3872 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 56, NO.

10, OCTOBER 2009

PID-Like Neural Network Nonlinear Adaptive

Control for Uncertain Multivariable
Motion Control Systems
S. Cong and Y. Liang

Abstract—A mix locally recurrent neural network was used output (SISO) plants, satisfactory control can be achieved by
to create a proportional-integral-derivative (PID)-like neural net- using established tuning rules. These rules can be applied to
work nonlinear adaptive controller for uncertain multivariable multivariable plants with SISO characteristics as well. Many
single-input/multi-output system. It is composed of a neural net-
work with no more than three neural nodes in hidden layer, and multivariable plants, particularly the single-input/multi-output
there are included an activation feedback and an output feedback, (SIMO) plants, however, show significant internal interaction.
respectively, in a hidden layer. Such a special structure makes the Since the tuning rules are developed for SISO system, their
exterior feature of the neural network controller able to become a applications to such “true” multivariable plants are ineffective.
P, PI, PD, or PID controller as needed. The closed-loop error be- Furthermore, in the case of multivariable plants, the number of
tween directly measured output and expected value of the system
is chosen to be the input of the controller. Only a group of initial PID parameters becomes quite a lot. Trial and error techniques
weights values, which can run the controlled closed-loop system are thus inadequate to derive a good compromise between
stably, are required to be determined. The proposed controller can controller performance and robustness. Second, assume that the
update weights of the neural network online according to errors controller has a PID structure, and find the PID parameters by
caused by uncertain factors of system such as modeling error and using some well-known optimization methods, e.g., H∞ [5],
external disturbance, based on stable learning rate. The resilient
back-propagation algorithm with sign instead of the gradient is mixed H2 /H∞ [6], and semidefinite programming approaches
used to update the network weights. The basic ideas, techniques, [7]. These methods can be used to obtain the PID controller
and system stability proof were presented in detail. Finally, actual parameters such that the controllers have good time-domain
experiments both of single and double inverted pendulums were performance and frequency-domain robustness. The main prob-
implemented, and the comparison of effectiveness between the lem with this approach is that the resulting controllers are state-
proposed controller and the linear optimal regulator were given.
space controllers of high-order rather than low-order controllers
Index Terms—Neural network, nonlinear adaptive control, with a fixed structure. Although one can reduce or approximate
proportional-integral-derivative (PID), single-input/multi-output it with a PID-like structure, it is not so far the reduced-order
(SIMO), uncertain multivariable system.
controller.
With advantages like the abilities of adaptive, parallel, and
I. I NTRODUCTION
fault tolerance, artificial neural networks plays an important

I N industry applications, proportional-integral-derivative

(PID) control is a very popular control strategy due to its
simple architecture and easy tuning. Despite their widespread
role in many fields. From the end of the 1980s, introductions
to neural network control are successively given. Feedforward
controllers in which the neural network learns to mimic the
use and considerable history, PID tuning is still an active area inverse of the plant are intensively discussed by [8]–[11].
of research, both academic and industrial. During the past five Currently, many different neural network control methods are
decades, a comprehensive PID tuning literature has been de- proposed which include the adaptive neural networks to model
veloped. Roughly speaking, there are two different approaches the plant [12], the analytical redundancy method using neural
to obtain PID and PID-like controller parameters. First, tune network modeling of the induction motor in vibration spectra
the parameters of the PID structure by following one of sev- for machine fault detection and diagnosis [13], the method of
eral available tuning techniques: Ziegler–Nichol method [1], in direct adaptive control using neural network model of the
internal-model-control-based method [2], optimization method process and its inverse [14], and adaptive controller with neural
[3], and gain-phase margin method [4]. For single-input/single- network uncertain compensation [15].
In the real world, the system parameters are often time
Manuscript received May 8, 2008; revised October 7, 2008, December 12, varying, so the conventional PID and PID-like control scheme
2008, and January 29, 2009. First published April 7, 2009; current version becomes not flexible in dealing with the systems in which
published September 16, 2009. This work was supported by the National there exist uncertain factors such as modeling error and external
Science Foundation of China under Grant 60774098.
The authors are with the Department of Automation, University of Science disturbance. In addition, some modern control approaches are
and Technology of China, Hefei 230027, China (e-mail: [email protected]; also proposed, but the design and analysis procedures are
[email protected]). complex and difficult. From these factors, the specifications
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. required for real applications of control theories are that the
Digital Object Identifier 10.1109/TIE.2009.2018433 control structures and algorithms should be simple enough to be

0278-0046/$26.00 © 2009 IEEE

CONG AND LIANG: PID-LIKE NEURAL NETWORK NONLINEAR ADAPTIVE CONTROL FOR MOTION CONTROL 3873

and derivative node a3 , respectively. In the output layer, there

is only one neural node, which outputs the control value. The
activation functions in the hidden layer and output layer are all
linear functions. Nodes in input layer and hidden layer are fully
connected. The integral node a1 in hidden layer has an output
feedback, which is realized by delaying the output of a1 with a
unit sampling period and then feeding it back to the weighted
sum node n1 . The derivative node a3 in hidden layer has an
activation feedback, which is realized by negatively feeding
back the output of weighted sum node n3 with a unit sampling
period delay to the place of the input of the node n3 . The
proportional node a2 in hidden layer is a general node without
any feedback.
Fig. 1. Structure of the neural network controller.
According to the structure described above and as shown in
implemented and understood. The control algorithms should Fig. 1, the output of nodes in hidden layer ai (k), (i = 1, 2, 3)
have the following features: learning ability, flexibility, robust- at sampling times k may be calculated, respectively, as follows:
ness, and nonlinearity. A neural network-based PID real-time ⎛ ⎞
tuning method was proposed and used in practical motor control s
a1 (k) = f ⎝ w1j (k)ej (k) + a1 (k − 1)⎠
[16], but the method is only suitable for SISO system. The treat-
j=1
ment of the SIMO case is practically meaningful, since the plant
to be controlled has only one actuator with two or more sensors, 
s

which commonly appear in real control applications to get = w1j (k)ej (k) + a1 (k − 1) (1)
the better control performance by putting extra sensors. Some j=1
⎛ ⎞
results on performance limitations of a single-input two-output 
s 
s
system, special form of the SIMO system, can be found in [17]. a2 (k) = f ⎝ w2j (k)ej (k)⎠ = w2j (k)ej (k) (2)
Early in 2005, Cong et al. [18], [19] proposed a novel multivari- j=1 j=1
able PID-like neural network control scheme for SIMO system, ⎛ ⎞
and then discussed its design method in 2006, but the sufficient 
s 
s

stability analysis of the close-loop system is still lacking. a3 (k) = f ⎝ w3j (k)ej (k) − w3j (k − 1)ej (k − 1)⎠
The main contribution of this paper is that based on a mix j=1 j=1

locally recurrent neural network proposed in the [18], which 

s 
s
was used to create a PID-like neural network adaptive controller = w3j (k)ej (k) − w3j (k − 1)ej (k − 1). (3)
(PIDNNC) for multivariable SIMO systems with uncertainty, j=1 j=1
the sufficient condition of close-loop stability is proved, and
two typical actual motion control system experiments are The final output of neural controller is
implemented.

3
This paper is organized as follows. A novel PID-like neural u(k) = wi (k)ai (k). (4)
network controller scheme for multivariable SIMO systems is i=1
first described in Section II. Then, we derive in Section III
updating weights rules of neural network controller according Taking z −1 as an unit delay factor, we get a(k − 1) = a(k) ·
to the improved resilient back-propagation algorithm. The sta- z −1 , e(k − 1) = z −1 · e(k). From (1)–(4) and the output feed-
bility of closed-loop system is proven in Section IV, in which back, (1) of the first neural node in hidden layer can be
the Lyapunov stability theorem is applied to obtain the stable written as
learning rate of controller. In Section V, we give numerical 
s
applications to both single and double inverted pendulum sys- w1j (k)ej (k)
j=1
tems by using of the proposed controller. The comparison of a1 (k) = (5)
control results without and with disturbance between PIDNNC 1 − z −1
and linear optimal controller are also given. Finally, conclusions from which one can see that (5) has the relationship of integral,
are given in Section VI. so we call a1 (k) is an integral node.
On the other hand, due to the activation feedback, (3) of the
II. N OVEL PID-L IKE N EURAL N ETWORK C ONTROLLER third neural node in hidden layer can be rewritten as
The neural network controller we proposed is composed of ⎛ ⎞
 s
an input layer, a hidden layer, and output layer, as shown in a3 (k) = ⎝ w3j (k)ej (k)⎠ (1 − z −1 ) (6)
Fig. 1. In input layer, there are s input neural nodes, where j=1
s is the number of measurable output variables of controlled
system. In the hidden layer, there are no more than three neural from which one can see that (6) has the relationship of differ-
nodes: we call them integral node a1 , proportional node a2 , ential, so we call a3 (k) a differential node.
3874 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 56, NO. 10, OCTOBER 2009

On the one hand, the basic idea of PID is that the control
action u should be proportional to the error, the integral of the
error over time, and the temporal derivative of the error. From
(4)–(6), one can see that when the neural controller adjusts
input/output relation of the system, it shows the characteristic
of proportional (the second node a2 ) + differential (the third
node a3 ) + integral (the first node a1 ), and it becomes actually a
PID-like controller. The conventional PID controller is usually
only suitable for the system with SISO system and without
consideration of environmental and disturbing factors, while Fig. 2. Structure of the control system with network controller proposed.
the PID-like neural network controller proposed in this paper is
hoped to be suitable for SIMO system and can online regulate The output error of closed-loop system is defined as
controller’s parameters adaptively under the effect of uncertain
and unmodeled factors. ej (k) = rj (k) − yj (k), j = 1, 2, . . . , s. (8)
On the other hand, conventional neural networks can have
It is assumed that yj (k) (j = 1, 2, . . . , s) are output variables
several layers. There are usually two main types of multilayer
of system that are measurable directly by sensors, where j is
networks: feedforward and recurrent ones, in which, recurrent
the number of measurable output variables, and rj (k), (j =
networks have at least one feedback loop. This means an output
1, 2, . . . , s) are given command inputs. Then, the cost function
of a layer feeds back to any proceeding layer. This gives the
to output yj (k) at k is defined as
network partial memory. This makes recurrent networks pow-
erful in approximating function depending on time. The neural 1 1
network controller proposed has two layers, in which the hidden Jj (k) = (rj (k) − yj (k))2 = (ej (k))2 . (9)
2 2
layer play does a multi-input PID function so it has the explicit
signification of PID, while the output layer plays the rule of The total cost function J(k) at k is
sum, so in the controller proposed the weights of output layer

s
are set as J(k) = Jj (k). (10)
j=1
wi = 1, (i = 1, 2, 3). (7)
Equation (10) is the final cost function used in every sam-
pling period.
Although two layers of the network controller here use both Without loss of generality, the rules of updating weights are
the linear activation functions, the mix local recurrent network obtained by means of minimizing the cost function J(k). Here,
used in hidden layer makes the PIDNNC have the properties we adopt the method of gradient descent to obtain updating
there do not exist in the general linear networks or conventional rules. The rule of updating weights in the hidden layer is
local recurrent networks.
⎛ ⎞
Just because it is composed of a neural network with no more s 
∂J(k) ∂Jj (k) ∂yj (k) ⎠ ∂u(k)
than three neural nodes in hidden layer, and in the same hidden =⎝
layer there are an activation feedback and an output feedback, ∂wij (k) j=1
∂yj (k) ∂u(k) ∂wij (k)
respectively, such a special structure of the proposed controller ⎛ ⎞
makes it a P, PI, PD, or PID controller as needed. In such a 
s
∂y (k)
= −⎝ ⎠ ej (k).
j
way, the proposed controller has both the property of simplicity ej (k) (11)
j=1
∂u(k)
of PID control and, at the same time, the neural networks’
powerful capability of self-learning.
Equation (11) is obtained as follows: From (4), one can see that
u is independent
 of j, so ∂u(k)/∂wij (k) can be placed the out
side of the sj=1 (•), and for a fixed i, then according to (1)–(3):
III. U PDATING W EIGHTS R ULES OF P IDNNC AND I TS
One has (∂u(k)/∂uij (k)) = (∂ai (k)/∂wij (k)) − ej (k).
I NITIAL W EIGHTS D ESIGN
Considering the complexity and uncertainty of functional
The structure of closed-loop control system with the neural relation between the output yj (k) and the input u(k), the
controller proposed is shown in Fig. 2. Like general controllers, partial derivative ∂yj (k)/∂u(k) is approximated by using the
the goal of the control system with the PIDNNC is to minimize concept of difference, that is as follows: (∂yj (k)/∂u(k)) ≈
the error between the given command input and the output of (Δyj (k)/Δu(k)) = ((yj (k) − yj (k −1))/(u(k) − u(k − 1))).
closed-loop system. Because physical units used for each con- Moreover, in order not to consider the actual result of
trolled variable in practical system are different (for example calculation the same concept using in the resilient back-
the meter for the distance and meter per second for the veloc- propagation algorithm with sign instead of the gradient value is
ity), the cost function Jj (k) to different controlled variables used to calculate the Δyj (k)/Δu(k). That is to say, the result
yj (k) (j = 1, 2, . . . , s) at k should be defined, respectively, of Δyj (k)/Δu(k) is set to be ±1 according to the positive-
and the total cost function J(k) is the sum of all defined cost negative characteristic of yj (k) − yj (k − 1)/u(k) − u(k − 1).
functions. In such a way, we not only avoid the problem of stopping
CONG AND LIANG: PID-LIKE NEURAL NETWORK NONLINEAR ADAPTIVE CONTROL FOR MOTION CONTROL 3875

updating weights owing to the gradient decent being too small, where
but also greatly simplify the calculation and satisfy the request ⎛ ⎞⎛ ⎞
 s   s
of online control. Therefore, (11) can be rewritten as Δy (k)
q(k) = ⎝ ⎠⎝ e2j (k)⎠ .
j
ej (k)sgn (16)
⎛ ⎞ Δu(k)
s  j=1 j=1
∂J(k) Δy (k)
≈ −⎝ ⎠ ej (k).
j
ej (k)sgn (12)
∂wij (k) j=1
Δu(k) Proof: In order to stabilize the closed-loop control system
with the PIDNNC, the Lyapunov stability theorem is used here.
According to (12), the final rule of updating weights in The Lyapunov function is defined as
hidden layer is
1 2
s

∂J(k) V (k) = e (k). (17)

Δwij (k) = − ηij (k) 2 j=1 j
∂wij (k)
⎛ ⎞

s  Thus, the change of the Lyapunov function is
Δyj (k) ⎠
= ηij (k) ⎝ ej (k)sgn ej (k)
1 2
s
Δu(k)
j=1 ΔV (k) = V (k + 1) − V (k) = e (k + 1) − e2j (k) .
2 j=1 j
i = 1, 2, 3; j = 1, 2, . . . , s (13)
(18)
in which ηij (k) is the learning rate of the hidden layer. According to the structure of PIDNNC in Figs. 1 and 2, one has
The weights in the PIDNNC proposed in this paper need not
be trained a prior. The only thing requested is that all of its 
s 3 
∂ej (k)
initial weights are selected which can stabilize the controlled Δej (k) = Δwil (k) . (19)
∂wil (k)
system. Just because the special structure of the controller pro- i=1
l=1
posed, there exist corresponding relations between the weights
From (1)–(3), one can get
of PIDNNC and those of conventional controllers. In practice,
some simple conventional controllers, such as PID or linear ∂ai (k)
quadratic regulator (LQR) ones, can be first designed for the ∂ej (k) ∂wil (k) el (k)
= ∂ai (k)
= . (20)
controlled system and make it to be stable, whose parameters ∂wil (k) wij (k)
∂ej (k)
can be then used as the initial weights in PIDNNC according to
correspond relationships between them. Equation (13) is used According to (13) and (20), (19) can be rewritten as
to online adjust weights to adapt parameters change caused ⎛ ⎛
by the effects of nonlinearity and uncertainty to get better s  3
control performance of the system and stronger robustness to Δej (k) = ⎝ ⎝ el (k) ηil (k)
wij (k)
some external disturbance. In fact, in the case without (13) or l=1 i=1
Δwij (k) = 0, the PIDNNC is just the conventional controller ⎛ ⎞ ⎞⎞

s 
used to design the initial weights. Δyj (k) ⎠
×⎝ ej (k)sgn el (k)⎠⎠ . (21)
j=1
Δu(k)
IV. S TABILITY OF C LOSED -L OOP S YSTEM
Let
Theorem 1: If all the learning rate ηij (k) at the sampling
s 
 
time k have the same value η(k), and η(k) satisfies the follow- Δyj (k)
p(k) = ej (k)sgn . (22)
ing (14)–(16), then close-loop
 control
 system with the PIDNNC Δu(k)
j=1
in Fig. 2 is stable if ( sj=1 3i=1 (ej (k)/wij (k)))q(k) < 0
then Then, (21) can be rewritten as

s 
3
3 
ej (k) 
s  p(k)e2 (k)
wij (k) l
j=1 i=1 Δej (k) = ηil (k) . (23)
0 < η(k) ≤ −2  2
(14) l=1 i=1
wij (k)

s 
3
1
wij (k) q(k)
j=1 i=1 Let all the learning rate ηij (k) in the sampling time k have the
same value η(k), according to (23), one has
 
if ( sj=1 3i=1 (ej (k)/wij (k)))q(k) ≥ 0 then

3
p(k)  2
s


s 
3 Δej (k) = el (k) η(k)
ej (k)
i=1
wij (k)
wij (k) l=1
j=1 i=1 ⎛ ⎞
−2  2
≤ η(k) ≤ 0 (15) 3 s

s 
3
1 =
1 ⎝p(k) e2j (k) ⎠ η(k). (24)
wij (k) q(k) w (k)
j=1 i=1 i=1 ij j=1
3876 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 56, NO. 10, OCTOBER 2009

Let V. A CTUAL A PPLICATIONS ON I NVERTED

P ENDULUM S YSTEMS
⎛ ⎞

s
In this section, real-time experiments of the control sys-
q(k) = ⎝p(k) e2j (k) ⎠ . (25) tem with PIDNNC are implemented. They are applied to the
j=1
stabilizing controls of single and double inverted pendulum
systems.
Then, (24) can be represented as
A. Real-Time Implementation of PIDNNC Strategy

3
1
Δej (k) = q(k)η(k). (26) Under the description and design above, the proposed control
i=1
wij (k) system can be realized as the following procedure.
1) Determine the number of input layer nodes in PIDNNC
Since according to the number of output variables of the con-
trolled SIMO system.
ej (k + 1) = ej (k) + Δej (k). (27) 2) Choose a suitable controller structure (P, PI, PD, or PID)
through the selection of hidden layer nodes according to
the practical control necessity.
According to (26) and (27), (18) can be rewritten as 3) Initialize the weights wij (0) in hidden layer of PIDNNC
by means of conventional control method such as LQR or
1
s
PID, or neural network training based on the input and
ΔV (k) = ((ej (k + 1) + ej (k)) (ej (k + 1) − ej (k))) output data of the close-loop SIMO system controlled
2 j=1
by any other stabilizing controller. The output layer
1
s weights wi = 1, (i = 1, 2, 3). Let u(−1) = 0, a1 (−1) =
= ((2ej (k) + Δej (k)) Δej (k)) 0, ej (−1) = 0, w3j (−1) = 0.
2 j=1
4) Cascade the controlled SIMO system and PIDNNC to
s 
 form the close-loop SIMO system showed in Fig. 2.
1
= ej (k)Δej (k) + (Δej (k))2 5) In every sample step k(k ≥ 0) when system is running,
2
j=1 do the following.
a) Read the system output yj (k) through A/D sensor.

s 
3
1
= ej (k) q(k)η(k) b) Calculate the error ej (k) as inputs of PIDNNC
j=1 i=1
wij (k) according to (8).
2
⎞ c) Calculate the output of the hidden layer a1 (k), a2 (k),
1 
3
1 and a3 (k) according to (1)–(3).
+ q 2 (k)η 2 (k)⎠ d) Calculate control signal u(k) according to (4).
2 i=1
wij (k)
e) Determine the learning rate ηij (k) according to
⎛ ⎞ Theorem 1.
 s 3
ej (k) ⎠
=⎝ q(k)η(k) f) Calculate Δwij (k) according to (13), and get wij (k +
j=1 i=1
wij (k) 1) through wij (k + 1) = wij (k) + Δwij (k).
⎛ ⎞ g) Output control signal u(k).
2
1⎝ 
s 
3
1 h) k = k + 1.
+ ⎠ q 2 (k)η 2 (k). (28) i) Return to step (1) in 4) and repeat the online operating.
2 j=1 i=1
wij (k)

B. Stabilizing Inverted Pendulum Control System With

According to Lyapunov stability theorem, only if ΔV (k) ≤ 0
PIDNNC Controller
in any sampling time k, the stability of close-loop control
system with PIDNNC will be guaranteed. From (28), one can The Inverted Pendulum is one of the most important classical
concludes the sufficient condition for ΔV (k) ≤ 0 is that η(k) problems of Control Engineering. Broom Balancing (Inverted
satisfies (14)–(16). Pendulum on a cart) is a well-known example of nonlinear
Remark: From conditions (14) and (15) on the one hand, unstable control problem. The problem of balancing an inverted
one can see that the reciprocal of weights wij (k) is needed to pendulum is of the delicate control subset of problems in
calculate. More deep analysis can deduce that no matter how the greater field of robotics: It is almost purely mechanical,
weights are, too great or too small, η(k) can be a reasonable precluding the need for higher level planning or behavioral
value. In practice, usually η(k) can be fixed as an enough small programming that a more complex robot might require. Inputs
value, and it does not calculated at each sampling time in order over (limited) time in the form of pendulum position readings
to save the time. On the other hand, one can see also from (14) are mapped algorithmically to output in the form of motor
and (15) that η(k) even can be selected the negative value. We control. The aim of this paper is to stabilize the Inverted
will do the experiments in both situations. Pendulum such that the position of the carriage on the track
CONG AND LIANG: PID-LIKE NEURAL NETWORK NONLINEAR ADAPTIVE CONTROL FOR MOTION CONTROL 3877

is controlled quickly and accurately so that the pendulum is

always erected in its inverted position during such movements.
Good performance of system comes from the ability of
rejecting external disturbance and uncertainty of system. In
order to verify effectiveness and advantages of PIDNNC, actual
experiments of single and double inverted pendulum systems
are both implemented.
The system devices used in the experiments, which are based
on the inverted pendulum produced by Googol Technology
(Shenzhen) Ltd., are composed of three parts. The first part is
the equipment including a cart, a drive, and links. The second
one is an electrical control box with motor drives and interface
circuit. The third one is a general four-axis movement control
board fixed in a PC.
1) Comparison of Control Performance Between LQR Con-
troller and PIDNNC in Single Inverted Pendulum System: In
this case, variables in a single inverted pendulum system are
two: position of cart x and offset angle of link θ. The inverted
pendulum has two inputs and two outputs. In order to have full
state feedback control, two PID controllers have to be used.
Neural networks have a big advantage here due to their parallel
nature. In such a case, only one PIDNNC can be used instead of
two PID controllers. The initial weight values of PIDNNC are
obtained by LQR whose control law is
u(k) = −KX = −(k1 x + k2 ẋ + k3 θ + k4 θ̇).
The objective of such state regulation may be attained by
selecting the control input u(k) to minimize the following
performance index of the type:
∞

JLQR = X(k)T QX(k) + u(k)T Ru(k)
k=1

where Q ≥ 0 and R > 0 are symmetric weighing matrices.

LQR has been a standard and typical control design which can
be solved easily under MATLAB. Regarding how to design
the feedback matrix K, please refer to the [20]. The feedback Fig. 3. (a) Displacement under LQR and PIDNNC without disturbance.
matrix K is designed as (b) Displacement under LQR and PIDNNC with disturbance. (c) Displacement
under PIDNNC with positive and negative learning rates.
K = [kx kẋ kθ kθ̇ ] = [−21.49 14.79 54.74 9.00].
control of LQR controller, while the equilibrium position of
Because the controller of stabilizing inverted pendulum is a PIDNNC is 0.006 m in the same condition. It is evident that the
regulator, there are ri = 0. The inputs of PIDNNC are > PIDNNC is superior to the LQR. This is because the inverted
pendulum is actually a nonlinear system. Although LQR is an
e1 = r1 − x = −x e2 = r2 − θ = −θ.
optimal controller, it is effective only for the linear system. The
In order to exam the robustness and recover ability of elim- experiment verifies that the PIDNNC is a nonlinear controller.
inating the external disturbance of the system, some external When there is the external disturbance applied to the system,
pulse force with small limited amplitude is added during the as shown in Fig. 3(b), it demonstrates that the single inverted
experiment. The control performances are obtained from ac- pendulum under the control of PIDNNC has much smaller
tual experiment of single inverted pendulum under LQR and displacement and much shorter time of coming back to the
PIDNNC with or without external disturbance. The sampling equilibrium position compared to the LQR. Fig. 3(c) shows the
time of the experiment is 0.005 s. For the experiment imple- displacement responses with the positive and negative learning
mentation simplicity, we selected the fixed stable small learning rates, respectively, from which one can see that system can
rate value as η = 0.001. The experimental results are shown work even better than it works with the positive one, which
in Fig. 3, in which Fig. 3(a) shows the cart displacement satisfies the conditions of (14) and (15) in Theorem 1.
under the control of PIDNNC and LQR controller, respectively, 2) Comparison of Control Performance Between LQR Con-
without external disturbance, from which one can see that when troller and PIDNNC in Double Inverted Pendulum System: In
there is no external disturbance applied to the system, the cart this case, variables in double inverted pendulum system become
stabilizes at the equilibrium position of 0.014 m under the three: x, position of cart; θ1 , offset angle of link 1 (short link);
3878 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 56, NO. 10, OCTOBER 2009

Fig. 4. Structure of PIDNNC used in a double inverted pendulum system.

and θ2 , offset angle of link 2 (long link). The inverted pendulum

has three inputs and three outputs. The LQR control law is
Fig. 5. Control system of double linear inverted pendulum device.
U = −KX = −(k1 x + k2 ẋ + k3 θ1 + k4 θ̇1 + k5 θ2 + k6 θ̇2 )

in which feedback matrix K is designed as

K = [kx kẋ kθ1 kθ̇1 kθ2 kθ̇2 ]

= [32.1522 28.1289 129.3426
1.9087 − 234.6127 − 33.4221].

The values of such a K are used as the initial weight values of

the neural network controller. The structure of neural network
controller is shown in Fig. 4 in which there are three nodes in
input layer, two nodes in hidden layer, and one node in output
layer. Notice that there are only two nodes in hidden layer,
which are proportional node and derivative node. The integral
node a1 (as be shown with dashed line in the Fig. 4) needs not
be used in the given application, which is taught by the LQR
controller.
The inputs to neural network are

e1 = −x e2 = −θ1 e3 = −θ2 .

Initial weight values are

w21 (0) = kx = 32.1522 w31 (0) = kẋ /T = 5625.78
w22 (0) = kθ1 = 129.3426 w32 (0) = kθ̇1 /T = 381.74
w23 (0) = kθ2 = −234.6127 w33 (0) = kθ̇2 /T = −6684.42 Fig. 6. (a) Displacement under LQR and PIDNNC without disturbance.
(b) Displacement under LQR and PIDNNC with disturbance.
in which T is the sampling period and T = 0.005 s. Control
law is the maximum displacement is 0.06 m. While under the control
⎡ ⎤ of LQR controller, the equilibrium position is about 0.07 m,
s and the maximum displacement is 0.08 m. Fig. 6(b) shows the
u(k) = w2 (k) · ⎣ w2j (k)ej (k)⎦ + w3 (k) cart displacement with external disturbance, from which it can
j=1 verified that under the control of PIDNNC, the displacement of
⎡ ⎤ the cart depart from center (or starting position) is smaller and

s 
s
·⎣ w3j (k)ej (k) − w3j (k − 1)ej (k − 1)⎦ . the time of coming back to the equilibrium position is shorter.
j=1 j=1
Owing to the online learning capability, the proposed
PIDNNC can adjust its network weight parameters according
Fig. 5 is the actual experiment device of the double inverted to the changes of system structure parameters and external
pendulum control system. environment. Actual experiments indicate that PIDNNC has
The experimental results of double inverted pendulum sys- stronger antidisturbance ability than LQR does, which will give
tem are shown in Fig. 6, in which Fig. 6(a) shows the cart dis- system better adaptive capacity toward uncertainty of system
placement versus time under the control of LQR and PIDNNC and external disturbance in actual experiment. That is to say,
without external disturbance. It is evident that under the control controlling inverted pendulum system by PIDNNC needs a
of PIDNNC the cart stabilizes in the position of 0.05 m, and shorter horizontal rail. Fig. 6 also shows under the stability
CONG AND LIANG: PID-LIKE NEURAL NETWORK NONLINEAR ADAPTIVE CONTROL FOR MOTION CONTROL 3879

condition, two controllers have almost the same control results [12] D. Wang and J. Huang, “Neural network-based adaptive dynamic surface
of angles. This is reasonable because angles in stable control control for a class of uncertain nonlinear systems in strict-feedback form,”
IEEE Trans. Neural Netw., vol. 16, no. 1, pp. 195–202, Jan. 2005.
region usually do not change greatly. [13] H. Su and H. T. Chong, “Induction machine condition monitoring using
neural network modeling,” IEEE Trans. Ind. Electron., vol. 54, no. 1,
pp. 241–249, Feb. 2007.
VI. C ONCLUSION [14] S. Madhusudan and S. Smriti, “Identification and control of a nonlinear
system using neural networks by extracting the system dynamics,” IETE
An adaptive PID-like neural network controller suitable for J. Res., vol. 53, no. 1, pp. 43–50, 2007.
a large class of SIMO system is proposed in this paper. Rules [15] C. J. Lin and C. H. Chen, “A compensation-based recurrent fuzzy neural
network for dynamic system identification,” Eur. J. Oper. Res., vol. 172,
of updating weights for the controller are derived. The effect no. 2, pp. 696–715, 2006.
of learning rates to stability of closed-loop system is analyzed. [16] T. J. Ren and T. C. Chen, “Motion control for a two-wheeled vehicle
The actual experiments on single and double inverted pendulum using a self-tuning PID controller,” Control Eng. Pract., vol. 16, no. 3,
pp. 365–375, Mar. 2008.
are implemented, and the results are compared with that of [17] T. Bakhtiara and S. Hara, “Regulation performance limitations for SIMO
the LQR, which demonstrates the advantages of controller linear time-invariant feedback control systems,” Automatica, vol. 44,
proposed in simple structure, easy-design, and higher accuracy no. 3, pp. 659–670, Mar. 2008.
[18] S. Cong, G. Li, and B. Ji, “A novel PID-like neural network controller,” in
response performance for multivariable SIMO systems with Proc. 16th IFAC World Congr., Prague, Czech Republic, 2005, pp. 1–6.
disturbance and uncertainty. [19] S. Cong, Y. Liang, and G. Li, “The design of multivariable adaptive
PID-like neural network controller and its design method,” Inf. Control,
vol. 35, no. 5, pp. 568–573, 2006.
R EFERENCES [20] Y. Liang, “Nonlinear adaptive control of time-varying uncertain electro-
[1] J. G. Ziegler and N. B. Nichols, “Optimum settings for automatic mechanical motion system,” Ph.D. dissertation, Univ. Sci. Technol. China,
controllers,” Trans. ASME, vol. 64, pp. 759–768, 1942. Hefei, China, 2008.
[2] M. Morari and E. Zafiriou, Robust Process Control. Englewood Cliffs,
NJ: Prentice-Hall, 1989.
[3] M. Zhuang and D. P. Atherton, “Automatic tuning of optimum PID con-
trollers,” Proc. Inst. Elect. Eng.—Control Theory Appl., vol. 140, no. 3,
S. Cong received the B.S. degree in automatic con-
pp. 216–224, May 1993.
trol from Beijing University of Aeronautics and As-
[4] W. K. Ho, C. C. Hang, and L. S. Cao, “Tuning of PID controllers based
tronautics, Beijing, China, in 1982 and the Ph.D.
on gain and phase margin specifications,” Automatica, vol. 31, no. 3,
degree in system engineering from the University of
pp. 497–502, Mar. 1995.
Rome “La Sapienza,” Rome, Italy, in 1995.
[5] M. T. Ho, “Synthesis of H-infinity PID controllers: A parametric
She is currently a Professor with the Department
approach,” Automatica, vol. 39, no. 6, pp. 1069–1075, 2003.
of Automation, University of Science and Technol-
[6] H. J. Uang and C. C. Lien, “Mixed H2 /H∞ PID tracking control design
ogy of China, Hefei, China. Her research interests
for uncertain spacecraft systems using a cerebellar model articulation
include advanced control strategies for motion con-
controller,” Proc. Inst. Elect. Eng.—Control Theory Appl., vol. 153, no. 1,
trol, fuzzy logic control, neural networks design and
pp. 1–13, Jan. 2006.
applications, robotic coordination control, and quan-
[7] G. Cheng and K. Peng, “Robust composite nonlinear feedback con-
tum system control.
trol with application to a servo positioning system,” IEEE Trans. Ind.
Electron., vol. 54, no. 2, pp. 1132–1140, Apr. 2007.
[8] D. Psaltis, A. Sideris, and A. A. Yamamura, “A multilayered neural
network controller,” IEEE Control Syst. Mag., vol. 8, no. 2, pp. 17–21,
Apr. 1988. Y. Liang received the B.S. degree in automation
[9] T. Pajchrowski and K. Zawirski, “Application of artificial neural network from Southwest University of Science and Technol-
to robust speed control of servodrive,” IEEE Trans. Ind. Electron., vol. 54, ogy, Mianyang City, China, in 2002 and the Ph.D.
no. 1, pp. 200–207, Feb. 2007. degree in control theory and control engineering
[10] S. S. Ge and F. Hong, “Adaptive neural network control of nonlinear from the University of Science and Technology of
systems with unknown time delays,” IEEE Trans. Autom. Control, vol. 48, China, Hefei, China, in 2008.
no. 11, pp. 2004–2010, Nov. 2003. He is currently with the Department of Automa-
[11] C.-H. Lu and C.-C. Tsai, “Adaptive predictive control with recurrent tion, University of Science and Technology of China.
neural network for industrial processes: An application to temperature His research interests include the neural network
control of a variable-frequency oil-cooling machine,” IEEE Trans. Ind. controller design and implementation, motion con-
Electron., vol. 55, no. 3, pp. 1366–1375, Mar. 2008. trol, and robust control for nonlinear system.