776, Nov. 2015, doi:
10.1016/j.apenergy.2015.03.070.
[39] W. Zong, G.-B. Huang, and Y. Chen, “Weighted extreme learning
machine for imbalance learning,” Neurocomputing, vol. 101, pp.
229–242, Feb. 2013, doi: 10.1016/j.neucom.2012.08.010.
[40] K. Li, X. Kong, Z. Lu, L. Wenyin, and J. Yin, “Boosting weighted
ELM for imbalanced learning,” Neurocomputing, vol. 128, pp. 15–
21, Mar. 2014, doi: 10.1016/j.neucom.2013.05.051.
[41] R. Fletcher, Practical methods of optimization, 2nd ed. Chichester,
West Sussex England: John Wiley & Sons, Ltd, 2013.
12
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3000152, IEEE Access

[42] W. Cao, J. Gao, Z. Ming, and S. Cai, “Some Tricks in Parameter
Selection for Extreme Learning Machine,” IOP Conf. Ser. Mater.
Sci. Eng., vol. 261, p. 012002, Oct. 2017, doi: 10.1088/1757-
899X/261/1/012002.
[43] M. Li and D. Wang, “Insights into randomized algorithms for neural
networks: Practical issues and common pitfalls,” Inf. Sci., vol. 382–
383, pp. 170–178, Mar. 2017, doi: 10.1016/j.ins.2016.12.007.
[44] W. Cao, J. Gao, X. Wang, Z. Ming, and S. Cai, “Random
Orthogonal Projection Based Enhanced Bidirectional Extreme
Learning Machine,” in Proceedings of ELM 2018, vol. 11, J. Cao,
C. M. Vong, Y. Miche, and A. Lendasse, Eds. Cham: Springer
International Publishing, 2020, pp. 1–10.
[45] Yimin Yang, Yaonan Wang, and Xiaofang Yuan, “Bidirectional
Extreme Learning Machine for Regression Problem and Its
Learning Effectiveness,” IEEE Trans. Neural Netw. Learn. Syst.,
vol. 23, no. 9, pp. 1498–1505, Sep. 2012, doi:
10.1109/TNNLS.2012.2202289.
[46] W. Cao, Z. Ming, X. Wang, and S. Cai, “Improved bidirectional
extreme learning machine based on enhanced random search,”
Memetic Comput., vol. 11, no. 1, pp. 19–26, Mar. 2019, doi:
10.1007/s12293-017-0238-1.
and alternative fuel combustion, dual-fuel applications, clean & efficient
combustion enabling, and advanced engine development. He has published
a number of Journal articles and peer-reviewed Conference papers
worldwide. Dr. Han is a member of the Society of Automotive Engineers,
American Society of Mechanical Engineers, and a licensed Professional
Engineer of Ontario.
Jimi Tjong received the B.Sc., M.Sc., and Ph.D.
degrees in Mechanical Engineering from the
University of Windsor, ON, Canada, in 1980,
1984, and 1993, respectively. He is currently an
Adjunct Professor at the University of Windsor,
University of Toronto, and McMaster University.
His research interests include production and
comprehensive design validation of components
and systems for conventional and electrified
vehicles, and optimization of automotive test
systems for cost, performance, and compatibility.

Weiying Zeng (Student Member, IEEE)

received the B.Eng. and the M. A. Sc. degrees
from China Agricultural University, Beijing,
China, in 2007 and 2009, respectively. He also
earned the B. Ec. degree from Peking University,
Beijing, China, in 2009. He was with Beijing
Automotive Group Co., Ltd. (BAIC Group),
Beijing, China, from 2009 to 2013. He is currently
working toward the Ph.D. degree in Electrical
Engineering at the University of Windsor,
Windsor, ON, Canada. His research interests include neural networks,
automated testing, fault diagnostics and optimization techniques for real-
time systems.

Mohammed A. S. Khalid (Senior Member,

IEEE) received the Ph.D. degree in Computer
Engineering from the University of Toronto in
1999. He has over 25 years of experience in
teaching, research and development in academia
and industry. Before joining the University of
Windsor in August 2003, he worked for 4 years as
a Senior Member of Technical Staff in the
Verification Acceleration R & D Group (formerly
Quickturn), of Cadence Design Systems, based in
San Jose, California. His research and development interests are in
architecture and Computer-Aided Design (CAD) for field programmable
chips and systems, CAD algorithms and tools, Electronic Design
Automation (EDA), power-efficient FPGA-based hardware acceleration for
HPC, High-Level Synthesis (HLS) for FPGAs and Automotive Embedded
Systems (AES), especially for Autonomous Vehicles (AV) and Advanced
Driver Assistance Systems (ADAS). He has published several papers in
these areas and holds a U.S. Patent in the area of the architecture of
reconfigurable systems.

Xiaoye Han received the Ph.D. degree in

Automotive Engineering from the University of
Windsor, Windsor, Canada. He is currently a
Research & Advanced Engineering Supervisor at
the Powertrain Engineering Research and
Development Centre of Ford Canada and an
Adjunct Associated Professor with the
Department of Mechanical, Automotive, and
Materials Engineering at the University of
Windsor. He has expertise in advanced powertrain
controls development, engineering and testing. His primary research focuses
on clean & efficient powertrain systems for pollutant reduction and
greenhouse gas mitigation. The research topics include low temperature
combustion, generic ECU development, active combustion control, biofuel

VOLUME XX, 2020 13

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.