Sunteți pe pagina 1din 16

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/281003301

Neural Network Modeling of a Tuned PID Controller

Article · February 2012

CITATIONS READS
5 4,541

2 authors:

Michael Ayomoh Mosud Ajala


University of Pretoria International Islamic University Malaysia
31 PUBLICATIONS   115 CITATIONS    10 PUBLICATIONS   11 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Gas powered firefighting robot applications View project

Robot Vehicle Path Planning View project

All content following this page was uploaded by Michael Ayomoh on 17 August 2015.

The user has requested enhancement of the downloaded file.


European Journal of Scientific Research
ISSN 1450-216X Vol.71 No.2 (2012), pp. 283-297
© EuroJournals Publishing, Inc. 2012
http://www.europeanjournalofscientificresearch.com

Neural Network Modeling of a Tuned PID Controller

M. K. O. Ayomoh
Department of Systems Engineering, University of Lagos
Lagos, Nigeria
E-mail: mikeayomoh@yahoo.com
Tel: +234-803-582-7359

M. T. Ajala
Department of Industrial Maintenance Engineering
Yaba College of Technology, Lagos, Nigeria
E-mail: majala26@yahoo.com
Tel: +234-805-628-1911

Abstract

In this research, a Neuro-PID controller model has been developed to improve on


the response and performance of a conventional Proportionate Integral Derivative (PID)
controller in a nonlinear dynamic environment by developing a self-tuning/adaptive
Neural-PID controller. The proposed Neuro-controller was developed using the back
propagation algorithm. The gradient descent method was employed for the learning rate, to
obtain the initial weight and used in each iteration in order to accelerate the speed of
convergence. Runge-kutta’s algorithm was employed to predict the system behaviour over
time for subsequent weight adaptation. Simulations were conducted on the proposed model
for some specific application areas namely: a mass-spring-damper system and a DC motor.
Results obtained so far from the different simulations have shown that our model is capable
of adapting to dynamic non-linear application domains. It can cope well in the presence of
changing parameters and system disturbances.

Keywords: Neuro-PID, Back propagation, Gradient Descent, Runge-Kutta method,


Feedback, Compensated-Controller

1. Introduction
Despite the recent proliferation of alternative control theories and techniques, a relatively larger
percentage of industrial control processes are still hinged on the proportional-Integral-Derivative (PID)
controllers. This follows the simplicity in structure, robustness in operation and easy comprehension in
principle of the PID controller. Like any other system, the PID controller has its own inherent
limitation which dominantly is its inability to give reliable results in highly non-linear and time varying
application domains. To improve this control performance, several schemes of self-tuning PID
controllers have been proposed in the past however, the artificial neural network currently constitute a
very large research interest.
Various types of modified PID controllers have been developed over the years ranging from the
conventional PID controllers, adaptive/self-tuning controllers to self-tuning predictive PID controllers.
Neural Network Modeling of a Tuned PID Controller 284

A widely used training method for feed-forward multilayer neural networks (MNN) is the back-
propagation (BP) algorithm. Zaheer-uddin and Tudoroiu (2004) adapted the Neuro-PID model to
process control of a laboratory combustor while Yildirim (2004) used neural network in the control of
a bus suspension system. Ali et al. (2006) developed a novel multiple-controller that incorporated the
generalized learning model (GLM) while Savran et al. (2006) developed a neural network (NN) based
adaptive flight control system for a high performance aircraft. Shu and Pi (2005) analyzed the
characteristics of the temperature control systems in industry using the Proportional, Integral and
Derivative neural networks (PIDNN) while Hsu and Lee (2011), proposed an adaptive PID (APID)
controller which is composed of a PID controller and a fuzzy compensator.
Ye (2008), developed an adaptive controller of nonlinear PID-based analog neural networks for
the velocity- and orientation-tracking control of a nonholonomic mobile robot. Beyhan and Alci (2010)
presented a novel model with radial basis functions (RBFs), which was applied successively for online
stable identification and control of nonlinear discrete-time systems. Chen (2011), developed an
intelligent transportation control system (ITCS) using wavelet neural network (WNN) and a PID-type
learning algorithm to increase the safety and efficiency in transportation process. Fang et al. (2010),
presented a mathematical model and a self-tuning PID controller based on the neural network theory to
adjust optimal stabilizer fin angles for reduction of ship roll motion in waves while Shu and Guo
(2004) proposed a new PID neural network (PIDNN) system to decouple and control a strong coupled
time varying system. The paper displayed the perfect performances of the PIDNN in the multivariable
time-varying systems.
Abbassi and Bahar (2005) designed a neural network controller that could effectively replace
PID controllers in thermal systems. Andrasik (2004) developed an adaptive intelligent controller using
neural network while Ahn and Thanh (2005) applied the Nonlinear NeuroPID Control concept to
Improve the Control Performance of Manipulators. In the area of thermal control logics, Moon et al.
(2011) developed three AI-based control logics viz: i) Fuzzy-based control; ii) ANFIS-based (Adaptive
Neuro-Fuzzy Inference System-based) control; and iii) ANN-based (Artificial Neural Network-based)
control. Ab Malek and Mohamed (2009), showed how Metamodeling techniques can be utilized to
tune the PID controller parameters quickly using the linear and nonlinear model. Yazdizadeh et al
(2009), studied two novel adaptive PID-like controllers capable of controlling multi-variable, non-
linear Multi-Input Multiple-Output (MIMO) systems.
Sumar et al (2010), designed a PID controller based on the universal model of the plant in
which there was only one parameter to be tuned. Leva and Maggio (2011) developed a systematic
method to tune an ideal PID and then use its parameters in a real one. Ke et al. (2010) developed two
types of neuro-fuzzy response alleviation control laws for a flexible large-aspect-ratio wing model. In
their paper, Bevrani, et al. (2011), addressed the problem of a new robust control strategy for synthesis
of proportional-integral-derivative (PID) based power system stabilisers (PSS) to find an optimal gain
vector. Hasanien et al. (2010), presented a novel adaptive neuro-fuzzy controller on transverse flux
linear motor for controlling its speed.
Coelho and Grebogi (2010), presented the synchronization of two identical discrete chaotic
systems subject to the different initial conditions by designing a proportional–integral–derivative (PID)
controller. Sharkawy (2010) developed a self-tuning PID control scheme with an application to anti-
lock braking systems (ABS) via combinations of fuzzy and genetic algorithms (GAs) while
Iruthayarajan and Baskar (2010), designed a centralized PID controller using Covariance Matrix
Adaptation Evolution Strategy (CMAES).

Table 1: Table of Notations

S/N SYMBOL DEFINITION S/N SYMBOL DEFINITION


1 e Tracking Error 12 kd Derivative Gain
2 r Set Point 13 ki Integral Gain
3 x State Variable 14 M Moment of Inertia
285 M. K. O. Ayomoh and M. T. Ajala
Table 1: Table of Notations - continued

4 A State Matrix 15 B Bias for the hidden node


5 B Input Matrix 16 P Parameter set for the neural Network
6 C Output Matrix 17 η Learning rate
7 D Transmission Matrix 18 σ Activation function
8 u Control Signal 19 Z Net Sum of a neural network
9 y Actual Output of the plant 20 J Performance Index
10 t Time 21 W Weight connecting the input and the hidden nodes
11 kp Proportional Gain 22 V Weight connecting the hidden and output nodes

2. Problem Definition
The traditional PID algorithm is prone to errors and difficult to deal with in highly non-linear and time
varying environments. Shown below in fig. 1 is the traditional PID controller. The problem at hand is
that of minimizing the error (e) which is a function of the differential between the actual and desired
system response. More often than not, the goal of most control systems is to establish a zero variation
level between the measured plant output and the desired. As shown in fig. 1 this comparison is done at
a junction called the summing point which could be performed by a sensor in real time. The error
generated is fed into the controller which has the sole aim of adjusting the measured output based on
the error generated. This control action performed is sent as a control signal to the plant. The
generalized control diagram in fig. 1 integrates the: summing point which can be represented by an
error detector, where the actual output of the plant is compared with the desired output, PID controller,
and plant or system.

Figure 1: Block diagram of a PID controller

r + e u P la n t


PID
controller
x = A x + Bu
y = Cx
y

where,
r = the set-point also called the reference or desired point
y = the actual output of the plant (also called the measured output)
e = the error between the actual value and the desired value and
u = the control signal
Fig. 1 is represented in (1) and expressed in terms of the desired output, actual output and
system error.
e=r−y (1)
According to Ogata (2004) the control signal of a PID controller could be expressed as shown
below:
∞ d
u = k p e + ki ∫
0
e(t ) dt + kd
dt
(e ) (2a)
hence,
Neural Network Modeling of a Tuned PID Controller 286

d d d2
(u ) = k p e + k i e + k d 2 (e) (2b)
dt dt dt
where,
Kp = proportional gain
Ki = integral gain
Kd = Derivative gain
e = tracking error
u = the control signal (output of the controller)

3. The Governing Equations of a PID Controller


Traditionally, in a state space system representation, two sets of equations are required to describe the
system. They include the state equation and the output equation. According to Ogata (2004) optimal
control problems are better solved using the state space representation. Fig. 2 is a comprehensive
systemic diagrammatic representation from which the state space equations would be derived.

Figure 2: Block diagram of the linear, continuous-time control system represented in state space source: Ogata
(2004)

D(t)

u(t) x(
 t) x(t ) + y(t )
B(t)
∫ dt
+ C(t) +
+

A(t)

For a linear time–invariant system, the equations are thus expressed as the state equations and
the output equation as shown in (3a) and (3b) respectively:
x = Ax + Bu (3a)
y = Cx + Du (3b)
where,
x and u are respectively an n-dimensional state vector and m-dimensional input vector
A = the state/ system matrix
B = the input matrix
C = the output matrix
D = the feed-forward/transmission matrix
y = the output vector
x = state vector
x = change in x
u = input vector
The control diagram in Fig. 2 describes the possible flow scenarios in the control loop of a
controller mechanism. However, following the assumption of Ogata (2004) with the affirmation that
most real systems are feedback dependent, the feed forward element in Fig. 2 is assumed to be a
dummy i.e (D = 0) hence reducing (3b) to:
287 M. K. O. Ayomoh and M. T. Ajala

y = Cx (4)
Substituting (4) in (1) gives
e = r − Cx (5)
Differentiating (1) wrt (t) gives
de dy
= − = − y
dt dt
from (4), y = Cx
de
hence, = − y = −Cx (6)
dt
Substitute (3a) in rhs of (6) and take the second derivative, we have
d 2e du
2
= −CAx − CB (7)
dt dt
du
= −CA2 x − CABu − CB (8)
dt
Substituting equations (6), (5) and (8) respectively into (2b) gives
du
= [ k d CB + 1] −1 ( k i r − x ( k p CA + k i C + k d CA 2 ) − u ( k p CB − k d CAB )) (9)
dt
Equation (9) represents the governing dynamic non-linear equation of the system wrt both the
system state and input. Our proposed Neuro-PID controller is a function of this governing equation.
Equation (9) could further be reduced to the expression:
du
= f ( r , x, u ) (10)
dt

4. Structure of the Neuro-PID Controller


The proposed Neuro-PID controller (NNPIDC) as shown in Fig. 3 consists of the plant, NNPID
controller and the state estimator. This integrated system is proposed for the desired adaptive controller
system.

Figure 3: Proposed Neuro-PID Controller block diagram

r Neuro-PID u Plant
x = Ax + Bu
Controller y
y = Cx
x
State Estimator

A two-layered feedforward neural network (a special case of Multilayer neural network) is


proposed as an integrator to the PID controller. The learning process will be based on a modified back
propagation algorithm (BP).
Neural Network Modeling of a Tuned PID Controller 288
Figure 4: Structure of the Neural Network (MIMO model)

j w i v qi q
x1 ji u1
bi

xn u2

r1

rl um
bH
o u tp u t la y e r
in p u t la y er ( lin e a r u n b ia s e d )
( lin e a r u n b ia s e d ) h id d e n la y e r
( s ig m o id b ia s e d )

The BP will be modified using the 4th order runge-kutta method, a method suitable for most
initial value problem. A generalized structure of the neural network architecture showing the multiple
input multiple output (MIMO) model is as shown in Fig. 4.
where,
th
x j = input to the i neuron in the hidden layer j= 1, 2, - - - n
r k = input to the kth neuron in the hidden layer k = 1, 2, - - l
th
b = bias for the i hidden load i = 1, 2, - - - H
i
th th
w ji = weights connecting the j input node to the i hidden node
viq = the weight connecting the ith neuron of the hidden layer to the qth neuron of the output
layer.
u q = the output from the output layer q = 1,2,3 - - - m
Assuming the following conditions hold i.e. the
dimension of r (the dimension of the system input) = l
dimension of x (the system state) = n
dimension of the output = m
dimension for the controller input = L + n
No of input neuron = l + n
No of output neuron = m
Following the works of Yonghong and Cauwenberghe (1996) and Hecht-Nielsen (1990), where
it has been deduced that a neural net with only one hidden layer is able to represent any function
mapping Rn → Rm, we decided to adopt a single hidden layer neural network. The MIMO generalized
model represented in Fig. 4 now reduces as shown in Fig. 5 below. This follows that our simulation
would be carried out using a single input, single output (SISO) model.
where,
r = single input
vi = the weight connecting the ith neuron of the hidden layer to the output layer
u = the output from the output layer
289 M. K. O. Ayomoh and M. T. Ajala
Figure 5: Neural Network for a SISO model

↓j ↓i
w ji
x1
bi vi
xn
bi u

r
output layer
bH
hidden layer (linear Unbiased )
input layer
( linear U nbiased ) ( sigmoid biased )

The activation function of both the input and the output layer are linear and unbiased while that
of the hidden layer is sigmoid and biased. The generalized form of our proposed integrated (NNPIDC)
is shown in (11) in terms of the output ‘u’ input ‘r’ and state estimator ‘x’
i.e. u = f (r , x, u )
Assuming (11)
where,
U = the output of the neural network

p = parameter set of the network including its weights and biases.
Integrating the neural network model to the conventional PID model in (9) results in:
du
= [ k d CB + 1] −1 ( k i r − x ( k p CA + k i C + k d CA 2 ) − u ( k p CB − k d CAB )U ) (12)
dt
The training of the Neural Network requires that initial values of the variables be known prior
to training iterations. The initial conditions have to be established before the 4th order Runge-kutta
scheme is introduced.
Assuming that the output of the controller is not changing initially
du
i.e. =0 (13)
dt t = 0
Thus (12) reduces to
r (0) − 1 ( k CA + k C + k CA2 ) x − 1 ( k CB + k CAB )U = 0 (14)
k p i d 0 k p d
i i
where, x0 is the initial condition of the state vector.
In solving (14), a performance index (J) is introduced. This is achieved by normalizing the
average of the quadratic form of (14) resulting in:
2

J = 1 r (0 ) − 1 ( k C A + k C + k C A 2 ) x − 1 ( k C B + k C A B )U (15)
2 k p i d 0 k p d
i i
where,
J= teaching signal to be minimized.
The neural network parameters are updated using the Gradient Descent Method i.e.
Neural Network Modeling of a Tuned PID Controller 290

 ∂J
∆p = −η  , (16)
∂p
where,

p is the network parameters which include the weights and biases

i.e. p = {w ji ,Vi , bi } and η is the learning rate
Now we have
∂J ∂N
= −  r (0) − 1 (k pCA + ki C + kd CA2 ) x0 − 1 (k pCB + K d CAB) N  1 (k p CB + K d CAB )  (17)
∂p 
 k i k i 
 ik ∂p
∂N  = derivative of the ANN with respect to its parameters
∂p
The initial weight update for the neural network is achieved using the following equations
H
N = ∑ V σ (Z ) (18)
i i
i =1
n +1
Z = ∑ w x +b (19)
i ji j i
j =1
where,
Zi =net sum of a neural network
Also,
∂N = σ (Z ) (20)
∂V i
i
∂N = V σ ' (Z ) (21)
∂b i i
i
∂N = V σ '( x ) (22)
∂w i j
ji
Equations (17) to (22) can thus be used to train the network initially before subsequent
application of Runge-Kutta’s method.

4.1. Weight Adaptation using Runge-Kutta’s Method


The eqn. for the weight adaptation is achieved using Runge-Kutta’s Algorithm as shown below.
Step 1 Input (ri , xi )
Step 2 Compute
k1 = hf ( ri , xi ), U i )
h k
k 2 = hf [( ri , xi ) + ,U i + 1 ]
2 2
h k
k3 = hf [( ri , xi ) + , U i + 2 ]
2 2
k 4 = hf ( ri , xi ) + h, U i + k3 )
Step 3 Output
k1 + 2 k 2 + 2 k 3 + k 4
U i +1 = U i +
6
where,
i = 1, 2, 3 --- ٨
291 M. K. O. Ayomoh and M. T. Ajala

5. Results and Discussions


The developed NNPID model was validated through the process of simulation using: (i) a mass-spring-
damper system and (ii) a DC motor. Fig.6 shows a generic Neural Network structure for both systems.
Simulations were conducted using MATLAB 7.0 version on a Pentium IV system, 1.6Ghz, 1Gbyte
ram.

Figure 6: Neural Network structure for the MSD and the DC motor.

input
x1
output

x2 y

input layer output layer

hidden layer

The mass spring damper (MSD) system is schematically represented in Fig. 7. In reality, this
could be likened to the shock absorber of an Automobile system. For this system the input layer was
designed to have three neurons as represented in the generic model shown in Fig. 6 viz: u, x1 and x2
each of which respectively represents the applied force, displacement and velocity. The hidden layer
consists of ten neurons while the output layer has a single output represented as y.
Figs. 8 and 9 respectively show the step response and controller action of the uncompensated
plant. The system is having a periodic instability when given a reference input of 1N. This is obvious
from the sinusoidal nature of the generated curves. At this point, the system is highly nonlinear
following that it does not obey either the principles of superposition or homogeneity. The
uncompensated system displayed a very low rise time and a very significant steady state error. This is a
display of the system’s behaviour without the influence of the Neuro controller.

5.1. The Mass-Spring-Damper (MSD) System

Figure 7: Mass-spring-damper system (Richard, 1989)

k
u

y
b
Neural Network Modeling of a Tuned PID Controller 292

The MSD system is defined in terms of the following underlisted parameters:


m = mass of the object
u = external force on the system (input)
y = the displacement (output)
k = spring constant
b = friction resistance
Using the Newton’s second law of motion, we have ∑ F = ma . In the generalized form,
my + by + ky = u (23)
Assuming x1 = y then, x1 = y = x2 (24)
x2 = 
y (25)
Therefore we will only have two state variables namely: x1= displacement and x2 = velocity thus
eqn (23) becomes:
mx2 + bx2 + kx1 = u (26)
b k 1
x2 = − x2 − x1 + u (27)
m m m
recall
x1 = x2 , and
y = x1
From the above, the two governing equations will be:
b k 1
x2 = − x2 − x1 + u (28)
m m m
y = x1 (29)
The matrix representation of eqns 28 and 23 is thus written as:
 0 0  0
 x1   k b   x1   1 


 x  = − m − m  x  + − m u (30)
 2   2   
   
   
with
x 
y = [1 0]  1  (31)
 x2 
Comparing this with the generalized state space equation given in eqns 3a and 3c we conclude
that:
 0 1 0
A= k b , B =  1 , C = [1 0]
− −   
 m m   m 
Substituting the values
b = 10Ns/m, k = 20N/m, m = 1kg and r = 1N, we have A, B and C given as:
 0 1  0
A= , B= , C = [1 0],
 −20 − 10  1
293 M. K. O. Ayomoh and M. T. Ajala
Figure 8: Step response of the uncompensated plant Figure 9: Controller action of the uncompensated plant

Step Response: Plant Controller Action


0.2 2.6

0.18 2.4

0.16 2.2

0.14 2

0.12 1.8
Amplitude

0.1 1.6

0.08 1.4

0.06 1.2

0.04 1

0.02 0.8

0
0 2 4 6 8 10 12 14 16 18 20 0 20 40 60 80 100 120
Time (sec)

The PID gains, after careful design yields the following;


Kp = 500; Kd = 600 and Ki = 100
Fig. 10 shows the learning curve of the neural network architecture, which is meant to improve
the performance of the conventional PID controller. The graph here shows how the performance index
collapsed to zero after 100 iterations where the steady state condition was achieved. After several trials,
the learning rate was chosen as 0.000001.

Figure 10: The learning curve for the initial training Figure 11: Step response of the compensated plant
of the NNPID controller

Initial Training ( Learning Rate = 0.1) Step Response


0.2 1

0.18 0.9

0.16 0.8

0.14 0.7
Performance Index

0.12 0.6
Amplitude

0.1 0.5

0.08 0.4

0.06 0.3

0.04 0.2

0.02 0.1

0 0
0 20 40 60 80 100 120 0 2 4 6 8 10 12 14 16 18 20
Time (sec)

Fig. 11 however shows the step and controller action for the compensated MSD system. It is
glaring that the performance specification for the system has been met. The system has a fast rise time
without any overshoot and attained steady state after 5secs. The above graphs have shown the
effectiveness of the NNPID controller action and how it conforms to the dynamics of the MSD system.

5.2. The DC Motor


The DC motor is a common actuator in control systems. By its action, it directly provides rotary
motion which in turn could be converted to a desired translational motion of a given system. The
circuitry diagram of the armature and the free body diagram of the rotor are shown in Figs. 12 and 13
respectively. Also in this case problem, the following PID parameters are assumed to be valid: Kp =
100; Kd = 10 and Ki = 200.
Neural Network Modeling of a Tuned PID Controller 294
Figure 12: Electric Circuit Armature (Carnegy, 2004) Figure 13: Free body diagram of the rotor (Carnergy,
2004)

The uncompensated motor can only rotate at 0.1 rad/sec with an input voltage of 1V. Some
basic requirements of a motor are viz: (i) rotate at a desired speed and (ii) the motor should be able to
accelerate to its steady-state speed quickly after it is turned on. (iii) steady-state error less than 1% and
further have a settling time of 2 seconds. Since a speed faster than the requirement may damage the
equipment, an overshoot smaller than 5% is desired. From Figs. 12 and 13, we can write the following
equations by combining both Newton and Kirchhoff’s laws as respectively depicted in (32) and (33).
Mθ + bθ = ki (32)
di
L + Ri = V - kθ (33)
dt
where,
M = Moment of inertia of the rotor
b = Damping ratio of the mechanical system
K = Ke = Kt =Electromotive force constant
R = Electric resistance
L = Electric inductance
V = Source Voltage (Input)
θ = rotating speed (the output)

On representing (32) and (33) in state-space form, (34) and (35) emerged. The rotating speed
and electric current are chosen as the state variables while the voltage and rotating speed are
respectively chosen as the input and output components.
 b  K
d θ  −  θ   0 
   +  1 V (34)
J J
= 
dt  i   K  i   
R
− − L
 L  L
 b K   0
θ  −
θ = [1 0]   therefore A =  J J 
, B =  1  , and C = [1 0] (35)
i   
− K −
R  L
 L L 
Substituting the values:
M = 0.01 kg.m2/s2, b = 0.1 Nms, K = 0.01Nm/Amp, R= 1ohm, L = 0.5H give
 10 1  0
A= , B =   , C = [1 0],
 − 0.02 − 5   2
Three neurons V, θ and i were used as input parameters for the input layer while ten neurons
were used in the hidden layer and one for the output layer
295 M. K. O. Ayomoh and M. T. Ajala
Figure 14: Step response of the uncompensated plant Figure 15: Controller Action for the uncompensated
plant
Controller Action Step Response: Plant
0.1 0.1

0.09

0.09
0.08

0.07
0.08
0.06

Amplitude
0.05
0.07
0.04

0.06 0.03

0.02

0.05
0.01

0
0.04 0 0.5 1 1.5 2 2.5 3
0 20 40 60 80 100 120 Time (sec)

where,
V = source voltage,
θ = rotor speed and
i = current.
Figs. 14 and 15 represent the learning curve and controller action of the uncompensated plant.
The system herein is seen to be nonlinear with an endless transient asymptotic curve for the step
response and controller action.

Figure 16: The learning curve for the initial Figure 17: Step response of the plant training of the
compensated NNPID controller
Initial Training ( Learning Rate = 0.001) Step Response
2500 1.4

1.2
2000

1
Performance Index

1500
0.8
Amplitude

0.6
1000

0.4

500
0.2

0 0
0 20 40 60 80 100 120 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (sec)

Fig. 16 shows the learning curve of the neural network collapsed to zero after 100 iterations
After several trials, the learning rate was chosen as 0.001.The controller’s action during the training
period first rises and gets to the peak value. This is followed by a gradually drop in value until the 100th
iteration is reached where it collapses. A relatively fast rise time of 1.5secs was recorded with an
overshoot of 1% however within the accepted tolerance limit and seen to settle at after 5secs.

6. Conclusion and Recommendation


A self-tuning Neuro-PID algorithm was developed and validated through simulation trials using
MATLAB software. This research work has shown the possibility of improving the performance of a
Neural Network Modeling of a Tuned PID Controller 296

conventional PID-based control system without significantly changing the control structure of the PID.
Simulation results have shown that our model is capable of adapting a PID controlled system to
dynamic nonlinear conditions by making it cope in the presence of changing parameters and noise. The
back propagation algorithm was deployed in the training of the neural network. Thereafter the rules of
updating the weights were derived using the fourth order Runge-Kutta method. Our model
demonstrates efficiency by combining Runge-Kutta’s method and feedforward artificial neural
networks which actuates as process models to predict future values of the controlled variables.
The controller showed an efficient result when simulated using the shock absorber i.e. a mass-
spring-damper system and a DC motor. In both case studies, the performance of the proposed Neuro-
PID scheme had superior outputs with respect to set point and settling time attainment. Areas of future
recommendations include testing of our proposed model on pilot or industrial scale to obtain
experimental validation of the scheme.

Acknowledgement
We wish to acknowledge Distinguished Emeritus Prof. V.O.S. Olunloyo for his constructive criticisms
and contributions to this paper.

References
[1] Abbassi A. and Bahar L. (2005), Application of neural network for the modeling and control of
evaporative condenser cooling load, Applied Thermal Engineering, Volume 25, Issues 17-18,
Pages 3176-3186.
[2] Ab Malek, M. N. and Mohamed A. M. S. (2009), “Evolutionary Tuning Method for PID
Controller Parameters of a Cruise Control System Using Metamodeling,”Modelling and
Simulation in Engineering, Article ID 234529, doi:10.1155/2009/234529.
[3] Ahn, K. K. and Thanh, T. D. C., 2005a, “NOnlinear PID Control to Improve the Control
Performance of PAM Manipulators Using Neural Network,” in KSME, Int., Jour., Vol. 19, No.
1, pp. 106~115.
[4] Ali S. Zayed, Amir Hussain and Rudwan A. Abdullah (2006), A novel multiple-controller
incorporating a radial basis function neural network based generalized learning model
Neurocomputing, Volume 69, Issues 16-18, Pages 1868-1881
[5] Andrasik A. Mészáros and De Azevedob S. F. (2004), On-line tuning of a neural PID controller
based on plant hybrid modeling, Computers and Chemical Engineering 28, pg. 1499–1509.
[6] Bevrani, H., Hiyama, T. and Bevrani, H. (2011), Robust PID based power system stabiliser:
Design and real-time implementation International Journal of Electrical Power & Energy
Systems, 33(2), Pp. 179-188
[7] Beyhan, S. and Alci, M. (2010), Stable modeling based control methods using a new RBF
network, Elsevier Ltd.,Vol. 49, Issue 4, pp. 510-518.
[8] Chen, C.H. (2011), Intelligent transportation control system design using wavelet neural
network and PID-type learning algorithms, Expert Systems with Applications, 38(6),pp. 6926-
6939.
[9] Coelho, L. S. and Grebogi, R.B. (2010), Chaotic synchronization using PID control combined
with population based incremental learning algorithm, Expert Systems with Applications, 37
(7), pp. 5347-5352.
[10] Fang, M.C. Zhuo,Y.Z. and Lee, Z.Y. (2010), The application of the self-tuning neural network
PID controller on the ship roll reduction in random waves, Ocean Engineering, Volume 37,
Issue 7, Pp. 529-538
297 M. K. O. Ayomoh and M. T. Ajala

[11] Hasanien, H. M., Muyeen, S.M. and Tamura, J. (2010), Speed control of permanent magnet
excitation transverse flux linear motor by using adaptive neuro-fuzzy controller, Energy
Conversion and Management, 51(12), pp. 2762-2768.
[12] Hecht-Nielsen, R. (1990), Neurocomputing. Addison Wesley, Reading MA.
[13] Hsu, C. and Lee, B. (2011), FPGA-based adaptive PID control of a DC motor driver via
sliding-mode approach, Expert Systems with Applications, Article in Press,
doi:10.1016/j.eswa.2011.02.185, DOI, Copyright © 2011 Elsevier Ltd,
[14] Iruthayarajan, M. W. and Baskar, S. (2010), Covariance matrix adaptation evolution strategy
based design of centralized PID controller, Expert Systems with Applications, Elsevier Ltd,
37(8),Pp. 5775-5781.
[15] Ke, S., Zhigang, W., Chao, Y. and Lei, C. (2010), Theoretical and Experimental Study of Gust
Response Alleviation Using Neuro-fuzzy Control Law for a Flexible Wing Model, Chinese
Journal of Aeronautics, Volume 23, Issue 3, June 2010, Pages 290-297
[16] Leva, A. and Maggio, M. (2011), A systematic way to extend ideal PID tuning rules to the real
structure, Journal of Process Control 21(1), Pp. 130-136.
[17] Moon, J.W., Jung, S.K., Kim, Y. and Han, S. (2011) Comparative Study of Artificial
Intelligence-Based Building Thermal Control Methods-Application of Fuzzy, Adaptive Neuro-
Fuzzy Inference System, and Artificial Neural Network, Applied Thermal Engineering,
Elsevier doi:10.1016/j.applthermaleng.2011.04.006, Copyright © 2011 Elsevier Ltd, Applied
Thermal Engineering, Article in Press, Available online 20 April 2011.
[18] Ogata, K. (2004), Modern Control Engineering, 4th Edition, Pearson Education (Singapore) Pte.
Ltd, India.
[19] Richard C. D. (1989), Modern Control Systems, 5th Edition, Addison-Wesley Publishing
Company.
[20] Savran, A., Tasaltin, R. and Becerikli, Y. (2006), Intelligent adaptive nonlinear flight control
for a high performance aircraft with neural networks. ISA Trans. 45(2):225-247.
[21] Sharkawy, A. B. (2010), Genetic fuzzy self-tuning PID controllers for antilock braking
systems, Engineering Applications of Artificial Intelligence, 23 (7), pp. 1041-1052
[22] Shu, H. and Guo, X. (2004), Decoupling control of multivariable time-varying systems based
on PID neural network Control Conference, 5th Asian Vol.1, pp. 682 - 685
[23] Shu, H. and Pi, Y. (2005), Decoupled Temperature Control System Based on PID Neural
Network, ACSE 05 Conference, CICC, www.icgst.com, Cairo, Egypt. Pp.107-111
[24] Sumar, R. R. Coelho, A. A. R. and Coelho, L.S. (2010), Computational intelligen-ce approach
to PID controller design using the universal model, Information Sciences,180,(20), pp. 3980-
3991.
[25] Yonghong, T. and Cauwenberghe, A. R. V, Optimization techniques for the design of a neural
predictive controller. Neurocomputing 10(1): 83-96 (1996)
[26] Yazdizadeh, A., Mehrafrooz, A., Jouzdani, J. and Barzamini, R. (2009). Adaptive Neuro-PID
controller design with application to nonlinear water level in NEKA power plant. J. Applied
Sci., 9: 3513-3521.
[27] Ye, J. (2008), Adaptive control of nonlinear PID-based analog neural networks for a
nonholonomic mobile robot, Progress in Modeling, Theory, and Application of Computational
Intelligenc - 15th European Symposium on Artificial Neural, Neurocomputing Volume 71,
Issues 7-9, March 2008, pp. 1561-1565.
[28] Yildirim S. (2004), Vibration control of suspension systems using a proposed neural network,
Journal of Sound and Vibration, Volume 277, Issues 4-5, Pages 1059-1069.
[29] Zaheer-uddin M. and Tudoroiu N. (2004), Neuro-PID tracking control of a discharge air
temperature system Energy Conversion and Management, Volume 45, Issues 15-16, Pages
2405-2415.

View publication stats

S-ar putea să vă placă și