Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Abstract: This article incorporates the wavelets functioned neural network (wavenet) and auto-regressive
with exogenous input (ARX) approaches in a parallel combination for nonlinear system identification.
Wavenet composition has recently used identification of nonlinear systems with agreeable results. ARX
structure also well known base for linear system identification in least squares sense. In this work; the
power of nonlinear identification ability of wavenet and the linear identification ability of ARX structure
are merged to have better results to govern identification problem. The new model is tested and compared
for some benchmark problems and then it seen that the resulting model approximates better than the other
well known nonlinear identification models.
optimization method and initial parameters (Haykin 1998). So
1. INTRODUCTION
there is no guarantee about the trained system that is the
System identification and modeling is an essential and correct system model.
important subject for controlling the systems without human
In this study, by considering these drawbacks of the neural
intervention. A mathematical model of the system or artificial
networks, first it is desired to improve the approximation
intelligent model which has same input-output characteristic
capability of neurons by using wavelet functions. Second
with model is necessary to analyze and control the system.
identification ability is also increased with parallel ARX
Modeling which is based on physical laws forms a
structure. Thus combination of the models may have better
mathematical model for the system. But for the identification,
identification ability.
there is no need to use previous knowledge and physical
structure of the system, so they are known as the black-box 2. WAVENET-ARX BASED SYSTEM IDENTIFICATION
identification process (Ljung 1999). If the behavior of the
system is much complicated, it is a big troublesome for the
identification. Because of wide variety of the systems, there is 2.1 ARX Model
not a universal solution to identification of systems. Due to
ARX is one of simpler time series models where the output is
nonlinear behavior, in the beginning of the identification task,
expressed as a linear function of past input and output values.
selection of the identification method is usually the most
difficult part. Appropriate methods can be chosen according
to system behavior and desired goal of identification. A( q ) y (t ) B (q )u (t nk ) e(t ) (1)
Therefore, it is necessary to use enhanced identification
methods to handle the model that approximates the system Here e(t) is the ‘change’ component used to account for
correctly. uncertainty in the relationship and assumed to be identically
distributed with zero mean and finite variance, 2 [Thum
After the back propagation algorithm developed, neural M.T. 1999]. A and B are polynomials of order na and nb
networks are utilized in many problems as classification, respectively and nk is the time delay in the input.
function approximation, control etc. Even though its power
proved in many areas, neural networks are enhanced with new
theories everyday and compared with the other fields. The A(q) 1 a1q 1 ... ana q na
main advantage of the neural network is its capability for the (2)
B(q) b1 b2 q 1 ... bnb q nb 1
self learning from the previously collected input-output data.
The capability of its neuron to make approximation has
In this approach, the model parameters are estimated by the
caused the neural networks used to provide mapping between
least-squares method (Ljung 1999).
input-output data (Nazaruddin et al. 2006). However
identification with neural networks has some difficulties. First
there is not an explicit mathematical model of the system 2.2 Wavelet Function
handled from trained neural network and it is difficult to
analyze. Secondly, in training part there can be many different
Wavelet functions are important for the modeling wavelet
resulting alterable neural network system models for the same
networks because of their different basis characteristics. A
data with same input-output characteristic according to
wavelet function is usually called mother wavelet and its yˆ (k ) p(1) yˆ1 (k ) p(2) yˆ 2 (k ) (4)
different dilation and translation parameterized functions are
called daughter wavelets (Burrus et al. 1998). By updating the The basic identification scheme with wavenet-ARX model is
translation and scaling parameters different daughter wavelets shown in figure 3. In the algorithm the model parameters are
are obtained in the algorithms. Some of the wavelet functions updated through minimisation of the mean square error
are Gaussian function derivatives. In this work Polynomials (MSE). Neurons (wavelet windows) with wavelet parameters
WindOwed with Gaussian (Polywog) wavelet function is used are gradually changed to cover the time-frequency region
in wavenet as used before (Lekutai G. et al. 1997) (Edison et occupied by the signal to be identified (Narendra et al. 1990).
al. 2004), (Nazaruddin et al. 2006) and its function is shown
below.
Plant
0.8
The jabobians of the all parameters are formulated in
0.6
equations 9-13 as follows.
0.4
J N
yˆ1 (k )
0.2
(10)
p1 k 1 0
-0.2
J N
yˆ 2 (k )
-0.4
(11)
p2 k 1 -0.6
-0.8
J N
yˆ1 (k ) * h( ) (12) -1
0 50 100 150 200 250
w j k 1
J N
p(2) * y (k j ) (16) -5
d j k 1
0 20 40 60 80 100 120 140 160 180 200
Wavenet Identification
5
The parameters vector is
0
P [ p1 , p2 , c1 , c2 , d1 , d 2 , w1 ,.., wn , a1 ..an , b1 ...bn ]T (17)
where n is the number of the neurons. -5
0 20 40 60 80 100 120 140 160 180 200
Wavenet-ARX Identification
5
3. SIMULATION RESULTS
0
In simulations one nonlinear model and two benchmark
problems are used to identify and compare proposed model -5
with other methods. In the simulations for the faith 0 20 40 60 80 100 120 140 160 180 200
comparison neuron numbers or similar structures are used. In
results wavenet model, neural network model, wavenet-ARX Figure 5. Output Comparison for the first test
model and fuzzy-ARMAX models are compared. This fuzzy-
ARMAX model also has parallel structure with proposed
model (Alcı et al. 2007). In the end, there is seen that the
proposed model has effective resulting MSE values. In all Table 1. The resulting MSE values for the fist test
simulations the models are constructed with two inputs one
output system. Inputs are original input signal and other input Model MSE (Test)
signal is the one delayed term of the output signal.
Neural Network
3.228
3.1 Simulation Results-1 (One-Hidden-Layer-5 Neurons)
Wavenet
In the first simulation a nonlinear static mapping test model is 0.455
(One hidden-layer -5 Neurons)
chosen [Alcı, 2008]. This system has a linear input and the
output is the triangular plus sine wave. The input-output data Wavenet-ARX
0.301
is shown in figure 4. (One hidden-layer -5 Neurons)
3.2 Simulation Results-2 y(k ) [ y(k 1) y(k 2) y(k 3)u (k 2)...
Here, Box-Jenkins furnace (Chung et al. 2006) identification *(( y (k 3) 1) 0.5) u(k )]/(1 y(k 2)2 y(k 3)2 )
is used. First 200 input-output data is used for the training the (19)
system and last 90 input-output data are used for the testing In this simulation totally 1000 samples are used. First 600
and the resulting comparison is shown in Table 2. samples for training the system and last 400 samples are for
the testing. The resulting comparison is shown in Table 3 and
simulations are shown in Fig. 6.
ARX Model Estimation
5
0 ARX Model Estimation
2
-5
0 50 100 150 200 250 300 0
Neural-Network Identification
5 -2
0 100 200 300 400 500 600 700 800 900 1000
0 Neural-Network Identification
2
-5
0 50 100 150 200 250 300 0
Wavenet Identification
5 -2
0 100 200 300 400 500 600 700 800 900 1000
0 Wavenet Identification
2
-5
0 50 100 150 200 250 300 0
Wavenet-ARX Identification
5 -2
0 100 200 300 400 500 600 700 800 900 1000
0 Wavenet-ARX Identification
2
-5
0 50 100 150 200 250 300 0
-2
0 100 200 300 400 500 600 700 800 900 1000
Figure 5. Box-Jenkins Furnace Identification