You are on page 1of 6

A Design of Nonlinear PID Controllers with a Neural-Net Based System Estimator

Yoshihiro Ohnishi Department of Electrical Engineering and Information Science Kure National College of Technology 2-2-1 I Aga-Minimi, Kure, Hiroshima, 737-8506, Japan Email: onishi@kure-nct.ac.jp

Tom Yamamoto
Department of Technology and Information Education Graduate School of Education Hiroshima University 1-1-1 Kagamiyama. Higashihiroshima, Hiroshima, 739-8524, Japan Email: yama@hiroshima-u.ac.jp classified into the two groups. The one is that control input is directly calculated by the neural network. As the control input is given by the output of the neural network, it is easy to employ for the controlled object. However, it is difficult to express the structure of the controller as the transfer function. The other is that control parameters are calculated by the neural network, and this control input can be calculated by these control parameters. Although this scheme makes easy to grasp the physical meanings of control parameters, the properties of controlled object can not be directly understood. Furthermore, these schemes require the information of system Jacobian to update weighting factors of neural networks. It is difficult to obtain the system Jacobian of such the nonlinear systems. In this paper, the neural network is utilized for the purpose of system identification. That is, the system parameters are estimated by using the neural network. Based on these estimates, PID parameters an: determined in an on-line manner. According to the newly proposed scheme, PID parameters are adequately adjusted corresponding to the nonlinear properties. This paper is organized as follows. The neural-net based system estimator is first proposed. Next, the design scheme of PID controller is explained, whose PID parameters are adjusted based on the relation to the generalized minimum variance control(GMVC). Finally, the effectiveness of the newly proposed scheme is numerically evaluated on a simulation example.
11. NEURAL-NET EIASED SYSTEM ESTIMATOR

Abstract-" control schemes based on the dassieal control theory, have been widely used for various process control systems for a long time. However, since such processes have nonlinear properties, it is difficult to determine 'optimal' PID parameters. In this paper, a system identification scheme by using a neural network is proposed. Furthermore, a PID control scheme based on the estimates is considered. According to the newly proposed scheme, it is possible to employ to nonlinear systems. Finally, the behavior of the newly proposed control scheme is investigated on a numerical simulation example.

I. INTRODUCTION
PID control schemes based on the classical control theory, have been widely used for various industrial control systems for a long time[l],[2].This is mainly because PID controllers have simple control structures, and simple maintain and tune. However, since such processes have nonlinear properties and uncertainties caused by modeling errors and process fluctuation, it is difficult to determine 'optimal' PID parameters. Some PID control schemes have been proposed based on the self-tuning control algorithm for the uncertain systems. However, as many self-tuning PID control algorithm calculate their PID parameters based on the estimates by least squares method, it is difficult to employ for the systems which can be used the least squares method. In the real industrial control systems, the linear mathematical model is used to calculate the

PID parameters. The parameters of mathematical model are


made variable. and these are assigned by using the reference table generated by the priori information. This procedure can clear on the system properties, and can employ the linear control theory for the nonlinear systems. However, it is very difficult to generate the reference table. The generation of this table is required cat-and-try for each processes. On the other hand, the effectiveness of neural networks is discussed for nonlinear systems[3]. Some control schemes by using the neural network have been proposed up to now [4]. The conventional neural-net based control schemes can be 0-7803-7906-3/03/$17 02003 IEEE. .oO

Consider a single-input and single-output system described as the following nonlinear discrete-time model:

1938

input
u(t-k-1)

hidden

output

u(t-k-m)YO-1)

. . . . .
=

the input to the neuron j and the output from the neuron j , respectively. And f(.) is the sigmoidal function used by

----)

Fig. 1. Schematic figure of a neural network.

2 f ( z ) := __ - 1. (4) 1 eca2 Since it is assumed that the absolute values of the system parameters are under 1.0, the sigmoidal function described ns eqn.(4) can be adopted. The output range of the sigmoidal function (4) should be tuned in proportion to the absolute value in the case where the absolute value in order 1.0. The weighting factors included in the neural network are updated based on the back-propagation method whose cost function E, is given by:

where U and y denote the control input signal vector and the corresponding output signal vector, which are given by
U

En

2 { ~ ( t )fi(t)}'. -

(5)

[u(t- k - 1),u(t- k - 2 ) , . . . ,u(t - k - m ) ]


(2)

The' update rule of Wkj can be derived by the following procedure. First, differentiating E, with Wkj yields

Y = [ y ( t - l ) , y ( t - Z ) , . . . ,Y(t-n)l.

Moreover, k denotes the time-delay, and G denotes a nonlinear function. In this paper, a parameter estimation scheme by using the neural network is proposed. First, the following identification model is introduced:

Furthermore,

6k

is given by

B(t)

-iiy(t - 1 ) - & y ( t - 2) bou(t - k - 1 ) , +&u(t - k", - 2) + . . .


+bm,u(t - k , - m' - 1)
(3)

aEn = e a m-) aok 6k := -@ . (7) anetk aOk anet, On the other hand, the update rule of W j i can be derived as follows. Differentiating E, with Wji yields

_where

aE, aWji

aE, &etj &etj aW,,

-6,O,,

(8)

where, k , denotes the minimum value of estimated timedelay. In this paper, the neural network as shown in Fig.1 is constructed for the parameter estimation. Note that every unit included in the output layer of the neural network corresponds to the estimated parameter. In Fig.1, 0 at the hidden and the output layers mean the neuron shown in Fig.2. 12ig.2, wji, netj and 0, mean the weighing factor between the neuron i and the neuron j , the summation of

6 . .- _3'p &etj

= C6kWkj03(1- Oj).

(9)

Thus, update rules of Wkj and Wji are given by

Wkj(t + 1) = Wkj(t)+ r j 6 k O k
Wji(t

+ C?Awk,(t)

(lo)
(11)

+ 1 ) = Wj,(t)+ q6jOj +aAWj,(t)

9
Oi,
neti

Fig. 2. Mathematical model of a neuron.

Furthermore, r j and OL denote the learning rate and the mOmentUm rate, reSpeCtiVeb'. The system identification can be realized via the above procedure. Note that according to the proposed scheme, a priori information about the system Jacobean is not necessary. This gives us a feature that it is possible to deal with unknown time-delay systems. In most industrial processes, it is relatively difficult to identify the time-delay exactly. Therefore, this is an advantage in designing process control systems. According to the newly proposed scheme, variable system parameters can be obtained corresponding to nonlinear properties. Therefore, by regarding the controlled object as the linear

1939

model, the controller can be designed the similar way to linear control scheme. In the following section, PID controller is designed, where PID parameters are adjusted based on the relation to the GMVC. 111. PID CONTROL The following velocity-type PID controller is employed for the controlled object given by eqn.(l):

Next, (12) is rewritten as:

C ( z - l ) y ( t )+ A u ( t )- C(z-')w(t) = 0
where

(20)

C(z-1)

+q t - 1 +

c2.5-2

Au(t)= k,-e(t) - k,(A + - A T,


TI
T S

2 )y(t),

(12)

where e ( t ) denotes the control error signal given by


e ( t ) := r ( t )- y(t)

Then, by considering the steady property of the control system, E ( z - ~ ) B ( z -in )eqn.(l7) can be replaced to the ~ static gain E ( l ) B ( l ) .

(13)

and k,, TI and TD are the proportional gain, the reset time and the derivative time, respectively. And, Ts denotes the sampling interval. These PID parameters are strongly depend on the control performance. The design method is considered based on the relation to the GMVC by following procedure. First, consider the following cost function to derive a GMVC control Iaw[5]:
J = E[@(t k , + I ) ]

F(r-')y(t)

+ { E ( l ) B ( l )+ X}Au(t)- R(z-')w(t) = 0
(22)

is defined as:

v := E ( l ) B ( l ) X
(22) can be rewritten as:
-y(t)
F(2-l)

(23)

+' k ( t )
~

R(z-')
~

w(t) = 0

(24)

(14)

where
+(t

Therefore, by comparing eqn.(20) and eqn.(24), R(2-I) and C(2-l) can be designed .as:
R(2-l) =

+ k, + 1 )

:=

P(z-')y(t

+ k, + 1 ) + XAu(t)
(15)

F(2-I)

-R(z-')w(t)

In this cost function, X is the weighting factor of the control input signal which is the user-specified parameter. P ( 2 - l ) is the user-specified polynomial defined as:
P(2-1) = 1

By this procedure, PID parameters can be calculated based


on GMVC. Therefore, by using eqn.(21) and eqn.(26), PID parameters can be obtained as the following equations.

+p*r-'

+p2t-2

(16)

k,

--U1 + 2 j 2 )
I/

Furthermore, the polynomial R(z-') is designed by the relationship with PID control law. Minimizing the cost function J yields the following control law:

F(e-')y(t)

{E(z-')B(z-')

+ A}Au(t)
=0

This algorithm of proposed scheme is summarized below. [Proposed PID control algorithm]
1.

-R(t-')w(t)

(17)

Obtain the input-output date for the learning of neural network. Design the structure of identification model (3). Learn the system property by using the neural network. Design P(2-I) and A. Calculate ii and

where F ( 2 - I ) and E ( 2 - I ) is obtained by solving the following Diophantine equation:

2.
3.
4.

P(z-') = A A ( z - ' ) E ( t - ' )


where

+ z - ( ~ * + ' ) F(t - I )

(18)

5.
6.
1940

21, by using the neural network.


F(2-l)

Solve E ( 2 - l ) and

based on eqns.(lS).

7.

Solve

based on eqn.(23).
......: . .

8.
9.

Solve k,, TI and TD based on eqn.(27) Calculate u(t)from eqn.(l2). Update t and return to 5..
0

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

10.

10

Y)

IV. SIMULATION EXAMPLE


In order to investigate the behavior of the proposed control scheme, a numerical simulation example is illustrated in this section. [Case 11 The controlled object is given by the following second order Hammerstein model:
2.5

.......................................... .... : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . : ... ,.. ....................................... al ...................... ..: ......: ............


I 1 0

a .
I [*=PI

Y)

10

y ( t ) = 0.9y(t - 1 ) - 0.3y(t - 2 ) 0.4z(t - 1) +0.2z(t - 2)

Fig. 4. Cooml result by using the proposed scheme in Case 1

z ( t ) = u(t) 0.5uZ(t).

(28)

The following model was used for system identification:

& ( t ) --81y(t =

- 1 ) - &y(t - 2 )

+ i0u(t'- 1 ) + 6,u(t - 2 )
(29)

Fig.3 shows the simulation result by using the fixed PID parameters, whose parameters are k , = 0.17, TI = 1.0 and To = 0.25. These parameters can be obtained using CHR scheme[2]. The number of units included in the inputs layers and the hidden layers of the neural network were set as 4 and 6. The parameters used in the training were set as

Fig.4 shows the simulation result by using the proposed scheme. and Fig.5 shows the trajectories of the calculation results of PID parameters. From these results, it is clear that the proposed control scheme can adjust the PID parameters corresponding to the nonlinear property. [Case 21 The controlled object is given by the following third order Hammerstein model:

y(t)

0.6y(t - 1)- O.ly(t - 2 )

+ 1 . 2 ~ (- 1 ) t
- 2)

-O.lz(t

z ( t ) = 1.5u(t)

+ 1.5u2(t) + 0.5u3(t).

(31)

a = 1.0, q = 0,001, (Y = 0.001.


This learning was carried on 50000 iterations.

(30)

Fig.6 shows the static property of the controlled object eqn.(31). From Fig.6, it is clear that this controlled object has the nonlinear property. Fig.7 shows the simulation result by using the fixed PID parameters, whose parameters are k , = 0.25, TI = 0.2 and

...

Fig. 3.

e *.r* . . . .

,,, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
0

'T c *l
Y) I*)

.....:......

..:.. . . . . . . . . . . :. ..

.i . . . . ...j

\w-r
0

10

20

.o

70

e. .,
II

. . . . : ...

: ...

......

.......

.....

........ ....................
m
.o
Y)

........ ........
m

DI

10

I*)

:...........;

...........

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

0 '
0

....

........

......

...

...

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

10

a t IWWI

Y)

10

a
0

...

70

20

.O

Y)

f a

70

TD = 0.25).

Control result by fixcd PID parameters(kc = 0.17.TI= 1.0 and

t [step]

Fig. 5 . Trajectories of PID parameters i case 1 o

1941

,~.. .................................
..............

J/ l

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

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

I,I

.iM
,
....................
.I.
0
10

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

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

I1
, .
I I

D
~~

.............................................
m
5

a
I (step1

IO

input signa1

Fig. 6. Stltic property of controlled object.

Fig. 8 . Coritrol result by using LSM.

TD= 1.0. These parameters ware designed by cut-and-try. It is difficult to obtain the good control result by the fixed PID
parameters. The following model was used for system identification:

parameters used in the training were set as

a=1.5,~=0.001,a=0.001.

(33)

Y ( t )= --81y(t

- 1) - &y(t - 2)

+ 6,U(t

- 1)

+ 61u(t - 2)

This learning was carried on 50000 iterations. Fig.10 shows the simulation result by using the proposed scheme. and Fig.11 shows the trajectories of the calculation results of PID parameters. From these results, it is clear that the proposed control scheme works well for nonlinear systems.

F i g 3 shows the simulation result by using the least squares method, whose parameters are i1 = -1.2296,& = +0.2618,60 = 0.4334,61 = -0.3953. These parameters are estimatied in the neighborhood of 1.0. It is also difficult to obtain the good control result by using LSM. On the other hand, the learning of the neural network was carried out by using input-output data shown as Fig.9 T h e number of units included in the inputs layers and the hidden layers of the neural network were set as 4 and 10. The

V. CONCLUSION
In this paper, an identification method by using the neural network and a design method of PID parameters based on the estimates have been considered. Furthermore, the effectiveness of the proposed method has been numerically evaluated on a simulation example. The main features of the proposed control scheme are summarized as follows.

81 L

k..;.
........ .. ....

.:.. . . . . . . . . . . . j

..-r--: Kd...;.
lo,l

*.I

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

-[ .....\Ti " I
n
.........

( 2

...................
........
~

....

To&
o , 2 .

.... : . . . . . . . ~ ... , ....


m

,~

...... .. ;... ............. I ....... i ..'I. . I . . . .i ........... ~.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .


Im

zm

211

91

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

,
YJ 40

,,J 3
o,I

p r
.....

. . . . . .

... . ; . . . . . . . . .

;........

"'

. . . . . . . . . . . . . . . . . ;..
YJ

10

io

11

ea

70

t [rtepl

I m

,SO

I ISIeQl

Fig. 7. Control result by fixed PID parameters(& = 0.25, TI = 0.2 and1 T~ = 1.0).

Fig. 9. Input-output date sets used for learning

1942

(I

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

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

........................................
0 1 0 m Y ) a a Y ) 7 o a o

[41 K. S. Narendra and K. Parthasarathy, Identification and Canhol of Dynamical Systems Using Neural Networks, IEEE Trans. Neural Networks. Vol. I , No.1. pp.4-27(1990). [SI D.W.Clarke and P.I.Gawthrop : Self-tuning conholler ; IEE Pmc., V01.122D;No. 9, pp.929-934 (1975)

.......

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

10

a
I lstePl

Y)

UI

E o

Fig. IO.

Control result by using Ule proposed scheme in case 2

This scheme can be employed for the nonlinear systems. PID parameters can be obtained based on the relation to the GMVC. PID parameters are adjusted corresponding to the nonlinear properties.

We are now in employing this scheme to a real system. The control results will be shown in that day.

ACKNOWLEDGMENT
This research was partially supported by The Japan Society for the Promotion of Science, Grant-in-Aid for Young Scientists (A), 14750382.

REFERENCES
[I1 1. G. Ziegler and N. B. Nichols. Optimum Setting for Automatic Conhollers, Trans. ASME, Vol.@%, pp.759-768 (1942)

I21 K. L. Chien, J. A. Hrons and J. B. Reswick, On the Automatic Control of


Generalized Passive Systems, Trans. AMSE, Vo1.74, pp.175-I85 (1972) [31 S.Haykin, Neural Networks. Macmillan College Publishing, New York, (1994).

~~

. 5..>.. r
0.6

.................. .................... . . . . . . . . . . . . . . . . . . ...... . . . . . . . . . . . . . . . . o2 . . . . . . . . . . . . . .

Or

G>., ........

............... :. ........:

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

o,l

........

=~ , ~ . . .
01.'

b-

0,s-

. . . . . . . . . . . . :. ............................... ...................... . .......................... ....; . .

1943