APPLICATION OF AN EXTENDED KALMAN FILTER TO PARAMETER IDENTIFICATION OF A N INDUCTION MOTOR

TETSUYA IWASAKI and TERUO KATAOKA, Senior Member, I E E E Tokyo Institute of Technology Department of Electrical and Electronic Engineering 0-okayama Meguro, Tokyo Japan 152
Abstract
Identification of the parameters of an induction motor is very important in point of the performance prediction, simulation analysis and control application. In this paper an extended Kalman filter is employed to identify the parameters using measurements of the stator voltages and currents and the rotor speed of the induction motor. A model of the induction motor in the state space and the Kalman filter algorithm are shown. This filter is applied to the parameter identification of an inverter fed induction motor. A simple and practical method of setting the covariance matrices of the noises, which are important in the Kalman filter algorithm, is proposed. The starting values of the state and parameter vectors as well as the covariance matrix of the estimation error are then be shown, and finally the results of parameter identification arc shown. The results demonstrate that the filter is capable of identifying the parameters.

INTRODUCTION

As a result of the technological developments in semiconductors, especially power transistors and GTO thyristors, ac adjustable speed motors fed by inverters have become widespread. The induction motor is one of the most important ac machines used because of its low cost and high reliability. Parameter identification of an induction motor is very important with respect to the performance prediction, simulation analysis and control application. For the performance prediction and simulation analysis of an induction motor, the parameters are usually identified through n d o a d tests and locked rotor tests. The Kalman filter approach to the parameter identification can be used for the purpcse of knowing the actual values of the parameters of the induction motor under load.

As for the control applications of an induction motor, field oriented control is receiving much attention. Parameter identification is essential in order to keep the advantage of field orientation, i.e., almost instantaneous torque response by means of the decoupling between the torque and flux components of stator current. A number of methods of estimating the parameters of an induction motor for the purpose of control application have been proposed. One of them [l] is utilizing an extended Kalman filter. In reference [l],only the rotor time constant is estimated using the filter with a view of controlling an induction motor. In the present paper, an extended Kalman filter is employed to identify various parameters of an induction motor simultaneously using measurements of the stator voltages and currents and the rotor speed of the induction motor. The parameters identified here are rotor time constant, total leakage inductance, magnetizing inductance and stator resistance. This filter is applied to the parameter identification of an inverter fed induction motor. It is necessary to set the covariance matrices of noises in the Kalman filter algorithm. In this paper only the pure disturbances are considered as noises, while the harmonic contents of thc input voltages were regarded as noises in reference [l]. In general, it is difficult to know the exact values of the covariance matrices of noises because every measurement usually has noises, and noises themselves cannot be measured. A simple method t o determine the covariance matrices of pure noises proposed in the present paper gives a good result.

MACHINE MODEL

The matrix equation of the induction motor in the stator B coordinates is given by fixed cy - ,

89CH2792-0/89/0000-0248$01 .OO 0 1989 IEEE

where 2 = A,x M = Lo2/LR where

+ B,u

(5)

LI = Ls - M
T

= LR/RR

In,

x=

'8'

A ,

,state vector; u =

[ ]

,input vector

(6)

a-axis stator voltage up, : @-axis stator voltage in, : a-axis stator current ip, : P-axis stator current A , : w a d s rotor flux linkage times Lo/LR A , : p-axis rotor flux linkage times Lo/LR Rs : stator resistance R R : rotor resistance Ls : stator self inductance LR : rotor self inductance Lo : mutual inductance LI : total leakage inductance M : magnetizing inductance r : rotor time constant W R : rotor electrical angular frequency p : time differential operator.
U , ,

- A ,

(7)

The equivalent circuit of the induction motor in the steady state with sinusoidal symmetrical three-phase supply voltages corresponding to equation( 1) is depicted in Figure.1. Note that the parameters contained in equation(1) are clearly indicated in this equivalent circuit. The parameters in this circuit are the ones to be identified in the present paper.

In the state equation ( 5 ) the rotor angular frequency is a variable, but it can be regarded as constant because the mechanical system response is much slower than that of the electrical system. Then the system matrix A, is constant and the state transition matrix can be expressed as matrix exponential function of A,. The following discrete model of the induction motor is derived using the zerc-order-hold for the input for one sampling period.

where Ad = exp(&A)

RS

L1

Bd =

[LA

exP(Act)dt]Bc

A : sampling period

M1
T S

where only the first and second order terms of A, is considered, which is necessary and sufficient to maintain accuracy. Subscript k as in xk means the discrete time t = kA.

Figure.1 Eguivalent Circuit of the Induction Motor

KALMAN FILTER

In order to utilize the Kalman filter, it is necessary to transform the o - p model given in equation(1) to the state space model. Selecting the stator currents and rotor flux linkages for the state variables since the time derivatives of them are inclyded in the o - B model, the state space model can easily be obtained as
249

System Model

The discrete Kalman filter is applied to the discrete model of the induction motor in the state space. Because the number of the parameters to be identified is four, a forth order parameter vector is defined a s

filter can be applied.

where superscript T means the transpose of the vector or matrix. These parameters are all that are needed t o determine the equivalent circuit depicted in Figure.1. Considering the state vector, which has four components, the following eighth order vector should be estimated.

Figure.2 System Model for the Kalman Filter In the real system, the system noise should be taken into account. This noise is generally treated as being located inside the system.[2] In this paper, however, the noise superimposed on the input voltage is assumed to be the system noise. This assumption makes it easier to determine the ccvariance matrix of the system noise. T h e system model is

Filtering Algorithm The-principle of the Kalman filter is as follows. At time stagek, the predictions of thestate vector and the parameter k and the estimation vector are calculated from the input u k assuming that there is no noise. The of the state vector x covariance matrix of the prediction error Mk+l, that of the estimation error P k + 1 and M m a n gain matrix Kk+l are updated. At time stage k + 1, the estimations of the state vector and the parameter vector are obtained by revising the predictions by the difference between the measurements of the stator currents and the predictions via the I<alman gain matrix, which is determined so that the covariance matrix of the estimation error is the smallest value. T h e Kalman filter algorithm is summarized as follows.

where

and wk is the system noise and assumed to be a zercmean white Gaussian noise which is independent of Xk, with covariance Q. Let the available discrete time measurements be modeled

3. Fk= 1

Ad(6)li) A(Ad(6k)kk -k Bd(6k)Uk) 0 I FkPkF:

4. Mk+i 5.

+ Gk$Gf +

as

= Mk+lHT(HMk+lHT R)-'

where

H=

1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0

(17)

9. go t o 2. ir : estimation of the state vector % : prediction of the state vector 6 : estimation of the parameter vector I : unit matrix
0 : zero matrix M : covariance matrix of prediction error
250

and the measurement noise vk is a zeremean white Gausk and wk, with covariance sian noise that is independent of x R. Equation(l7) means that the stator currents are measured. T h e system model is shown in Figure.2. Now that the complete system model i s obtained, an extended Kalman

P : covariance matrix of estimation error K : Kalman gain matrix ( = [ LT N T L : Kalman gain matrix for the state vector N : Kalman gain matrix for the parameter vector

1')

The Covariance M a t r i c e s of Noises

The starting values c the estimations of the state and parameter vectors and the covariance matrix of the estimation error, and the covariance matrices of the system and measurement noises should be set before filtering. Figure.3 shows the structure of the Kalman filter. 2 - l indicates the first order delay element in discrete system. The discrete system matrices A d and B d are updated at every step using the estimation of the parameter vector.

I ^

The covariance matrices of system noise Q and measurement noise R are set from the noises superimposed on the stator voltages and currents respectively. One of the frequently used filters to remove the noise from the measured waveform is the low pass filter. But it is not suited because the low pass filter regards only the high frequency domain as the noise. It is assumed that the system noise and the measurement noise are white Gaussian noises; i.e., their spectra are spread out all the frequency domain uniformly. Then another way of filtering is suggested: The waveform obtained by connecting the points of the data smoothly can be regarded as the signal, and the deviation of the data from the signal as the noise. As one of the methods of separating the noise from the data, the simple and practical three point average scheme is proposed. This scheme is to regard the average of the successive three points at t = k - 1, k,IC + 1, of the data as the signal at t = k. It should be noted that this method is not applicable where the voltage waveform changes abruptly. In this case the voltage data themselves are regarded as signals. Figure.4 shows the signal and the noise of the stator voltage and current thus obtained.

n .
"d 0

(signa

(noise)
I

H -

-300
0

10

-0.5 20

tmsl

Figure.3 Structure of the Kalman Filter

10.05

RESULTS
id
0

(signal)

se)
I

Conditions for t h e E x p e r i m e n t
-20'

I-0.05

In the experiment, a squirrel-cage induction motor with a rating of 1.5 kW,4 poles, 50 Hz, 1420 rpm was used. The real time data of the stator voltage and current in load test, which were measured for eight periods successively, were o b tained by means of division of one period into 1024 parts; i.e., the sampling period of A/D converter was about 20 p s. Then these data are applied to an extended Kalman filter. Rotor angular speed was measured using a speed detector. The load test was carried out at rated frequency, rated voltage ( 6 step wave ) and 1442 rpm.
25 I

10

[msl

20

Figure.4 Signal and noise of the stator voltage and current

The variances of the above noises are calculated. Then assuming that there is no correlation between the a-axis and B-axis components, the covariance matrices of the system noise Q and of the measurement noise R are set to

4 = ObO9

20
0 : 9 ] [AI
0

id

R=

20

o.o:02

0 . ; 0 2 ]

S t a r t i n g Values The starting values of the estimations of the state vector and the parameter vector are set to zero-vectors. For the covariance matrix of the estimation error,
0

50

100

P o=

1 0 0 0

0 1 0 0

0 0 1 0

0 0 0 1

0 0 0 0

0 0 0 0

0 0 0 0

0 0 0 0

[=I

0.5

o o o o a o o o o o o o o a o o o o o o o o a o
0 0 0 0 0 0 0 a

xdo

can be used. a should be large enough. In the experiment, a was set to a = 1.0 * 106 The normal convergence of the parameters was observed 0 " . for the order of a from IO6 to 1

-0.5

p v v, v
0
50

v ,v I
loo [ms]

2 . 0

Results of the Parameter Identification

1 / ~l ,/ L l , M and Rs calculated with the filter algorithm

1 .o

off-line. The error between the measurement of the stator current and the estimation is very small. The flux converges to the sine wave in 50 ms. Each identified value of the parameters is normalized by the value for the fundamental component which is identified through no-load tests and locked rotor tests using F F T [3] and which is reliable because the performance prediction using these values resulted in a good agreement with the performance of the real system. The parameters converges in relatively short period to very close values to those derived from the no-load tests and the locked rotor tests. This means that the effects of harmonics on the parameters is small for the induction motor used in the experiment.

50

Figure.5 Results of Estimation

CONCLUSION

linkage, and the parameters of an induction motor only measuring the stator voltages and currents and the rotor speed. It was verified that for the induction motor used in the experiment the values of the parameters were very close to those identified through nwload tests and locked rotor tests. This paper gives a basic step to the parameter identification of induction motors using the Kalman filter and the
252

This exper8,ment proves that an extended Kalman filter can be used to estimate the stator current, the rotor flux

followings are left for further research. In the experiment of this paper, 6-step waveform of voltage was used. Other inputs such as sinusoidal and PWM waveforms should be tested to confirm the applicability of the extended Kalman filter. A model of the induction mctor considering iron loss should also be examined although this was not dealt with in this paper. For the control application, on-line parameter identification is necessary and research on the parameter variation with the temperature is required.

Acknowledgment

The authors wish to thank Dr. R. G. Hoft, professor at University of Missouri-Columbia, for helpful discussions. They also thank Mr. H. Arayama and Mr. M. Sakamoto for their assistance in experiments.

References

[I] L.C.Zai and T.A.Lipo, " A n Extended Kalman Filter Approach To Rotor Time Constant Measurement I n PbvM Induction Motor Drives " IEEEIAS Annual Meeting Conference Record, 1987, pp.177-183.

[2] S. M. Bozic, "Digital and Kalman Filtering Edward Arnold, London, 1979

"

(book)

[3] T. Kataoka, H. Tai, Y . Kandatsu, T. Akasaka, " A New Method of Determining the Equivalent Circuit Parameters and Predicting the Steady State Performance of Inverter Fed Induction Motors '' I E E E IAS Annual Meeting Conference Record, 1987, pp.109-114.

253