International Journal of Advanced Computer Science, Vol. 2, No. 5, Pp. 200-203, May, 2012.

Manuscript
Received:
25,Jun., 2011
Revised:
25,Oct., 2011
Accepted:
8, Apr., 2012
Published:
15,Jun., 2012

Keywords
PI D,
decoupling
controller,
DRNN,
stability
condition
Abstract For the multivariant, nonlinear,
time variation and strong coupling plant, a
PID decoupling controller based on DRNN
is proposed. Firstly, a PID controller based
on DRNN is constructed. And then several
PID controllers based on DRNN are adopted
in parallel as the decoupling controller,
which can accomplish decoupling and
controlling simultaneously. Finally, the
convergence condition of the controller is
presented based on the Lyapunov theory.
Simulation results show that the proposed
PID decoupling controller is more effective.



1. Introduction
A PID controller is a generic control loop feedback
mechanism; it is still used widely in industrial control
systems because of its simple control structure. For
enhancing control effectiveness, many new types of PID
controller have been proposed [1]-[2]. For the traditional
PID controller, it is usually used to control SISO system,
but in industrial process control fields, most systems are
the MIMO nonlinear systems. Due to the coupling among
various inputs and outputs in MIMO systems, its control
problem is very complicated; the traditional control
methods cant acquire satisfying result. For solving the
control to MIMO systems, many decoupling controller
have been proposed [3]-[5].
It is well known that the neural network owns the
capability of self-learning and self-organization, and
theoretical works have proven that, even with one hidden
layer, neural networks can uniformly approximate any
continuous function over a compact domain [6], so it is
suitable for the nonlinear coupling control system. Over the
past years, neural networks, especially feed-forward
networks, have been widely applied to identify and control
nonlinear system [3], [7]. Compared to feed-forward neural
network, because Recurrent Neural Networks (RNN) owns
feedback neurons in hidden layers, it is more suitable to
identify and control dynamic systems [8]-[9]. Diagonal
Recurrent Neural Network (DRNN) is an RNNs
simplification form.

In this paper, DRNN is used as a plant identifier and a
nonlinear PID controller based on DRNN is constructed.
Several PID controllers are adopted in parallel as the PID

Guijin Han are with Xian University of Posts & Telecommunications,
Automation College(hgjin.123@126.com)
decoupling controller. The proposed PID Decoupling
controller not only can overcome the drawback of difficulty
in tuning the parameters of conventional PID controller, but
also can accomplish decoupling function. Finally, using the
Lyapunov stability theory, the convergence condition of the
controller is analyzed theoretically. Simulation shows that
the algorithm is effective and practical for nonlinear
multivariable process control.

2. DRNN structure and learning
The DRNNs structure is shown in Fig.1, which
consists of one input layer, one hidden layer and one output
layer, where
( ) x k
is the input vector, ( ) h k is the hidden
layer output vector and ( ) y k is the network output vector.
h
W ,
d
W and
y
W are connection weight matrices
between input layer and hidden layer, within hidden layer
and between hidden layer and output layer, respectively.
The structure of one hidden layer neuron with
feedback is shown in Fig. 2, the feedback vector within the
hidden layer includes 1 to v time lags to give delayed
values of the hidden layer output, which is used to
internally present the dynamics of the process to be
modeled.
The DRNN with q hidden layer nodes used in this
research can be governed by the following equations.

( )
( )
1
y
h k
y k W
(
=
(

(Equ. 1)
) (
1
1
) (
k z
e
k h

+
= (Equ. 2)
1,1 ,1
1, ,
1
0 0 ( 1)
( )
( )
1
0 0 ( )
( 1)
( )
( ) (
1
( )
d d
v
h
d
q v q
h d d
v
w w h k
x k
z k W
w w h k v
h k
x k
W diag w diag w
h k v
( ( ( (
( ( ( (
(
= +
( ( ( (
(

( ( (
(


(
(
(
( = +
( (

(

(Equ. 3)
A PID Decoupling Controller Based on DRNN
Guijin Han
Guijin Han: A PID Decoupling Controller Based on DRNN.
International Journal Publishers Group (IJPG)


201

Fig. 1. DRNN structure.


Fig. 2. Hidden layer neuron structure.

The usual method of modulating weight value of
DRNN is gradient descent algorithm with momentum term
(GDM). To optimize the algorithm, the proportional and
derivative function can add into GDM according to PID
control theory, and then get PIDGDM [10]. PIDGDM not
only has faster track and convergence characteristic than
GDM, but also has bigger possibility of jumping out local
extremum, which has been proved [11].The PIDGDM is as
Equations 5 and 6:
1

( ) ( ) ( ) ( ) ( )
2
1
( ) ( )
2
T
T
J k Y k Y k Y k Y k
E k E k
( (
=

=
(Equ. 4)
( ) ( )
( ) ( )
( ) ( )

( )
( )
( )
T
T
E k J k
W k E k
W k W k
Y k
E k
W k
q q
q
c c
A = =
c c
c
=
c
(Equ. 5)
( ) ( 1) ( )
[ ( ) ( 1)]
[ ( ) 2 ( 1)
( 2)] [ ( 1) ( 2)]
p
d
W k W k W k
W k W k
W k W k
W k W k W k
q
q
= + A +
A A +
A A +
A + c
(Equ. 6)
where q is the learning rate,
p
q ,
d
q is the
proportional and derivative parameter respectively, c is
the inertia coefficient. Z-D method is used to adjust
parameter q ,
p
q and
d
q , and then adjust parameter
c , (0 1) c < c < is used for assuring algorithms
asymptotic stability.
3. PID Decoupling Controller
It is well known that the input vector of traditional PID
controller is
1 2 3
( ) [ ( ), ( ), ( )] X k x k x k x k = and its output is
the control signal ( ) u k .
1
( ) ( ) ( ) ( ) x k e k r k y k = = (Equ. 7)
2
( ) ( ) ( ) ( 1) x k e k e k e k = A = (Equ. 8)
2
3
( ) ( ) ( ) 2 ( 1) ( 2) x k e k e k e k e k = A = +
(Equ. 9)
Where ( ) r k is the reference trajectory of controlled
plant, ( ) y k is the realistic output.
If we set the input vector of DRNN to
2
1 2 3
( ) [ ( ), ( ), ( )] [ ( ), ( ), ( )] X k x k x k x k e k e k e k = = A A , The
output of DRNN is the corresponding control signal
( ) u k ,then DRNN can be regarded as a nonlinear PID
controller.
For MIMO control system with n inputs and
n outputs, n PID controllers based on DRNN can be
adopted in parallel to constract PID decoupling controller.
The structure is shown as Fig.3.

Fig. 3. Decoupling Controller Structure.

The target function for the i th controller as Equ. 10
| |
| |
2 ,
2
1
2
1
1 1
( ) ( )
( ) ( ) ( )
2 2
1
( ) ( )
2
n j i
j j
i i i
j
n
i i
i
r k y k
J k r k y k
r k y k
=
=
=
(
= +
=

(Equ. 10)
In Equ.10, the second term serve as the decoupling part.
Define
T
n C
k e k e k E )] ( , ), ( [ ) (
1
= , we can rewrite
Equ.10 as the following:
) ( ) (
2
1
) ( k E k E k J
C
T
C i
=
(Equ.11)
Suppose ( )
i
V k is the weight vector of the i th
controller. Then PIDGDM can be used to train i th PID



NNC1

NNCn
x
11
x
12
x
13
u
1
x
n2
x
n1
x
n3
u
n
International Journal of Advanced Computer Science, Vol. 2, No. 5, Pp. 200-203, May, 2012.
International Journal Publishers Group (IJPG)


202
controller, According to (Equ. 5), we can get
( ) ( ) ( )
( )
( ) ( ) ( )
( ) ( )
( )
( ) ( )
i i i
i i i
i i i
T i
i C
i i
J k J k u k
V k
V k u k V k
u k Y k
E k
u k V k
q q
q
c c c
A = =
c c c
c c
=
c c
(Equ.12)
Let
T
U
U Y Y c c = /
.
) ( / ) ( k u k Y
i
c c
is the i th colu-
mn of
U
Y
, written as
i
U
Y
. While we regard ) ( ) (

k Y k Y ~ ,
then Y Y
U

~ , so PIDGDM for the i th controller is as
follows:
) (
) (

) ( ) (
k V
k u
Y k E k V
i
i i
U
T
C i i
c
c
= A q
(Equ.13)
( ) ( 1) ( ) ( ( ) ( 1))
( ( ) 2 ( 1) ( 2))
( ( 1) ( 2))
i i i pi i i
di i i i
i i i
V k V k V k V k V k
V k V k V k
V k V k
q
q
= + A + A A
+ A A + A
+ c
(Equ.14)
The control system structure is shown in Fig.4, where
the controlled plant is a coupling plant with n inputs and
n outputs, NNI is DRNN identifier, NNCi is PID
controller based on DRNN.
n
u u , ,
1
is control inputs,
n
y y , ,
1
is real outputs,
n
y y , ,
1
is the outputs of
NNI, and
n
r r , ,
1
is reference inputs.

Fig. 4 Control System Structure.

4. Convergence Analysis
Using the Lyapunov stability theory, the convergence
condition of the controller can be analyzed.
Convergence Condition Suppose ) (k V
i
of the i th
nonlinear PID controller based on DRNN is updated along
(Equ.14),
i
q is the learning rate for the weights. Then this
DRNN will be convergent in exponential speed, if
2
2
) (
) (

2
0
k V
k u
Y
i
i i
U
i
c
c

< <q
(Equ.15)
where - is Euclidean norm.
Proof Construct discrete-time Lyapunov function
) ( ) (
2
1
) ( k E k E k L
C
T
C i
=
(Equ.16)
Then, the change of the Lyapunov function due to the
training process is obtained
) ( ] 2 / ) ( ) ( [ ) ( ) 1 ( ) ( k E k E k E k L k L k L
C
T
C C i i i
A A + = + = A
) (
) (
) (

) (
) (
) (
) (
) (
) (
) (
) (
) (
k V
k V
k u
Y
k V
k V
k u
k u
k Y
k V
k V
k E
k E
i
T
i
i i
U
i
T
i
i
i
i T
i
C
C
A
(

c
c
=
A
(

c
c

c
c
= A
c
c
= A
(Equ.17)
From (Equ.13), this equation can be written as
2
) (
) (

) (

) (
k V
k u
Y k E Y k E
i
i i
U
T
C
i
U i C
c
c
= A q
(Equ.18)
Then
2
2
4
2 2
2
( )

( ) [ ( ) ]
( )
( ) 1

( )
2 ( )
T i i
i i C U
i
i T i i
i U C U
i
u k
L k E k Y
V k
u k
Y E k Y
V k
q
q
c
A = +
c
c
(


c
(Equ.19)
To ensure DRNN be convergent in exponential
speed, it should be 0 ) ( < A k L
i
, which leads to Equ.15.
This completes the proof.
5. Simulation Results
Now, we test the proposed PID decoupling controller
above using the following system with 2 inputs and 2
outputs as the controlled plant
( ) ( )
2
1 1 1 1 2
( ) 1.0/ 1 ( 1) 0.8 ( 1) ( 2) 0.2 ( 3) y k y k y k u k u k = + + +
(Equ.20)
( ) ( ) ) 2 ( ) 3 ( 3 . 0 ) 1 ( 9 . 0 ) 1 ( 1 / 0 . 1 ) (
2 1 2
2
2 2
+ + + = k u k u k y k y k y
(Equ.21)
The structure of NNI are selected as 2-9-2, and the
structure of two PID controller based on DRNN are all
selected as 3-7-1, and 2 = v , sampling time s T 1 = . First,
we take the setting value of the system output are 1,0
respectively. Simulation results are shown in Figure5 and
Figure6 is simulation results using PID decoupling control
strategy [1].

n



+

-

+

-


NNI
u
1
u
n
y
n
y
1
NNC1
NNCn

PLANT
r
1
r
n
+

-

+

-


Guijin Han: A PID Decoupling Controller Based on DRNN.
International Journal Publishers Group (IJPG)


203

Fig. 5. Results of PID decoupling control based on DRNN.


Fig. 6. Results of PID decoupling control.

We can see from the Fig.5 and Fig.6, compared with PID
decoupling control, the proposed PID decoupling controller
based on DRNN can trace the setting value faster.
Secondly, we take pulse signal as setting value of the
system outputs for testifying validity of the proposed
control strategy better. Simulation results as shown in
Fig.7.

Fig. 7. Results of PID decoupling control.

We can see from Fig.7, after the initial learn online,
both the outputs realize the tracing to the setting value
respectively and the function of the decoupling control.
System has better static state and dynamic state
characteristic, and the output is stable and the error is
small.
6. Conclusion
In this paper, a PID decoupling controller based on
DRNN is presented. The main result is that several PID
controller based on DRNN are adopted in parallel as the
decoupling controller to accomplish decoupling and
controlling simultaneously and the convergence condition
of the controller is presented based on the Lyapunov theory.
For testing the validity of proposed PID decoupling
controller, a nonlinear coupling plant is studied using the
proposed PID decoupling controller, simulation results
show it is more effective.
References
[1] Jinkun Liu, Advanced PID control and its MATLAB
simulation. Bei Jing: publishing house of electronics
industry, 2004, ch. 6, pp. 247-250.
[2] Chen, Z.Q., Mao, Z.X., Du, Sh.Zh., "Analysis of robustness
of PID-GPC based on IMC structure," (2003) Chinese J.
Chem.Eng, vol. 11, no. 3, pp. 55-61.
[3] Lianfei Zhai , Tianyou Chai, " Nonlinear decoupling PID
control using neural networks and multiple models, "(2006)
Journal of Control Theory and Applications, vol. 4, no. 1,
pp. 62-69.
[4] Jiao Zhu-qing, Qu Bai-da, Xu Bao-guo,"Decoupling Design
for Multivariate PID Controller Based on RBF Neural
Network, "(2008) Journal of System Simulation, vol. 20, no.
3, pp.627-631.
[5] LI Lei,MAO Zhi-zhong,LIU Fang. "Nonlinear Adaptive
PID Decoupling Controller Based on Multiple Models,
"(2010)Journal of Northe Astern University(Natural
Science), vol. 3, no. 11, pp.54-57.
[6] Narendra, K.S., Zafiriou, E., " Identification and control of
dynamical systems using neural networks,"(1990) IEEE
Trans.Neural Networks, vol. 1, no. 2, pp. 4-27.
[7] Mahmoud F. Hussin, Badr M. Abouelnasr and Amin A.
Shoukry, "Comparative Study of Neural Network
Controllers for Nonlinear Dynamic Systems," (2000)
Lecture Notes in Computer Sciencevol. 1822, pp. 357-
368.
[8] Delgado, A., "Dynamic recurrent neural networks for
systerm identification and control, "(1995) IEE Proc.
Control Theory Applvol. 142, no. 10, pp. 307-314.
[9] Parlos, A.G., Rais, O.T., Atiya, A.F., "Multi-step-ahead
prediction using dynamic recurrent neural network, "(2000)
Neural Networks, vol. 13, no. 4, pp. 765-786.
[10] Qing Yang, Xuanju Dang, " A Multivariable Decoupling
Control Based DRNN, "(2004) Control theory and
application, vol. 23, no. 5, pp. 14-17.
[11] Wenxiu Zhang, Rough set theory and method. Bei jing:
Science Press, 2003, chp.5, pp. 100-145.

Guijin Han was born in henna, China, in
1978. He received the B.Sc. and M.Sc.
degrees from Zhengzhou University in
2000 and Xian University of Technology
in 2006, respectively. He is the author or
coauthor of more than ten national and
international papers. He is currently
working in Xian University of Posts &
Telecommunications and pursuing his
Ph.D. degree in Xian University of Technology. His current
research interests include intelligent control.