The NARMAX Model for a DC Motor using MLP Neural Network

N.A. Rahim1 M.N. Taib2 A.H. Adom1 M.Y. Mashor1

1
Electronic & Biomedical Intelligent Systems Group (EBItS), School of Mechatronic Engineering
Kolej Universiti Kejuruteraan Utara Malaysia, 02600 Jejawi, Arau, Perlis, MALAYSIA
E-mail: [email protected]
2
Advanced Signal Processing (ASP) Research Group
Faculty of Electrical Engineering, Universiti Teknologi MARA, 40450 Shah Alam, Selangor Darul Ehsan
E-mail: [email protected]

Abstract input-output relationship as a nonlinear difference equation

of the form (see Equation (1)) [6-8].
The approach to the design of direct current (DC) motor
varies considerably using advanced methods such as yˆ (t ) = F [ y (t − 1),..., y (t − n y ), u (t − 1),..., u (t − n u ),
artificial intelligence (AI). However, accuracy issues ε (t − 1),..., ε (t − nε )] (1)
cannot be totally achieved using conventional methods.
This paper presents the study on nonlinear autoregressive
where, F is a nonlinear Function, y (t ) , u (t ) and ε (t ) are
moving average with exogenous input (NARMAX) model
using multilayer perceptron (MLP) neural network for DC output, input and noise signal respectively and n y , nu and
motor modeling. The simulation results show that MLP nε are their associated maximum lags.
neural network can fit the NARMAX model and identify the
DC motor drive system. 3. MLP Neural Network Approach
Keywords Artificial neural networks (ANN) have become a very
useful tool in a wide range of research area. It offers the
NARMAX, MLP, OSA, NSSE, LMA
advantage of a simple nonlinear modeling through learning
via a highly parallel and distributed processing paradigm
1. Introduction [9]. Neural networks were developed using the principles of
Recently, system identification has attracted a lot of operation of the human brain. It also inherits the
attention from the research community. Its application has intelligence and nonlinear characteristic of the brain. In this
grown beyond control system theory into a wider sphere work, the MLP neural network was employed for the
including biomedical, engineering, financial and operation NARMAX parameters’ optimization.
research.
ANN consists of interconnected network of elementary
This paper presents the NARMAX model development unit, called neurons, which are arranged in layers and
using MLP neural network applied to DC motor drive operate in parallel. Each neuron in the network operates by
system modeling. Usually DC motors can be used to taking the sum of its weighted inputs and passing the result
control speed, torque and position. In order to come up with through a nonlinear activation function. A feed forward
a robust and efficient motor control, various linear and network is made up of an input and output layer, having
nonlinear techniques were exploited to model the DC one or more layers of neurons in between called the hidden
motor. Examples of linear techniques include the layers, which propagate the inputs forward to the output
autoregressive moving average (ARMA) and autoregressive layer. The general structure of a multilayer feed forward
moving average with exogenous input (ARMAX) models; network is shown in Figure 1.
while nonlinear techniques may apply input output
linearization, artificial neural network (ANN) and feedback
linearization [1-5]. Many of these techniques were
successful to certain extent but their accuracy decreases as
the degree of nonlinearities increases.

2. NARMAX Model
Accurate model is required in order to have a successfully
system identification. Leontaritis and Billings [6] proposed
a NARMAX structure as a general parametric form for
nonlinear systems. This structure describes both the
stochastic and deterministic component of a system. The
NARMAX model is a generalization of the linear ARMAX
model and the SISO case. NARMAX structures models the
Figure 1 - MLP Architecture
The overall equation for figure above can be derives as: choosing to determine the proper lag space structure before
identifying a model of a dynamic system without estimating
⎛ q
⎞
yˆ i (w , W ) = Fi ⎜⎜ ∑ Wij h j (w ) + Wi 0 ⎟⎟ (2) the values of the parameters of the model. Unfortunately
⎝ j =1 ⎠ the Lipschitz method is very sensitive to the noise in the
data [16].
Only one network with one hidden layer are considered in
the present study because the results of Billings et al. [11] The main idea of the method is to estimate the maximum
and Taib et al. [12] show that this is sufficient to gradient of an unknown mapping from the given input-
approximate all continuous functions. The activation output data or in other words it can estimate the smoothness
function for output layer is selected to be linear and for an unknown mapping [17]. While smooth mapping (less
hidden layer is selected to be hyperbolic tangent. maximal gradient) is a desirable feature of a model, it can
be determine the order the model to create the smoothness
Previously, several variants of ANN were employed for DC mapping between the input and the output data.
motor modeling and control. Minkova et al. [10] used a
feed-forward ANN and Mohamed et al. [9] used a NARX 5. Modeling Process
model via a three layer feed-forward network. These efforts
attested to the viability of using ANN for motor modeling For the modeling process (see Figure 2), input-output data
and control. for a DC motor drive system were collected. Then, the
Lipschitz analysis was performed on the data to gauge the
3.1 Levenberg-Marquardt Algorithm (LMA) maximum lag of the data. Subsequently, NARMAX model
A Levenberg-Marquardt algorithm (LMA) is the standard structure and fitting using MLP network was carried out.
method for minimization of mean-square error criterion, The process is adjusted until the prediction output satisfied
due to its rapid convergence properties and robustness. A a model validation test.
version of this method described in Fletcher [13], Ljung
[14] and Nørgaard et al. [15]. Data Collection
This work uses prediction error approach, which is based
on the introduction of a measure of closeness in terms of
mean square error criterion as defined as: Selecting Model
Structure
( )
N

∑ ε(t , θ)ε (t , θ)
1 (3)
V N θ, Z N = T

2N t =1

where N is a number of training and ε t , θ is a difference ( ) Model Fitting

between desired and actual output, defined as:
ε(t , θ) = y (t ) − yˆ (t , θ) (4)
Model Validation
This algorithm is as follow:
1) Select an initial parameter vector θ (0 ) and initial
value λ(0 ) .
2) Determine the search direction from No
[R (θ (i ) ) + λ(i ) I ] f (i ) = −G (θ (i ) ) Model is
OK
λ(i ) (if predicted decrease is close to
3) r (i ) 〈 0.75 ⇒ λ(i ) =
2
actual decrease let the search direction approach the Yes
Gauss-Newton search direction while increasing the
step size). Prediction Output
4) r (i ) 〈 0.25 ⇒ λ(i ) = 2λ(i ) (if predicted decrease is far from
Figure 2 - Modeling Process
the actual decrease let the search direction approaches
the gradient direction while decreasing the step size).
5) If V N (θ ( i ) + f ( i ) , Z N )〈V N (θ ( i ) , Z N ) then accept 6. Model Validation
θ (i +1) = θ (i ) + f (i ) as a new iterate and let λ(i +1) = λ(i ) Three methods were employed to test model validity
and i = i + 1 . namely: one-step-ahead (OSA) prediction, correlation test
and normalize sum squared error (NSSE).
6) If the stopping criterion is not satisfied go to 2)
OSA prediction is based on the past actual output rather
4. Lipschitz Analysis than the past prediction output. OSA is given as:
One of the principle goals of nonlinear system yˆ (t ) = fˆ ( y (t − 1),..., y (t − n y ), u (t − 1),..., u (t − nu ),
identification is to determining the structure of an unknown
system from input and output data. Lipschitz method is ε (t − 1, θˆ),..., ε (t − n , θˆ); θˆ)
ε
(5)
where fˆ (.) is the estimate of F (.) , ε (t , θˆ) is the prediction y(t ) = [ y(t − 1), y(t − 2), u (t − 1),.u(t − 2) , ε (t − 1), ε (t − 2)] (14)
error or residual. Table 1 show the parameter used for fit the NARMAX
Correlation tests lie within 95% confidence limit, defined model using MLP neural network approach.
as ± 1.95 / N , where N is a number of data. Residual
Correlation test is given as: Table 1 - Parameter for MLP neural network training

θ εε (τ ) = E[ε (t − τ )ε (t )] = δ (τ ) (6) Parameter Value

nu = 2 , n y = 2 and
θ ε 2ε 2 (τ ) = E [ (ε 2 (t ) − ε 2 (t ))(ε 2 (t − τ) − ε 2 (t )) ] = δ(τ) (7) Lag Space for Input, Output and
Residual nε = 2
θ yε (τ ) = E[ y (t − τ )ε (t )] = 0, ∀τ (8) Hidden Layer Neuron 2

θ y ε (τ ) = E[( y 2 (t − τ ) − y 2 (t ))ε (t )] = 0, ∀τ
Maximum Iteration (Epochs) 500
(9)
λ =1
2
Lambda
θ y ε (τ ) = E[( y 2 (t − τ ) − y 2 (t ))ε 2 (t )] = 0, ∀τ
2 2
(10) Stopping Criterion 1 × 10 −5

θε (εy ) (τ ) = E [(ε (t )ε (t − 1 − τ ) y (t − 1 − τ ))] = 0,τ ≥ 0 (11) The predicted output after training (Figure 4) show the

[( )]
NARMAX approach can identify the data. The predicted
2 )(
θ ( yε )ε (τ ) = E yε (t ) − yε (t ) ε 2 (t − τ ) − ε 2 (t ) = k ε δ (τ ) (12) and actual outputs do not differ significantly. The OSA
prediction model shows high fitting accuracy of 98.70%.
θ ( yε )u 2 (τ ) = E [(yε (t ) − yε (t ))(u (t − τ ) − u (t ))] = 0, ∀τ
2 2 (13) The validation data (Figure 5) show the NARMAX
approach can also identify the data. The predicted and
ε (τ ) is a residual between the measured output and actual outputs do not differ significantly. It proves that,
prediction output. although the validation data is not same like the training
data the system can identify the model.
Meanwhile, NSSE is produced at each training steps and
can be calculated as equation (3). Figures 6 show the correlation tests. It can be observed that
all tests falls within the 95% confidence limit.
7. Results and Discussions The NSSE iteration plot at Figure 7 demonstrates the speed
Four thousand input-output data pairs were used in of convergence for the network training using the
modeling the DC motor. The data were divided into two prediction error method. This show that the MLP neural
sets, one for training and remaining for validation. The network is able to convergence at a high speed and the
NNSYSID tool [18] was use to analyze the input-output predicted output can be accepted.
data.
The Lipschitz analysis in Figure 3 shows that the maximum
lag space required is two. Hence, it is not unreasonable to
assume that the system can be modeled by a second order
since the slope of the curve decreases for the model orders
>= 2 (refer to figure 2).This information was used to define
the NARMAX structure (see Equation (14)) for
identification using the MLP network. Then the MLP
network was trained using LMA and a network with two
hidden layer neurons was found to be sufficient.

Figure 4 - Prediction output with residual for training data

Figure 3 - Lipschitz method for determine lag space

 8. Conclusion
The work presented in this paper has shown that a DC
motor drive system can be successfully modeled using the
NARMAX model with parameter fittings employed by
MLP neural network. The validity of the model was
illustrated by OSA prediction model, correlation test and
NSSE iteration.
The performance of the model can be improved by further
investigations on MLP neural network parameters.

References
[1] K. Sedghisigarchi, A. Hasanovic, A. Feliachi and A.
Davari, Evaluation of Three Algorithm for Nonlinear
Control of a DC Shunt Motor, IEEE Proceeding of
the 33rd Southeastern Symposium on System Theory,
March 2001, pp. 407-411.
[2] J. Chiasson and M. Bodson, Nonlinear Control of a
Shunt DC Motor, IEEE Transactions on Automatic
Figure 5 - Prediction output with residual validation data Control, Vol. 38, No. 11, November 1993, pp. 1662-
1666.
[3] M.D. Minkova, D. Minkov, J.L. Rodgerson and R.G.
Harley, Adaptive Neural Speed Controller of a DC
Motor, Electric Power Systems Research, 1998, pp.
123-132.
[4] S. Mehta and J. Chiasson, Nonlinear Control of a
Series DC Motor: Theory and Experiment, IEEE
Transactions on Industrial Electronics, Vol. 45, No.
1, February 1998, pp. 134-141.
[5] J. Zhou, Y. Wang and R. Zhou, Global Speed
Control of Separately Excited DC Motor, IEEE
Power Engineering Society Winter Meeting, Vol. 3,
2001, pp. 1425-1430.
[6] I.J. Leontaritis and S.A. Billings, Input-Output
Parametric Models for Nonlinear Systems Part I:
deterministic nonlinear systems, International journal
of Control, 1985, 41(2), pp. 303-328.
[7] J.J. Liu, I.C. Kung and H.C. Chao, Speed Estimation
of Induction Motor Using a Nonlinear Identification
Technique, Proceedings Natl. Dci. Counc. ROC (A),
Vol. 25, No. 2, 2001, pp. 107-114.
[8] N. Chiras, C. Evans and D. Rees, Nonlinear Gas
Figure 6 - Correlation test Turbine Modeling Using NARMAX Sructures, IEEE
Transaction on Instrumentation and Measurement,
Vol. 50, No. 4, 4 August 2001, pp. 893-898.
[9] H.A.F. Mohamed, S. Yaacob and M.N. Taib,
Induction Motor Identification using Artificial Neural
Network, Australasian Universities Power
Engineering Conference, October 1997, pp. 217-221.
[10] M.D. Minkova, D. Minkov, J.L. Rodgerson and R.G.
Harley, Adaptive Neural Speed Controller of a DC
Motor, Electric Power Systems Research, Vol. 47,
1998, pp. 123-132.
[11] S.A. Billings, H.B. Jamaluddin and Chen, S., A
Comparison of the Backpropagation and Recursive
Prediction Error Algorithm for Training Neural
Networks, Mechanical Systems and Signal
Processing, 1991,5(3), pp. 233-255.
Figure 7 - NSSE Iteration
[12] M.N. Taib, R. Andres and R. Narayanaswamy, [16] N. Chiras, Linear and Nonlinear Modeling of Gas
Extending the Response Range of an Optical Fibre Turbine Engines, Thesis Dissertation, PhD, The
pH Sensor using an Artificial Neural Network, University of Glamorgan, 2002.
Analytica Chimica Acta 330, 1996, pp. 31-40. [17] X. He and H. Asada, A New Method for Identifying
[13] R. Fletcher, Practical Methods of Optimization, Orders of Input-Output Models for Nonlinear
Wiley, 1987. Dynamic Systems, Proceeding of the American
[14] L. Ljung, System Identification- Theory for the User, Control Conference, San Francisco, California, June
Prentice Hall, Eaglewood Cliffs, New Jersey, 1987 1993,pp. 2520-2523.
[15] M. Nørgaard, O, Ravn, N.K., Poulsen, and L.K., [18] M. Nørgaard, Neural Network Based System
Hansen, Neural Network for Modelling and Control Identification Toolbox, Technical Report, 00-E-891,
of Dynamic Systems, Springer-Verlag, London, 2000. Department of Automation, Technical University of
Denmark, 2000.