IOP Conference Series: Materials Science and Engineering

PAPER • OPEN ACCESS

Development of system identification from traditional concepts to real-

time soft computing based
To cite this article: A M Kwad et al 2020 IOP Conf. Ser.: Mater. Sci. Eng. 767 012050

View the article online for updates and enhancements.

This content was downloaded from IP address 23.95.105.3 on 14/04/2020 at 12:46

1st International Symposium on Engineering and Technology (ISETech) 2019 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 767 (2020) 012050 doi:10.1088/1757-899X/767/1/012050

Development of system identification from traditional

concepts to real-time soft computing based

A M Kwad, D Hanafi, R Omar and H Abdul Rahman

Faculty of Electrical and Electronic Engineering (FKEE), Universiti Tun Hussein Onn
Malaysia (UTHM), 86400 Parit Raja, Batu Pahat, Johor, Malaysia
[email protected]

Abstract. Advancement in computer and data acquisition technology gave the system
identification a push forward, especially in the speed of the process and parallel calculation
techniques. This article reviews the progress of system identification methods and techniques
from traditional to modern methods and from offline to fast on-line analysis through previous
years till these days. It takes into account part of artificial intelligence techniques (soft computing
techniques), primarily artificial neural networks methods, for off-line system data analysis and
real-time system identification, including the linear and non-linear issues. This study will provide
a very concise survey for researchers about the development of system identification, its
implementations, and techniques that have been used.

1. Introduction
System Identification is steps of logical operation which are followed to build a mathematical model
that describes features of a dynamic system (or a process) depending on existing relations among its
parameters and variables or between the input and corresponding measured output data [1]. Each system
has parameters and variables that interconnected together to form its behavior. Sometimes, these
parameters and variables could be seen clearly [2], hence it could build a mathematical model to describe
a system with small error margin (although the mathematical model is involved, it cannot adequately
describe a real system), and that what is named in some references as a white box [3]. In contrast, a black
box [4] is a system that has unknown variables, parameter values, and their relations. All that can be
seen are the input (to the system) and output (from the system); so, the system is evaluated by observing
the output-input relationship. Finally, there is a gray box system [3], which is mixed from white box and
black box systems; there are some features known and others unknown. It is crucial to select appropriate
variables from the principle equations and tune the parameters by using regression methods to minimize
the variation between the derived mathematical model and the measured data [1]. Most modeling
operations are done after the data have been collected (off-line), but there is another type of modeling is
done within real-time to estimate the parameters of a specific system, which is named on-line system
identification [5,6]. The main equipment in the system identification are: measured data, derived model,
selected cost function, and the numerical method used to minimize the cost function (it must be fast to
convoy on-line system identification [7,8]).
Nowadays, researchers trend to use the Artificial Intelligence technique to solve many system
problems like non-linearity, time-varying, and ambiguity or inaccessibility [9,10]. This technique based

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
1st International Symposium on Engineering and Technology (ISETech) 2019 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 767 (2020) 012050 doi:10.1088/1757-899X/767/1/012050

on the learning concept, where weights that belong to a specific perception start with random values,
then these weights are tuned up to be suitable amounts. Whenever the learning increased (till a specific
limit), it makes weights close from appropriate value. Artificial intelligence techniques can be divided
into two types relative to teaching manner; supervised learning and unsupervised learning. Where the
first type compares the output of the network with the desired output (that means the desired output is
known), while the second type is learning without teacher, also it is can be divided into two kinds relative
to the size of network; single layer and multi-layer networks [11,12], and there are other classifications
according to training methods and used function.
The following sections will explain the off-line and online system identification, what is the
difference between them, and how the soft-computing algorithms were used in these two types instead
of the traditional methods.

2. Off-line System Identification

Off-line system identification is the analysis of the pairs of input-output data (with older information for
current pairs in the dynamic system) to find the internal mathematical structure of the system (model of
the system) and assess the parameters of that structure one step ahead, after collecting the data via
sensors. The steps of off-line identification can be summarized by:
1- Finding the mathematical model for the system from the principle physical rules, if the
system is a white or gray box. The black box has no known structure, so this step ignored.
2- Stimulating the system with a controlled input to get on system response.
3- Sensing the input and output signals and accumulating them in pairs form.
4- Implementing the system identification operation on the input/output data (in some cases,
the data needs to be treated before the identification operation) to estimate the value of the
parameter in the defined model.
5- After completing the identification operation, the model now ready to be compared to part
of the data through operation named model validation. If the output of the model is
convergent from the original output, it can consider the model represent the system relatively,
if not the step (4) must be returned.

Soft computing methods can substitute the traditional methods in creating Identifying a mathematical
model for either known (white box) or partially known (gray box) system’s properties, and unknown
system’s properties (black box).
De la Sen et al. [13] designed an expert network that had been created to estimate a discrete transfer
function for a linear time-invariant system with impulses sampling. They ran different methods for
system identification. One of these methods relied on recursive least squares (RLS);
A(q 1 ) yl  B(q 1 ) xl  vl
M
VM   vl (1)
l 1

ˆl  ˆl 1  k l ( yl  yˆl )

Where 𝑣𝑙 represent the calculated error, M is the amount of the samples, 𝜃̂𝑙 are the expected parameters
at time t = l, 𝑦𝑙 is the measured output at the same time t, and 𝑦̂𝑙 is the predicted output. Another method
relies on Leverrir’s algorithm;

2
1st International Symposium on Engineering and Technology (ISETech) 2019 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 767 (2020) 012050 doi:10.1088/1757-899X/767/1/012050

Y(q -1 )
G(q -1 )  -1
 yt  yt 1q 1   yt  n q  n
X(q )

  yt  r q  r
r  n 1 (2)
-1
B(q )

A(q -1 )
b1  b2 q 1  b3 q 2  ...

1  a1q 1  a2 q 2  a3q 3  ...

Where, q represents a shift time in the dynamic system q yt =yt+1 and q-1 yt= yt-1.
Artificial Intelligence is one of the best choices to solve the non-linearity problem; it has some attributes
that make it more adaptable. The soft robot was used for the motion of some parts of a body during
radiotherapy was used as a case study, using supervised deep learning for multilayers recurrent neural
networks [14]. de Canete et al. [15] indicated how to benefit from the Labview, which is a graphical
language for the data acquisition and the automation control, to Identify the distillation column process.
The nonlinearity of the system was solved by neural networks tactic.
Shamsudin and Chen [16] developed an off-line dynamics system identification approach for
unmanned copter depend on neural networks (NN). The Levenberg-Marquardt model form was applied
to teach the NNs with autoregressive exogenous (AXR). A combination of correlation analysis, cross-
validation, weights regulation, and Lipschitz criterion was used to enhance the form of the network.
Hanafi et al. [17] developed a mathematical model that relied on the physical principle laws, that
ruled the system, and the attributes of the parameters. The case study was a suspension system of a car
with nonlinearity in the damper and spring elements. The SI was based on the multilayers NNs to solve
the complex relations between the inlet and the outlet of the system. The car was examined using the
artificial bump on the road, the data were handled off-lined, and the NNs were taught by the weighted
least squares (WLS) method.
Mills and Zomaya [18] examined an off-line dynamic system identification with the assist of neural
networks. They utilized a backpropagation procedure for training the networks based on the linear
estimator of the algorithmic least-squares for the nonlinear system. Besides NNs, the Fuzzy Logic with
Gaussian member shape was used to be a six-layer fuzzy neural for improving the weights with pre-
optimization step using the genetic algorithm (GA) optimization method [19].
Cong et al. [20] followed a parametric system identification, which is one of the identification
approaches, to construct a nonlinear dynamic mathematical model representing a DC motor with the
nonlinear effect of the Coulomb friction. The suggested approach utilized GA, which is one of heuristic
and evolutionary methods, with the simplex method to solve the global optimization problems for
estimating the model. The validation, which was utilized to compare between the model and the DC
motor, was comparing the output of the motor with the output of the model (from the simulation).
Conforming to Herreros et al. [21], a straightforward model was employed to identify the suspension
system of a car applying non-interfering input signals. The amplitude difference between the real and
the estimated power magnitude of the system output for spectral density was taken as the main objective
to be reduced and to reduce a discrepancy of estimation due to distorted input (because of the noise and
ignoring the presence of dynamics), the GA was implemented as an optimizer tactic.
Jizhen et al. [22] suggested a model based on the gray-box principle for identifying an online
parametric model of wind turbine generation system (WTGS) in a closed-loop system. The system as an
aerodynamic system and a generation core of the WTGS entirely were modeled all together to obtain a
total identification model relied on a nonlinear state-space model. GA was implemented to assist in
identifying the system and assess the parameters of the suggested model.

3
1st International Symposium on Engineering and Technology (ISETech) 2019 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 767 (2020) 012050 doi:10.1088/1757-899X/767/1/012050

As stated by Cen et al. [23], a gray-box multilayers perceptron neural networks and integrator was
established for identifying a nonlinear dynamic system and estimating fault scheme. The model was
formed from simple rules of the neural networks, the essential criterion is the deviation between the
actual system output and the derived model output, which represents the error offset value for various
fault values.
The article [24] proposed a process to find the cross-section of neutron-induced reaction in a specific
condition, used weighted least squares method to regress the error and to update the neural network
weights and backpropagation technique for training the network off-line.
Kurita [25] examined the teaching procedures of multilayers NNs from the maximum likelihood
estimation (MLE) viewpoint. It was shown that Fisher information (FI) of the network, with subtraction
expectancy values of Hessian matrix (H), was gained from a weighted covariant tensor of input
variables. The training approach utilized Fisher’s scoring concept, which applied FI instead of H, that
can be symbolized by the weighted least squares technique for multilayers (with single hidden layer)
neural networks.
Puscasu and Codres [26] proposed a non-linear system identification and system control tactic
supported by structured NNs. To decrease the degree of computation complexity for whole the system,
in the research, splitting the system into subsystems by separating algorithm, was offered. Thus, each
nonlinear subsystem module has a nonlinear controller built from neural networks, and it was switched
among the networks of controllers by a dynamic switch which was done also by neural networks.
Narendra and Parthasarathy [26] presented the principles of model representation to identify and
control a non-linear dynamic system by utilizing NNs ability. The research tried to solve some problems
with applying the neural networks for higher-order multi-variable nonlinear systems, and the effects of
applying simpler models for identification and control. Quiroga et al. [27] explained utilizing neural
networks for identifying a permanent magnet synchronous motor (PMSM). The recurrent neural network
pattern used for maintenance to perform fault detection or integrity assessment. It was used
Matlab/Simulink with other techniques to implement and validate the model.
Sierra and Santos [28] examined implementing alternative tactics for identifying an unmanned quad-
copter system; artificial neural networks approach, parametric model estimation methods, neuro-fuzzy
structure, and a hybrid of some previous methods. The comparison among these methods revealed that
the hybrid method was retaining an appropriate balance between precision and computation time.
Ismeal et al. [29] investigated a permanent magnet DC motor identification using artificial neural
networks. Input-output data (armature voltage-current and angular speed, respectively) were taken to
use in building a non-parametric black-box model for identifying a dynamic system of motor. The neural
networks with hidden three-layers depended on a non-linear autoregressive network with exogenous
(NARX) input was applied. NARX network was taught by parallel constructions and scaled gradient
algorithm using the error derivative (first and second order derivatives) to decrease the error. Whereas,
Rahim et al. [30] examined utilizing a non-linear autoregressive moving average for exogenous input
(NARMAX) method with multilayers neural networks to identify a non-linear system for the DC motor.
The off-line learning of neural networks was improved by implementing the regularization method.
Lu et al. [31] built a semi on-line, locally linear model for an unmanned drone system. Angles of
attack and sideslip are measured as output for the system. The observer/Kalman filter identification
algorithm is used for identifying a fixed-wing unmanned aerodynamic system, built on the state-space
method, which developed from both lengthways and side force-components.
Pislaru and Shebani [32] utilized the feed-forward neural network established on radial basis
function, which is a function with real values that relies on the distance from the origin only, what is
known as (RBFNN) technique for identifying a nonlinear system. That technique was validated by
applying it on a DC motor, tank system, double tank system, and two lab’s models. The clustering
algorithm of K–means used to choose the RBFNN centers.

4
1st International Symposium on Engineering and Technology (ISETech) 2019 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 767 (2020) 012050 doi:10.1088/1757-899X/767/1/012050

Figure 1. RBFNN [32].

3. Real-time System Identification

Real-time (online) system identification is estimating the model parameters at the current moment
depending on the pairs of input/output data (with older information for the current pairs in the dynamic
system) to find the internal mathematical structure of the system (model of the system) one step ahead.
The steps of real-time identification can be summarized by:
1- Finding the mathematical model for the system from the principle physical rules, if the
system is a white or gray box. The black box has no known structure, so this step ignored.
2- Stimulating the system with a controlled input to get on system response.
3- Sensing the input and output signals and accumulating them in pairs form. The input signal
for the system is the same input for the model, and the subtraction of the model from the real
system output represents the current error.
4- Implementing the system identification operation recursively on the input/output data (in
some cases, the data needs to be treated before the identification operation), to estimate the
parameter's value in the defined model.
5- Using the calculating error to update the parameters’ value instantaneously.
6- After completing the identification operation, the model now ready to evaluate the model. If
the output of the model was convergent from the original output, it could consider the model
represent the system relatively; if not, it may return the whole process or start from the step
(2).

The recursive methods are dominant in real-time system identification.

An approach inspired by Takagi-Sugeno model for real-time training was proposed by Angelov and
Filev [33], it was constructed on supervised and unsupervised teaching algorithm that updated the model
parameters and structure repetitively. It was implemented to the data obtained from airconditioning
equipment operating in a certain construction.
Qian et al. [34] proposed a system identification scheme adopted a parametric online identification
method to get parameters of an actuator for a range of loads using hybrid simulation. A GA was built on
experimental data, which was loaded the input part to the actuator and the output used to assess the
convergence of estimated values for the model’s parameters from the original system behavior.
Al Seyab and Cao [35] developed an identification process based on recurrent neural networks
(RNN) for nonlinear predictive and control models. The process based on neural networks was derived
from a nonlinear state-space form to be a real-time process using a developed automatic differentiation

5
1st International Symposium on Engineering and Technology (ISETech) 2019 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 767 (2020) 012050 doi:10.1088/1757-899X/767/1/012050

algorithm. The technique was used for online optimization and predictive controller, as well. Wang et
al. [36] offered an unconventional motion-mode energy method to identify moving shapes of a vehicle
body (roll, bounce, and pitch motion modes) with a minimum error ratio by applying neural networks.
Pairan and Shamsudin [37] introduced experimental results based on the RBFNNs trained by
minimal resource allocating network approach (MRAN) for identifying the quad-chopper system in real-
time. The activity of the MRAN tactic was compared to the RBFNN tactic with a constant trace (CT)
process for 2500 pairs of input-output data samples. It was found that using MRAN appending or
eliminating hidden layer neuron tactic to optimize RBFNN improve weights estimation precision and
minimize the time of teaching.

Figure 2. System identifier based on NN [37].

Kirkpatrick et al. [38] proposed a linear system identifier by using neural networks for airplane
system, the identifier function was derived from state-space equations and the data used for training the
neural networks, to get matrices of the state space, were obtained from previous flight operation, the
same data had already been utilized for identifying the dynamic airplane system by Observer/Kalman
filtering approach. The suggested identifier compared with (O/K) filter method [39], which was locally
linear, for airplane dynamic nonlinear system identification in real-time.
Hussein et al. [40] introduced a model for identifying and controlling a DC motor speed; the input
voltage was provided from the solar power system through a DC-DC converter utilized two universal
learning networks (ULNs). The feed-forward NNs were applied to control, evaluate and update the
parameters of the nonlinear model in real-time. The essence of Rubaai and Kotaru [41] research was
how to deal with controlling a DC motor speed through utilized from feed-forward NNs to represent the
nonlinear system of the DC motor in an on-line system identification with on-line learning for neural
networks were done.
Griñó et al. [42] proposed a tactic depended on NNs structure for identifying a dynamic system which
has an unknown dynamical behavior; an adjustable NNs were added to take the place of unknown
dynamic part of the system. The model of NNs has linear parameters with updating weights
synchronously. Two techniques were used; the first technique adopted the gradient method and analysis
of the model’s output attitudes against calculated parameters, the second one used variant calculation,
which was leading to an offline statement and the invariant embedded way that converts it to an online
solution.

6
1st International Symposium on Engineering and Technology (ISETech) 2019 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 767 (2020) 012050 doi:10.1088/1757-899X/767/1/012050

Guo et al. [43] proposed a method based on (RNN) with shared and specialized memories (ShM and
SpM) or (SSMs) tactic for chaotic adaptable filtering to improve a simultaneous estimation. The SpM
involved weights of the last layer with linear specifications, whereas ShM designed for nonlinear
weights, which was pre-learned in off-line to a particular value, then implemented as fixed weights in
on-line, while the SpM is trained on-line. The method based on recursive least-squares teaching. Ibrahim
[44] suggested an approach for identifying the system depended on neural networks as the auto-adjusting
regulator. A real-time identification is implemented by RNNs to estimate parameters of discrete PID
control, which inspired by the Zeigler-Nichole method. It was tested in simulation software and
compared to the traditional PID controller.
Sadeghi and Farrokhi [45] presented a model for Identifying a nonlinear dynamic system in real-time
based on Wiener oriented blocks. The Levenberg-Marquardt algorithm with multilayers NNs were used
for the nonlinear static part in the Wiener model, the identifier of subspace with the multi-variable
approach was used for the linear dynamic part. The proposed model was simulated on the supposed
system and Lyapunov method utilized to improve the steadiness of model convergence.

Figure 3. Suggested online identification [45].

Axial Idea of PIVOŇKA et al. [46] studied the impacts on the identification procedure in high-speed
sampling and compared among three techniques of real-time identification: recursive least-squares,
gradient, and system identification approach based on neural networks with damped least-squares
minimizing method, then showed that the neural networks getting the better of the others. The structure
of neural networks was upgraded from [47] and [48], which utilized to identify a black-box system in
real-time updating parameters according to the stability of Lyapunov.
Chen et al. and Purwar et al. [51,52] utilized a model of neural networks for identifying ambiguous,
dynamic, nonlinear systems in discrete time. Purwar et al. [49] applied only one layer of neural networks
instead of multilayers and substituted that by distributing the input by Chebyshev-polynomials, and it
used the iterative least-squares technique with forgetting factor as real-time teaching algorithm to update
parameter. While, Chen et al. [50] adopted a multi-layer NNs (with a single hidden layer), and used a
developed algorithm derived from prediction error formulation.
Jafari et al., and Salahshoor and Jafari [51,52] produced an approach using the consecutive growth
and trimming radial basis function (RBF) to identify a nonlinear system online. It was a newly proposed
method for the growing and pruning neurons criteria to be proper for identifying online with modified

7
1st International Symposium on Engineering and Technology (ISETech) 2019 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 767 (2020) 012050 doi:10.1088/1757-899X/767/1/012050

teaching tactic based on unscented Kalman filter (UKF) for neural networks, the method implied using
a changeable forgetting factor technique embedded in UKF form to stimulate parameter assessment in
nonlinear dynamic systems.
Rocha Filho and de Oliveira Serra [53] presented an identification algorithm for a black-box
stochastic system in real-time built on neuro-fuzzy approach with impressive variables to model a
dynamic system. A continuous online neuro-fuzzy technique with Takagi-Sugeno developed structure,
which applies the maximum likelihood with iterative parameter estimation, was adopted as a procedure.
Copter with two degrees of freedom was used as a case study.
Aeronautics is a rich area for applying online system identification since it needs a quick reaction
and moving. By utilizing three layers unmanned airborne vehicle with fixed wings as a plant; Puttige
and Anavatti [54] found the differences and similarities between the real-time identification and off-line
identification using three layers of neural networks with a multiple input\output system that has a six-
output, where ARX was applied to minimize the error. The same previous strategy was implemented on
Eagle helicopter six degrees of freedom [55].
Feng et al. [56] determined a non-straight scientific model for the electro-water driven stage,
including the impact of Stribeck friction, where the friction dissipated power was determined from a
deliberate weight of pressure-driven chambers at different speeds extend. There are a few upgrades were
made on GA method, like the fitness, crossover, and mutation assumption. In real-time, a noticed speed
utilized to assess the actual velocity, and dynamic feed-forward remuneration for friction impact was
configured.

4. Conclusion
Identifying a dynamic system allows scientists and engineers to understand the nature of the system and,
as a result, gives them the ability to control that system or to gain full knowledge of the system and
improve it. Conventional linear analyses for identifying systems are developed to include some systems
with nonlinear behavior (using linearity assumption). The majority of the dynamic systems are not
linear, and implementing linearity presumptions lead to a false mathematical model. Determining a
precise nonlinear model is not a simple mission (to reach the genuine physical model), and evaluating
the parameters of that condition will be somewhat hard, so utilizing artificial intelligence can deal with
non-linearity and uncertainty issues. Estimating the value of parameters in the off-line analysis is a little
bit different from the on-line analysis, where the first leads to get fixed values, which have a considered
error in specific points, while the on-line analysis has flexible and adaptable parameters.

5. Future Work
Based on the previous, the soft-computing features could be employed in the system identification, either
off-line and online mode. The noted effectiveness in dealing with nonlinear systems makes the soft-
computing aspects be nominated as a perfect solver for the complexity of the system so that future work
will be using this approach from developing the system algorithm for gray-box, not only representing
the black-box system depending on the input-output data.

Reference
[1] Schoukens J, Pintelon R and Rolain Y 2012 Mastering system identification in 100 exercises (John
Wiley & Sons)
[2] Noël J-P, Esfahani A F, Kerschen G and Schoukens J 2017 A nonlinear state-space approach to
hysteresis identification Mech. Syst. Signal Process. 84 171–184
[3] Nelles O 2013 Nonlinear system identification: from classical approaches to neural networks and
fuzzy models (Springer Science & Business Media)
[4] Hunt K J, Irwin G R and Warwick K 2012 Neural network engineering in dynamic control systems
(Springer Science & Business Media)

8
1st International Symposium on Engineering and Technology (ISETech) 2019 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 767 (2020) 012050 doi:10.1088/1757-899X/767/1/012050

[5] Hong E Y, Meng T K and Chitre M 2013 Online system identification of the dynamics of an
autonomous underwater vehicle Underwater Technology Symposium (UT), 2013 IEEE
International (IEEE) pp 1–10
[6] Yang Q, Ren X, Zhao W, Lv Y and Wang S 2017 Optimal adaptive parameter estimation-based
tracking control of motor turntable servo system Modelling, Identification and Control (ICMIC),
2017 9th International Conference on (IEEE) pp 139–144
[7] Gensler A, Gruber T and Sick B 2015 Fast feature extraction for time series analysis using least-
squares approximations with orthogonal basis functions Temporal Representation and Reasoning
(TIME), 2015 22nd International Symposium on (IEEE) pp 29–37
[8] Amrani N, Serra-Sagristà J and Marcellin M 2017 Low complexity prediction model for coding
remote-sensing data with regression wavelet analysis Data Compression Conference (DCC), 2017
(IEEE) pp 112–121
[9] Worden K and Green P 2017 A machine learning approach to nonlinear modal analysis Mech. Syst.
Signal Process. 84 34–53
[10] Nechyba M C and Xu Y 1994 Neural network approach to control system identification with variable
activation functions Intelligent Control, 1994., Proceedings of the 1994 IEEE International
Symposium on (IEEE) pp 358–363
[11] Mehrotra K, Mohan C K and Ranka S 1997 Elements of artificial neural networks (MIT press)
[12] Daniel G 2013 Principles of artificial neural networks vol 7 (World Scientific)
[13] De la Sen M, Garrido A J, Barambones O and Maseda F J 2003 An expert network for obtaining
approximate discrete-time models for LTI systems under real sampling using parameter
identification Emerging Technologies and Factory Automation, 2003. Proceedings. ETFA’03.
IEEE Conference vol 1 (IEEE) pp 462–469
[14] Ogunmolu O, Gu X, Jiang S and Gans N 2016 Nonlinear systems identification using deep dynamic
neural networks ArXiv Prepr. ArXiv161001439
[15] de Canete J F, Gonzalez-Perez S and del Saz-Orozco P 2008 Artificial neural networks for
identification and control of a lab-scale distillation column using LABVIEW World Acad. Sci. Eng.
Technol. 47 64–69
[16] Shamsudin S S and Chen X 2012 Identification of an unmanned helicopter system using optimised
neural network structure Int. J. Model. Identif. Control 17 223–241
[17] Hanafi D, Tee K S, Johana E, Rahman H A, Rahmat M F, Wahid H and Ghazali R 2018 Neuro Model
for Passive Suspension of a Light Car Int. J. Integr. Eng. 10
[18] Mills P M and Zomaya A Y 1993 A neural network approach to on-line identification of nonlinear
systems Cybern. Syst. 24 171–195
[19] Attia A and Horáček P 2001 Modeling Nonlinear Systems by a Fuzzy Logic Neural Network Using
Genetic Algorithms 41 69–76
[20] Cong S, Li G and Feng X 2010 Parameters identification of nonlinear DC motor model using
compound evolution algorithms Proceedings of the World Congress on Engineering vol 1
[21] Herreros A, Baeyens E, Perán J R and Melgar A 2002 Parameter identification of a car suspension
system using non-intrusive signals IFAC Proc. Vol. 35 427–432
[22] Jizhen L, Junlin G, Yang H, Juan W and Hong L 2017 Dynamic modeling of wind turbine generation
system based on grey-box identification with genetic algorithm Control Conference (CCC), 2017
36th Chinese (IEEE) pp 2038–2042
[23] Cen Z, Wei J and Jiang R 2013 A gray-box neural network-based model identification and fault
estimation scheme for nonlinear dynamic systems Int. J. Neural Syst. 23 1350025
[24] Dubey B, Kataria S and Mohanty A 1997 Neural network fits to neutron induced reactions using
weighted least-mean-squares Nucl. Instrum. Methods Phys. Res. Sect. Accel. Spectrometers Detect.
Assoc. Equip. 397 426–439
[25] Kurita T 1994 Iterative weighted least squares algorithms for neural networks classifiers New Gener.
Comput. 12 375–394
[26] Narendra K S and Parthasarathy K 1992 Neural networks and dynamical systems Int. J. Approx.
Reason. 6 109–131

9
1st International Symposium on Engineering and Technology (ISETech) 2019 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 767 (2020) 012050 doi:10.1088/1757-899X/767/1/012050

[27] Quiroga J, Cartes D and Edrington C 2009 Neural network based system identification of a PMSM
under load fluctuation Dyna 76 273–282
[28] Sierra J E and Santos M 2018 Modelling engineering systems using analytical and neural techniques:
Hybridization Neurocomputing 271 70–83
[29] Ismeal G A, Kyslan K and Fedák V 2014 DC motor identification based on Recurrent Neural Networks
Mechatronics-Mechatronika (ME), 2014 16th International Conference on (IEEE) pp 701–705
[30] Rahim N, Taib M, Adom A and Halim M 2007 Nonlinear System Identification for a DC Motor using
NARMAX Model with Regularization Approach International Conference on Control, Instrument
and Mechatronics Engineering, CIM vol 7
[31] Lu H H, Rogers C, Goecks V G and Valasek J 2018 Online Near Real Time System Identification on
a Fixed-Wing Small Unmanned Air Vehicle 2018 AIAA Atmospheric Flight Mechanics Conference
p 0295
[32] Pislaru C and Shebani A 2014 Identification of Nonlinear Systems Using Radial Basis Function Neural
Network Int. J. Comput. Inf. Syst. Control Eng. 8 1528–1533
[33] Angelov P P and Filev D P 2004 An approach to online identification of Takagi-Sugeno fuzzy models
IEEE Trans. Syst. Man Cybern. Part B Cybern. 34 484–498
[34] Qian Y, Ou G, Maghareh A and Dyke S J 2014 Parametric identification of a servo-hydraulic actuator
for real-time hybrid simulation Mech. Syst. Signal Process. 48 260–273
[35] Al Seyab R and Cao Y 2008 Nonlinear system identification for predictive control using continuous
time recurrent neural networks and automatic differentiation J. Process Control 18 568–581
[36] Wang L, Zhang N and Du H 2015 Real-time identification of vehicle motion-modes using neural
networks Mech. Syst. Signal Process. 50 632–645
[37] Pairan M and Shamsudin S 2017 System identification of an unmanned quadcopter system using
MRAN neural IOP Conference Series: Materials Science and Engineering vol 270 (IOP
Publishing) p 012019
[38] Kirkpatrick K, May J and Valasek J 2013 Aircraft system identification using artificial neural networks
51st AIAA Aerospace Sciences Meeting including the New Horizons Forum and Aerospace
Exposition p 878
[39] Valasek J and Chen W 2003 Observer/Kalman filter identification for online system identification of
aircraft J. Guid. Control Dyn. 26 347–353
[40] Hussein A, Hirasawa K and Hu J 2003 Online Identification and Control of a PV-Supplied DC Motor
Using Universal Learning Networks. ESANN pp 173–178
[41] Rubaai A and Kotaru R 2000 Online identification and control of a DC motor using learning adaptation
of neural networks IEEE Trans. Ind. Appl. 36 935–942
[42] Griñó R, Cembrano G and Torras C 2000 Nonlinear system identification using additive dynamic
neural networks-two on-line approaches IEEE Trans. Circuits Syst. Fundam. Theory Appl. 47 150–
165
[43] Guo Y, Wang F and Lo J T-H 2017 Nonlinear system identification based on recurrent neural networks
with shared and specialized memories Control Conference (ASCC), 2017 11th Asian (IEEE) pp
2054–2059
[44] Ibrahim H E 2012 Adaptive control based on neural network system identification Proceedings of the
11th WSEAS international conference on Electronics, Hardware, Wireless and Optical
Communications, and proceedings of the 11th WSEAS international conference on Signal
Processing, Robotics and Automation, and proceedings of the 4th WSEAS international conference
on Nanotechnology (World Scientific and Engineering Academy and Society (WSEAS)) pp 84–
91
[45] Sadeghi M and Farrokhi M 2016 Online identification of non-linear dynamic systems by Wiener
model using subspace method and neural networks Trans. Inst. Meas. Control 0142331216663618
[46] PIVOŇKA P, VELEBA V and OŠMERA P 2007 Using of Neural Network Based Identification for
Short Sampling Period in Adaptive Control 11th WSEAS Int. Conf. Syst. Crete Isl. Greece 12–17
[47] Daachi B and Benallegue A 2003 A stable neural adaptive force controller for a hydraulic actuator
Proc. Inst. Mech. Eng. Part J. Syst. Control Eng. 217 303–310

10
1st International Symposium on Engineering and Technology (ISETech) 2019 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 767 (2020) 012050 doi:10.1088/1757-899X/767/1/012050

[48] Goedtel A, SILVA I N D and SERNI P J A 2005 Neural network based estimation of torque in
induction motors for real-time applications Electr. Power Compon. Syst. 33 363–387
[49] Purwar S, Kar I and Jha A 2007 Nonlinear system identification using neural networks IETE J. Res.
53 35–42
[50] Chen S, Billings S and Grant P 1990 Non-linear system identification using neural networks Int. J.
Control 51 1191–1214
[51] Jafari M R, Alizadeh T, Gholami M, Alizadeh A and Salahshoor K 2007 On-line identification of non-
linear systems using an adaptive RBF-based neural network Proceedings of the World Congress
on Engineering and Computer Science WCECS
[52] Salahshoor K and Jafari M R 2012 On-line identification of non-linear systems using adaptive RBF-
based neural networks Int. J. Inf. Sci. Manag. IJISM 5 99–121
[53] Rocha Filho O D and de Oliveira Serra G L 2016 Online identification based on instrumental variable
evolving neuro-fuzzy model for stochastic dynamic systems Fuzzy Systems (FUZZ-IEEE), 2016
IEEE International Conference on (IEEE) pp 9–16
[54] Puttige V R and Anavatti S G 2007 Comparison of Real-time Online and Offline Neural Network
Models for a UAV IEEE Int. Conf. Neural Netw. 412–417
[55] Samal M K, Anavatti S and Garratt M 2009 Real-time neural network based identification of a rotary-
wing UAV dynamics for autonomous flight Industrial Technology, 2009. ICIT 2009. IEEE
International Conference on (IEEE) pp 1–6
[56] Feng H, Qiao W, Yin C, Yu H and Cao D 2019 Identification and compensation of non-linear friction
for a electro-hydraulic system Mech. Mach. Theory 141 1–13

11