Application of Adaptive Neuro Fuzzy Inference System on the development of the Observer for speed sensor less Induction

Motor
Era Purwanto,MemberJEEE, Syamsul Arifin, Bian-Sioe So
L ~L, ,

: Stator and rotor self inductance

Abstract: Direct field-oriented induction motor drive system need rotor flux observer and rotor angular speed identifier. ANFIS (Adaptive Neuro Fuqy Znfrence System) used for identifying parameter dynamics and system variable estimation, linear either non-linear. ANFIS with backpropagation learning algorithm has applied to estimate flux rotor and identi@ rotor angular speed of three-phase induction motor.

M
0
7 ,
0,

: Mutualinductance : Leakage coefficient, C = ~ - M * / L ,L,

: Rotor time constant, rr = L, / R,

: Motor angular speed

3. BASICS OF ANFIS

3.1 Adaptive Networks

1. INTRODUCTION
To control induction motor which parameters are uncertain, for this case considered the controller is of known structure and uses the values of the state variable. But some state variables cannot be measured, such as rotor flux. Then we need an observer to estimate the state variable that cannot be measured. A flux observer based on ANFIS will be used to estimate the flux, for the purpose of i d e n t i w g motor induction angular speed.
2. INDUCTION MOTOR MODEL

The state equations can be expressed as written below in the stationary reference frame (dq model)

An adaptive network is a multilayer feed forward network in which each node performs a particular function (node function) on incoming signals using a set of parameters specific to this node [3]. The form of node functions may vary from node to node, and the choice of each node function depends on the overall function, which the adaptive network is designed to implement. To reflect different capabilities, we use both circle and square nodes in an adaptive network. A square node (adaptive node) has modifiable parameters while a circle node (fixed node) has none. The parameter set of an adaptive network is the union of the parameter sets of each adaptive node. In order to achieve a desired inputoutput mapping, these parameters are updated according to given training data and a gradient-based update procedure described below. Suppose that a given adaptive network has L layers and layer 1 (I = O,l, ..L;I = 0 represents the input layer) has No) node. Then the output and function of in layer I can be represented as x,,~ node i (i = 1, and fr, ,asshown in the figure. Since the output of a node depends on the incoming signals and the parameter set of the node, we have the following general expression for the node function: ~ ~ t = f(X ~ I -tI , I , ...., ~ l ~ , ~ ( ~ ) a , P , x . . (3.1) .) a$,y are the parameters of this node. Assuming that the given training data set has P entries, we can define an error measure for the pth (1< = p < = P) entry of the training data as the sum o f squared error :

...,No))

where dk : the kth component of the pth desired output vector, XLk : the kth component of the actual output vector produced by presenting thepth input vector to the network.

&,&

: Stator and rotor resistance

ZEEE Catalogue No. 01 CH37239 0-7803-7101-1/01/$lO.OO 0 2001 ZEEE.

409

( For notational simplicity, we omit the subscript p for both dk and XL,& ) Thus our task here is to minimize an overall error measure, which defined as

(3.10)

E=

xLl
V -.

E,

(3.3)

XQ1

where & is the step size, the length of each transition along the gradient direction in the parameter space. Usually we can change the step size to vary the speed of convergence.

x0,z

3 . 2 Adaptive Neuro-Fuzzy Inference Systems

I .

I .

L .

Figure 3.1. Feed forward adaptive network

We define the error signal q i as the derivative of the error measure Ep with respect to the output of node i in layer 1,
EL,

=aE,

ax,

(3.4)

This expression was called the ordered derivative by Werbos. The error signal for the ith output node at layer L could be calculated directly a+E, - a E~ (3.5) E , . $ = --

ax,.,

axL..,

For the intemal node at the ith position of layer 1, the error signal can be derived by the chain rule

Adaptive Neuro-Fuzzy Inference Systems (ANFIS) is a class of adaptive networks that are fimctionally equivalent to fuzzy inference systems. Structurally, the only limitation on the network configuration is that it should be feed forward type if we do not want to use the more complex asynchronously operated model. We assume that the fuzzy inference system under consideration has two inputs x and y, and one output z. For a first order Sugeno fuzzy model, a common rule set with two fuzzy if-then rules is the following: 1 : I f x is AI and y is B I , then5 =pIx+qly+rI 2 : I f x is A2 and y is B2, then5 =pzx+qg+r2

where 0 <= 1 <= L-I For any I and i (0 < = 1 < = L and 1 < = i < = NO)) the error signal can be found by first applying Equation (3.5) once to get error signals at the output layer, and then applying muation (3.6) iteratively until we reach the desired layer 1. The underlying procedure is called backpropagation since the error signals are obtained sequentially from the output layer back to the input layer. The gradient vector is defined as the derivative of the error measure with respect of each parameter. If a is a parameter of the ith node at layer 1,

fl=PPWW,
I

figure 3.2 A two input f w order Sugeno fuzzy model with two rules

figure 3.3 Equivalent ANFIS architecture

The derivative of the overall error measure E with respect to a

The update formula for the generic parameter a

&=-q- aE

(3.9)

aa

where 2 2 is the learning rate, which can be further expressed as

Figure 3.2 illustrates the reasoning mechanism for this Sugeno model, the corresponding equivalent ANFIS architecture is as shown in figure 3.3. (The output of the ith node in layer I as O,i) Luyer I : Every node i in this layer is an adaptive node with a node function fori = 1,2, or 0 1 , i = m Ai@) for i = 3,4 OI,i = m ~ - 2 6 9 , where x (or y) is the input to the node i and Ai (or Bi.2) is a linguistic variable. In other words, OI,i is the membership grade of a fuzzy set A (=AI,At, BI, Bd and it specifies the degree to which the given input x (or y)

410

satisfies the quantifier A. Here the membership function for A can be any appropriate parameterized membership function, such as Gaussian function: -1x-b e 2( a(3.1 1) where {ai, bi} is the parameter set. As the values of there parameters change, the Gaussian-shaped function varies accordingly. Parameters in this layer are preferred to as premise parameters. Luyer 2 : Every node on this layer is a fixed node labeled Fb, whose output is the product of all the incoming signals 02,i = wi = m m Bi(v) i = 1,2 (3.12) Each node output represents the firing strength of a rule. Layer 3 : Every node in this layer is a fixed node labeled N. The ith node calculates the ratio of the ith rule's firing strength to the sum of all rule's firing strengths (3.13) 03,i = wiJ = wi @q + q,y + ri, @i, q i . ri} is the parameter set of this node, known as consequent piuameters. Layer 4 : The first node in this layer computes the output of layer 3, the second node below computes the normalized firing strength of layer 2. Layer 5: The single node in this layer is a fixed node labeled 4, which computes the overall output as the division of the incoming signals:

P . @ )

%)'

We use Gaussian function for its smoothness, instead of triangular shape. We do not have domain knowledge (from a human operator's point a view) about the observer, then the number of fuzzy if-then d e s has to be decided by trial and error. The flux observer and angular speed identifier is implemented as ANFIS with four inputs on each. With substmctive clustering method, we have eight membership function for each input, and eight rules for each observer and identifier. (Though the number of h z y rules can be more than eight, the simulation indicates eight rules are enough for identification).
4.3 Training

owrallou~rr=Ou=~W~=i

ZiW4
CiWI

(3.14)

To speed up the convergence, we follow a strict gradient descent in the sense that each transition of the parameters will lead to a smaller error measure. If the error measure increases after parameter update, we back up to the original point in the parameter space and decrease the current step size by half. This process is repeated until the weight update leads to a smaller error measure. However, this step size update rule tends to use a small step size if the error measure surface encountered in the first few updates is not smooth. Therefore we multiply the step size by 4 after observing three consecutive transitions without any backup actions. The initial step size in the simulation is 2 and the learning process stops whenever the number of epoch given has reached.

4. DESIGN AND ANALYSIS

We use Matlab 5.3.1, which already provide the ANFIS function, for the observer design and simulation. Induction motor receive the inputs as real quantities, then proceed with mathematics model solution by Runge Kutta-Gill integral towards differential equation to produce real quantities. The observer was designed to estimate some induction motor outputs that can not be directly measured. Observer performance was measured by the difference between induction motor output and observer observer. More less the difference then can be said that observer already has high performance.
4.1 Data

fi-

4.1 ANFIS Architecture of direct rotor f l u x observer

Training d a t a taken while the overall system simulation running without the observer. The induction motor driven by PWM Inverter and proportional speed controller with gain=l, at the set point 250 rads. Simulation run for 5000 iteration, so we can get 5000 a t a . This data used for find the fuzzy inference pairs d system and the training.

4.2 Fuzzy Inference System Design

figure 4.2 ANFIS Architecture of quadrature rotor flux observer

41 1

figure 4 . 7 o v d system used for simulation

Piumeters used for observer and identifier:

Table 4 . 1 Parameters of observer and identifier

No

n ( ml

figure 4 . 3 ANFIS Architecture of rotor angular speed identifier

Figure 4.4 shows off-line configuration of the rotor flux observer, figure 4.5 for rotor angular speed identifier. On-line configuration at figure 4.6 shows us that estimated flux from flux observer make the input for i t h stator current. rotor angular speed identifier together w
0.11

a,

re*

e n v

I3.w

@ma

<e

figure 4 . 7simulation results for direct rotor flux estimation w i t h

training 200 epochs

figure 4.4training scheme of observer Iz,

Induction; Model

l-r*

-FW

she

W b f m 6 4s)

i.
A N F I S IdentiRer (2)

figure 4 . 8simulation results for quadrature rotor flux estimation w i t h training 200 epochs

figure 4.5training scheme of identifier

ANFIS FLUO b s m
ANFlS Swed

Identifier (2)

[-*a* r w * a w l )

& . B

tlW*

(*+

figure 4 . 9 simulation results for direct rotor flux estimation w i t h training 500 epochs

figure 4.6 Configuration of flux observer and speed idenflier at the simulation

4.4 Validation At this validation stage, observer and identifier applied connected with induction motor, as shown below. e r e running for Time sampling 0.00005 seconds, motor w 0.5 seconds.

-94

[-re -rss

"&&E*

1 1 , s
SXRII

Is)

figure 4 . 1 0 simulation results for quadrature rotor flux estimationwith training 500 epochs

412

a m . t* figure 4.11 simulation results for direct rotor flux estimation with traming 1000 epochs

I Qb

-w

-wsm-ra&re

as
~sltfl~

firmre 4.15 simulation results for rotor anrmlar - meed - estimation with 1000 epochs

r*

-s x r p l

._.--

? , -

timr Is)

fieure 4.12 simulation results for auadrature rotor flux estimationw i t h &g 1000 epochs

This plot shows the original rotor flux and the one predicted by ANFIS. The difference is so tiny that it is impossible to tell one from another by eye inspection. That is why we show the error curve also.
.............. : ............. ............,. .......... .........
~ ~

Standard Error Estimation (SEE) on each experiment are: l*: training 200 epoch results SEE direct rotor flux 0.0437 Wb, SEE quadratur rotor flux 0.0412 Wb, and SEE rotor angular speed identifier 19.2394ruds. 2nd:training 500 epoch results SEE direct rotor flux 0.0425 Wb,SEE quadratur rotor flux 0.0367 Wb, and SEE rotor angular speed identifier 16.7773ruds. 3d :training 1000 epoch results SEE direct rotor flux 0.0425 Wb,SEE quadratur rotor flux 0.0369 Wb, ,and SEE rotor angular speed identifier 16.7118 ruds.

5. CONCLUSION The results show that rotor flux observer and rotor angular speed identifier can estimate the pattem of flux rotor and rotor angular speed. Thus rotor flux observer and rotor angular speed identifier shown a good performance.
RATING AND PARAMETERS OF INDUCTION MOTOR USED FOR SIMULATION

j . . . . . . . . .

....

.............

J-VH

^Ylkht.b,W

OS time (ai

1 0 Watt , 115 Volt, 2 poles, 60 Hz

figure 4.13 simulation results for rotor angular speed estimationw i t h traming 200 epochs
............ ....................... ............r-........ ........... .............
1 .
~

R, = 176 R,= 190

L,= 3.079 L,= 3.31 M=3 . 2 1 J = 0.0000105 = 1.49e-5

rob1
[Ob1

N
[HI

N
Ekg.mZI *g.m2 / SI
REFERENCES

OS 1-4 - m e a t Wir 1 C 1 n m 8s) figure 4.14 simulation results for rotor angular speed estimation w i t h training 500 epochs

1. Boldea, I., Vector Control of AC Drives, CRC Press, Inc., 1992 2. Hisao Kubota and Kouki Matsuse, Speed Sensorless Field Oriented Control of Induction Motor with Rotor Resistance Adaptation, IEEE Trans. Industrial Applications, Vol. 30, No. 5, September/October 1994

413

Jang, J. S. R., Selfkarning Fuzzy Controllers Based on Temporal Back Propagation, IEEE Trans. On Neural Networks, Vo1.3, No. 5, Sept 1992 4. Jang, J. S. R., C. T. Sun and E. Mizutani, NeuroFuzzy and Soft Computing, Prentice-Hall International, Inc., 1997 5. S.L. Chi- Fuzzy Model Identification Based On Cluster Estimation, Journal of Intelligent and Fuzzy Systems, 2(3), 1994 6. Yusinnan I., Perancangan Observer Berbasis Jaringan Saraf Tiruan Untuk Identifikasi Kecepatan Motor Induksi Tanpa Sensor Kecepatan, Tugas Ahir S-1 Jurusan Teknik F i s h ITS, 1999
3.

414