Proceedings of Sixth International Conference on Machine Learning Cybernetics, Hong Kong, 19-22 August 2007

RESEARCH ON FLUX OBSERVER BASED ON WAVELET NEURAL NETWORK ADJUSTED BY ANTCOLONY OPTIMIZATION
CHENG-ZHI CAO, XIAO-FENG GUO, WEN-JING WANG School of Information Science & Engineering, Shenyang University of Technology, Shenyang Liaoning 110023, China. E-MAIL: caochengzhi169@sohu.com, gxf_1979@hotmail.com, tinajing8153@163.com

Abstract:
To improve the performance of extra-low speed in direct torque control (DTC) system, this paper applies wavelet neural network (WNN) to constitute flux observer by deep researching nonlinear mathematic model of stator flux of asynchronous motor. Furthermore, in order to improve rapidity and real time characteristics of WNN flux observer, the paper applies ant colony algorithm (ACA) with embedded deterministic searching strategy to optimize dilation factor translation factor and output weight of WNN. The paper compares this method with wavelet neural network flux observer optimized by gradient descent algorithm. Simulation shows that the former not only can reduce the node numbers of hidden layers and quicken the convergence rate of WNN, but also can improve on-line identification precision of flux observer, so it can effectively improve low speed performance of DTC system.

characteristic, rapidity, availability. Because wavelet neural network has four merits: self-construction networks, partial retrieval of approximated function, fast convergence and escaping local minimum. The wavelet basis function can be used to realize approach of random function by fast discrete wavelet transform. In additional, ant colony algorithm is a novel bio-inspired optimization algorithm that has the ability to find better solution (the algorithm, which is a parallel algorithm in essence, not only speed up the evolution process by positive feedback, but also exchange and transfer pheromone carried through different agents. So many ants may collaborate each other in favor of discovering better solution) to improve convergence rate. 2. Flux observer theory

Keywords:
Ant colony algorithm; Wavelet neural network; Flux observer

The DTCs U-N model in the stationary reference frame ( ) can be described as follows:

1.

Introduction

ra =

[ L
Rr
sa

Rr

(
(

sa

sa

ra

]dt
r

(1)

The direct torque control (DTC) systems development in recent years makes torque and flux as control object, control asynchronous motor in the stator axis system. The DTC technology is a high-performance speed method with fast torque response, The optimal PWM gating signals are generated to control the inverter according to the output of the bang-bang regulator. It mainly depends on stator flux direction. So it is a critical issue to real time and exactly estimate stator flux to improve static and dynamic performance of extra-low speed [1] in direct torque control (DTC) system. Nowadays there are three methods (U-I,U-N,I-N) of observing stator flux, while they all have some questions, such as flux drift [2] and torque ripple. Comparing the above three methods, wavelet neural network flux observer based on ant colony algorithm proposed in this paper can improve dynamic performance of the basic DTC at low speed and have real time

[ L

) +
s

] dt (2)

(u (u

i sa ) dt

(3) (4)

i s ) dt

Where, s , r are stator and rotor flux, u s is the stator voltage. Rs , Rr respectively denote stator and rotor resistance, i s is the stator current, L is the leakage inductance, , respectively denote axis and axis, denotes the rotor speed. Stator flux can be expressed as:

s = f (us , is , ) = f (usa , us , isa , is , )

(5)

1-4244-0973-X/07/$25.00 2007 IEEE 862

Proceedings of Sixth International Conference on Machine Learning Cybernetics, Hong Kong, 19-22 August 2007 3. Wavelet neural network structure

Here, q is the number of the training samples,

Fil (t ) is the desired value of yi non-linear function.
l

The wavelet transform is the method with differentiable rate and good localization in both time and frequency space. It is widely applied to the analysis of time and frequency and the object observation by the inner products of a set of wavelets with the input signal. So we constitute a new mapping network combined wavelet and neural network, it is called wavelet neural network (WNN) [3]. Fig 1 shows the structure of wavelet neural network:

(t) . yil (t ) = f (uil ) , f

is a

4.

Wavelet neural network optimization

Gradient descent algorithm is a typical algorithm which may be used to optimize wavelet neural network. It is provided with vivid structure and little calculation dimensions. However, its rate of convergence is slower and liable to get into local minimum. Ant colony algorithm is a novel bio-inspired optimization algorithm based on colony. It is proposed at the light of achievement which is acquired by studying the collectivity of real ant colony in nature. It belongs to random searching algorithm. The ability to search the best solution or better solution and better robustness will be provided with ant colony algorithm. So ant colony algorithm can solve the defects of gradient descent algorithm to some degree. Consequently, an embedded ant colony algorithm [4], [5] is proposed in this paper. 4.1. Ant colony algorithm Ant colony algorithm is a efficient optimized algorithm which is achieved by profoundly and broad researching ant colony action in nature. In order to be convenient to understand ant colony algorithm principia, Traveling Salesman Problem, which is used to solve n cities on the same plane (0,1,,-1 is shown to numbers of cities), will be proposed to show the ant system model. Above all, marks as follows will be used: m is the ant numbers of ant colony, d ij (i, j = 1,2, n ) denotes the distance between city i and city j , bi is the ant numbers in city i . m =
n

not

Where, neural function of wavelet neural network is sigmoid function but wavelet basis function.

us , us , is , is , are five inputs to stator resistance

detector based on wavelet neural network. Fifteen nodes exist in hidden layer between input and output. Change of flux is output. Input must be normalized. The math model can be defined as follow:
M K t bk y i = f W ki X m ( t ) h m ( ) (6) ak m =1 k =1

Where, yi (i = 1,2, N ) is the output for the i th training vector, M is the node numbers of input layer, K is the node numbers of hidden layers, Wki is the weight between the k th node of hidden layer and i th node of output layer, X m (m = 1,2, M ) is the input for the m th training vector X (t ) , h is wavelet function, in which a k , bk are respectively denote dilation factor and translation factor, f is the nonlinear function. Network parameters such as a k , bk , Wki can be optimized by LSE (least square error) function.

b
i i =1

ij

; ij expresses the pheromone

quantity of the ij route. At the beginning of program, pheromone quantity on all routes is the same. Supposing that ij (o ) = C (cons tan t ) . Ant k (k = 1,2, m ) , which is in the movement, selects transfer direction according to the pheromone quantity on different path., p ij is probability which Ant k transfer from i to j ,
k

min E =

1 q N l l [ F i ( t ) y i ( t )] 2 l =1 i =1

(7)

863

Proceedings of Sixth International Conference on Machine Learning Cybernetics, Hong Kong, 19-22 August 2007
ij ( t ) ij ( t ) j allowedk = is ( t ) is ( t ) 0 otherwise

k ij

(8)

4.2. Wavelet neural network adjusted By embedded ant Colony algorithm First, design matrix X : X = [a, b, w] , a, b , w represent dilation factor ,translation factor and output weight. So it X that network amounts to optimizing parameters ak , bk , Wki are optimized. W ( X ) is Wavelet neural network target function. To be optimized Wavelet 1 . m ants neural network target function: MaxZ = W (X ) stochastically distribute in X . Each ant has a neighbor, r denotes its radius. Each ant proceeds search in its neighbor. After all ants accomplishing local search, individual starts global search by pheromone intensity and heuristic function. After a cycle, pheromone intensity will be updated. Ant colony algorithm is combined transfer regulation with pheromone intensity update regulation. Supposed that (i, j ) is pheromone intensity between position i and j , initializations about , i, j are all constants. (i, j ) is visibility from i to j that is confirmed by some heuristic algorithm. The move rule that ant transfers position r to position s is:

Where, , show pheromone quantity which ants accumulate during the movement and different function which heuristic factor acts on the process that ants select paths, ij denotes expectation degree which ants transfer from city i to city j , it is confirmed concretely by some heuristic algorithm. Allowedk = {0,1, n 1} tabuk is used to denote cities which Ant k can select. Manpower ant colony, which is greatly different from real ant colony in nature, is provided with memory function. Aggregate tabu k (k = 1,2, m ) is used to register cities which ant

k has been to. It will be adjusted dynamic with evolution progress. When time flies, pheromone, which is old gradually will wear away. 1 is used to show the degree of pheromone evaporation. If updated, ant completes a cycle. Pheromone quantity on different routes will be adjusted by the formula as follows:
k k [ij ]new=[ij ]old +(1 ) ij

(9) (10)

on the ij

ij

=
k =1

k ij

arg max [ ( i , j )] a [ ( i , j )] q < q 0 J (i ) (12) s= S otherwise

Where,
q denotes stochastic number, 0< q <1;

k ij is pheromone increment which ant k left behind

route during the current cycle. ij

is

selected based on(15), J (i ) denotes ant in position i get to the next possible position. Supposing that X ' is the new position, if the value in new position is smaller than originally target function. The new position will be reserved, otherwise leave it. Local search radius r will decrease according to iterative numbers. So:

q0 denotes coefficient, 0< q0 <1; S is the next position

pheromone increment of ij route during the current cycle.

k ij

= 0

cond 1
k

(11)

otherwise

cond1: if ant k can pass by ij . Q is constant,

Lk denotes length which ant k has walked during the

current cycle. Initialization of formula (10) is constant, and initialization of formula (11) is zero. According to different k algorithms, the expression of ij , ij and pij may be different. They will be confirmed respectively by different problems. So different ant colony algorithms are come into being.

X X = X'

w( X ) > w( X ' ) otherwise

(13)

Global search demonstrates that each ant chooses staying at origin, transfer to other ants neighbor or start global search after accomplishing a local search. Supposing Act (i ) is the action of the ant i . W avg is the average value of m ants target function. so:

global radom search(w( Xi ) < wavg) (q < q0 ) Act(i) = S, otherwise

(14)

864

Proceedings of Sixth International Conference on Machine Learning Cybernetics, Hong Kong, 19-22 August 2007 Where, 0< q <1,0< q0 <1, S can be selected by move rule as follow:

p ij =
d ij = w ( X i ) w ( X j )

( j) e

dij T dij T

(15)

and

( j) e

searching strategy basic thought will be listed as following: firstly, initializing parameters, which make each ant be at the beginning of loop. Then, let each ant go on local search utilizing formula (18) and global search using formula (14) (15) (16). After that, the pheromone intensity will be updated with formula (17). If the ceasing condition is meet (otherwise repeat all), optimal solution or secondary solution will be output. 5. Simulation experiment

formula(12) ensures that ant i moves according to pij to ant j neighbor which has bigger target function value. Coefficient T decides the probability function slope. When ants move to origin having high pheromone intensity, they probably find new food resource at random location of transfer path, this case is defined as directional stochastic transfer. Formula ant i transfers to ant j neighborhood can be given by:

{( d ij

< 0 ) ( i j ) ( d ij = 0 ) i = j

} {

The

In the paper choosing MATLAB6.1 simulation means SIMULINK to carry out simulation to the direct torque control system of asynchronous motor. Simulation parameter as follows: p =1.7kw, ne =1440r/min, R1 =4.25,

R2

=3.24,

L1
2

=0.666H,

L2

=0.671H,

Lm =0.651H, J =0.02N.m , m =80, v0 =0.1, q0 =0.2, T =10, 0 =0.6, =0.7, cycle index:120, Mexico hat will
be regarded as generating function of wavelet.

od value C < 0 X j 's neighborho is random Xi = (16) X j + (1) Xi ,otherwise

Where, 0< <1, 0< 0 <1, 0< <1. Global search goes on in term of random mode and update pheromone intensity. If there is m ants moving to j , then:

( j ) = ( j ) + (1 )

(i )
i =1

(17)

Where, 0< <1: forgetting factor. In order to improve the defect of stochastic algorithm, deterministic searching strategy is embedded in ant colony algorithm. We adopt step length acceleration of direct method including two portions: exploring search and model transfer. The variation rule of target function can be obtained by exploring along every coordinate direction, then confirm search direction and transfer along it , repeating above two steps, until exploring appropriate step. Algorithmic local search rule:
local deterministic search using step length accelerationv < v 0 R= local random search based on formula (13)otherwise

Figure.2 flux caterpillar based on WNN optimized by ant colony algorithm

(18)

Where:0< v <1, 0< v0 <1. Ant colony algorithm with embedded deterministic Figure.3 flux caterpillar based on WNN optimized by gradient descend

865

Proceedings of Sixth International Conference on Machine Learning Cybernetics, Hong Kong, 19-22 August 2007 As can be seen in figure 2 and figure3, the direct torque control system fluxs running track based on wavelet neural network flux observer optimized by gradient descend has excursion, the flux has grave aberrance. But the joined of wavelet neural network flux observer optimized by ant colony algorithm makes the flux track steadiness. Stator flux track is similar to rotundity. The outcome of simulation shows that wavelet neural network based on ant colony algorithm has the faster convergence rate and the better approximation capability and the better on-line identification precision than flux observer based on wavelet neural network optimized by gradient descend. So utilizing ant colony algorithm to optimize wavelet neural network can improve the control performances. 6. Conclusions scientific research project of the Education Department of LiaoNing Province (20206331, 05L288). References [1] DONG Lei,LIYongdong, Vector control of a speed sensor-less induction motor at very low speed, Department of Electrical Engineering, Tsinghua University, Beijing, Vol. 98, No. 5, pp. 573-576, 2003. [2] D.Casadei,G.Serra, and A.Tani, Analytical investigation of torqueand flux ripple in DTC schemes for induction motors, in Proc. IEEEIECON97, vol. 2, pp. 552-556, 1997. [3] Caochengzhi, Lumuping, Wangnan, Wangxin, Parameters of Identification on DTC System Based on Wavelet Fuzzy Neural Networks, Transaction of electro-technical society, Vol.19, pp.18-22, 2004. [4] YANG Yong, SONG Xiao-feng, WANG Jian-fei , HU Shang-xu, Ant colony algorithm for continuous space optimization, Control and Decision, Vol. 98, No.5, pp. 573-576, 2003. [5] J.Maes and J. A. Melkebeek, Speed-sensorless direct torque control of induction motors using an adaptive flux observer, IEEE Transactions on Industry Application, Vol. 36, No. 3, pp. 778-785, May/June 2000.

Wavelet neural network optimized by ant colony algorithm embedded deterministic searching strategy, which has faster convergent rapidity and can reduce stator flux drift and make the overall control system steady. Adopting this method to observe stator flux can improve static and dynamic performance of extra-low speed in direct torque control (DTC) system. In the stator flux observer technique it is superior to other ways. Acknowledgements This research work is supported by the High School

866