The International Arab Journal of Information Technology, Vol. 1, No.

2, July 2004 209

A Hierarchical Neuro-Fuzzy MRAC of a Robot in

Flexible Manufacturing Environment
Kasim Al- Aubidy and Mohammed Ali
Computer Engineering Department, Philadelphia University, Jordan

Abstract: In one hand, the Model Reference Adaptive Control (MRAC) architecture has been widely used in linear adaptive
control field. The control objective is to adjust the control signal in a stable manner so that the plant’s output asymptotically
tracks the reference model’s output. The performance will depend on the choice of a suitable reference model and the
derivation of an appropriate learning scheme. While in the other hand, clusters analysis has been employed for many years in
the field of pattern recognition and image processing. To be used in control the aim is being to find natural groupings among a
set of collected data. The mean-tracking clustering algorithm is going to be used in order to extract the input-output pattern of
rules from applying the suggested control scheme. These rules will be learnt later using the widely used Multi-layer perceptron
neural network to gain all the benefits offered by those nets. A hierarchical neuro-fuzzy MRAC is suggested to control robots
in a flexible manufacturing system. This proposed controller will be judged for different simulated cases of study to
demonstrate its capability in dealing with such a system.

Keywords: MRAC, mean-tracking clustering algorithm, MLP neural nets, computer control, real-time systems, robots, FMS.

Received July 29, 2003; accepted March 8, 2004

1. Introduction controllers are not suitable for real-time FMS

applications. This leads to think about controllers with
Recently, low cost, small and middle production is intelligent capabilities to control the robot's operations
made possible by a Flexible Manufacturing System in uncertain environments.
(FMS). Flexible manufacturing systems represent There are several types of control algorithms that
efficiently grouped machine tools linked together for can be used for joint control of the robot. Some of
batch processing. The FMS consists of work cells, each these use classical controllers, such as Proportional-
cell is responsible of producing a group of parts with Integral-Derivative (PID) [7] and adaptive controllers
similar production processes [1, 16]. FMS is designed [10, 13], others use intelligent controllers based on
to accept raw materials at its input and automatically neural nets [5, 8] and/or fuzzy logic [11, 15].
processes these raw materials into a certain product, Conventional PID controller is still widely used in
which will be delivered at its output. The robotics. The performance of such a controller is not
manufacturing process of these materials may take optimal and its parameters require readjustment, since
place on different work cells. Hence the capability and the joint parameters are varying with time. Several
throughput of these systems are affected by the tuning methods [17] have been published to obtain the
efficiency of the robots that move the product to and controller parameters; however, most of these methods
from these work cells. Moreover, inside each cell require the mathematical model of the robot. The
several machines may share to complete the nonlinear dynamic interactions of the robot joints are
manufacturing process [2]. In this case, the robot will effectively minimized by applying sliding mode
play an important role in delivery, disposal and controllers [14]. Such a controller requires prior
transport systems between cells and machines inside information about the robot parameters.
each cell. This paper deals with design and implementation of
The dynamic equations of the robot are a set of a neuro-fuzzy controller extracted from a model
highly nonlinear differential equations. Therefore, the reference adaptive controller. The resultant functional
movement of the end effecter in a particular trajectory controller is built based on the rules derived from
requires an efficient controller, which generates control applying a certain correcting formula to drift the
signals applied by the robot joint actuators. There are system to behave as close as possible to the selected
many control strategies that can be applied to control model. This can be applied to any joint in the
robot joints. The traditional controllers cannot manufacturing system, which represents the control
effectively control the motion of the robot. A controller activity in hierarchical control in order to make it
based on the theory of the nonlinear control is suitable suitable for real-time applications. The proposed
for the robot control [6]. Unfortunately, such
210 The International Arab Journal of Information Technology, Vol. 1, No. 2, July 2004

controller will improve the system performance by • Balancing precision with complexity.
distributing the control tasks on multilevels. • Multiple time scale operations.
It is assumed that the higher levels in the hierarchy,
2. Hierarchical Architecture that is planning and sequencing, deal with a more
abstract view of the control problem and do so in less
In our previous work [1, 2], the design and precise terms. Moreover, the action-taking place at the
implementation of a hierarchical rout planner for FMS higher levels affects the behavior of the system over a
were proposed, as illustrated in Figure 1. The aim of longer time span whereas the lower levels in the
the FMS rout planner is to obtain the optimal hierarchy operate on a faster time scale.
manufacturing routes for jobs according to well- Figure 3 outlines the general layout of a robot
designed cost function. The sequencing and monitoring system. It consists of a manipulator, and an
module will monitor the competitive jobs to use the input/output interface. A feedback interface is required
manufacturing cells and the required machines and to convert the position sensor signal of each joint into a
robots. Also, this module can reveal the abnormal digital code. The actuating signal generated by the
conditions in the system and generates feedback control algorithm is loaded to the actuator of each joint
signals to the route planner to modify the old through a feedforward interface.
manufacturing routes to avoid the problems that may
PLANNING & SEQUENCING
occur.
FMS
KNOWLEDGE-BASE
ROBOT TRAJECTORY PLANNING

ROUTE PLANNING
AND SEQUENCING

RULE-BASED FUZZY TUNER

ROBOT TRAJECTORY
PLANNING & CONTROL

Figure 1. System organization. LOCAL LOCAL LOCAL

CONTROLLER CONTROLLER CONTROLLER

For an FMS, several tasks of measurement, control,

planning, operator communications, etc. can be JOINT
(n)
JOINT
(2)
JOINT
(1)

distributed among a number of computers linked

together and configured in a hierarchical structure. For Figure 3. Robot system layout.
the proposed system, given in Figure 2, five levels are
recommended, these are:
3. Model Reference Adaptive Control
• Measurement and actuation level: Provides on-line
measurement and actuation database for the whole The MRAC architecture has been widely used in the
system. linear adaptive control field as shown in Figure 4 The
• Control calculations level: Generates the required control objective is to adjust the control signal in a
control signal for each joint. stable manner so that the plant’s output y(t),
• Controller parameters tuning level: Updates the asymptotically tracks the reference model’s output
controller parameters according to the actual ym (t). The performance of this algorithm depends on
behavior of the joint and the required trajectory. the choice of a suitable reference model and the
• Robot trajectory planning level: Determines the derivation of an appropriate learning mechanism.
input commands for each joint according to the Researchers in the sixties found that simple gradient-
robot trajectory. based learning rules were sometimes insufficient and
• Planning and sequencing level: Obtains the optimal there is no reason why this should not also be the case
manufacturing routs for jobs, and then selects the for more general nonlinear plant models and
required manufacturing cells, machines and robots. controllers [3, 4, 12].
Plant
DRIVE CIRCUIT Reference
INPUT/OUTPUT
INTERFACE Model
MEASUREMENT
Learning
JOINT (1)
R Learning Signal Algorithm
O
DRIVE CIRCUIT
INPUT/OUTPUT B Desired
INTERFACE
MEASUREMENT
O O/P
O/P
T Controller PLANT
JOINT (N)

MICROPROCESSOR SYSTEM

Figure 4. Model reference control architecture.

This architecture
Figurehas several features:
2. Hierarchical architecture.
3.1. Control Strategy
• Information abstraction.
A Hierarchical Neuro-Fuzzy MRAC of a Robot in Flexible Manufacturing Environment 211

The main function of the learning algorithm is to of clusters is problem dependent. The mean-tracking
obtain the correct control signal (ud) corresponding to clustering algorithm was derived with the intention of
the desired output (yd ). The difference between the dealing efficiently with operating data collected from
desired response (yd ) and the measured process output high speed production machinery, the data is in the
(y) is called the learning error (eL). It is expected that form of variable information taken from sensors in the
this error will asymptotically approach zero, or a plant, i.e.; variables such as speed, tension,
predefined small region, with increasing number of temperature. In this case data is plotted in an n-
trials. The proposed learning scheme is shown in dimensional space, each data point in the space
Figure 5. corresponds to the machine state at a particular instant
Learning
Learning of time. Natural clusters of data points are then formed;
Error Algorithm all of the points within a cluster depict similar
eL
Ud operating conditions to other points within that cluster.
Reference
Input Plant Yd Y The center of gravity of the search data is then found
Reference + PLANT
Model by finding the mean value of all the points, which lie
-
within the same cluster [18], see Figure 6 for
Control Error
+ Short-Term description. In the proposed functional neuro-fuzzy
Memory
controller the controlled data, input-output variables,
Figure 5. Block diagram of the proposed MRAC system. are collected and clustered based on fuzzy-number
using mean-tracking algorithm. The choice of the
3.2. Learning Algorithm clusters will be clarified in simulation results section
specific ally in equation (9).
The object of the learning control is to determine the
control input u d (t) by repetitive trial such that the error Variable1

asymptotically tends to zero, or a prespecified small center of gravity

of each window
search
value, in the time interval of interest. The following window
error and derivative correction learning algorithm is x x x
x path described by
proposed: x x search method
x x x x
Uk+1 (t)= u k(t) + Pe Lk(t+?) + QeLk(t+ ?) (1) x
x
x x
Where k denotes the instant number, ? is the time x x x x x
x x
advances and p, q are learning gains for error signal x x x x x x
and its derivative respectively. x x
x x x x
x x x x
Variable2
It is noted that the error between the step command
signal and the controlled output y(t) cannot be used as (x)= measured data
a learning basis because such a learning objective (step
Figure 6. The mean-tracking cluster search method.
output) is clearly unrealistic. Therefore, a reference
model is introduced, which specifies an achievable
performance one would like to attain. Then, the 4.2. Controller Implementation
learning error is used as a learning signal; see Figure 5. Depending on the methods for converting qualitative/
Equation (1) is used throughout the simulation for linguistic labels into quantitative/ numerical values, the
Single Input Single Output (SISO), P and Q are just structures of the resulting controllers are significantly
scalar gains. different. Three possible controller input modes can be
The main bottlenecks of this algorithm reside in defined as in the following vectors:
choosing a suitable reference model and the time-
consuming trail and error procedure in finding the Z = [eml ceml, e m2 cem2 ,....., emT cemT] (2)
suitable settings of the learning gains. The complete Z = [eml seml, em2 sem2 ,...., emT semT] (3)
derivations of the previously discussed learning control Z = [eml ceml seml,,. …., emT cemT semT] (4)
algorithms are given by Linkens et al. [9]. where em , cem, and sem are the measured error, change
of error and sum of error. The three input types
4. NeuroFuzzy Controller Design determined by the above representations (2, 3, 4) are
called EC, ES, and ECS respectively. It is noted that
4.1. Mean-Tracking Algorithm they are analogous to classical PD, PI, and PID
Clusters analysis has been employed for many years in controllers respectively.
the field of pattern recognition and image processing, By explicitly embedding the meaning of the
the aim is being to find natural groupings among a set linguistic labels, the control jth rule can be written as:
of collected data. A main problem always is the If em is [Cj (em ), dj(em )] and em is [Cj (cem ), dj (cem )] Then
question of how many clusters there should be within a u is [C j (u), dj (u)] (5)
set of collected data. In practice, however, the number
212 The International Arab Journal of Information Technology, Vol. 1, No. 2, July 2004

where dj (em ), dj (cem ) are the input width for error and the system since it has low gain with a high damping
change-in-error respectively with centers Cj (em ) and ratio; moreover steady state error is detected. Thus the
Cj (cem ), while dj (u) is the width of the control action need is raised to include the effect of proportional,
with center Cj (u). derivative and integral actions (ECS) type.
Now, it is possible to teach a neural network (NN) A model reference of:
with only the central value vectors, i.e; the previous j th Gm(s) = 5.4/ s2 +5.4s +5.4 (8)
linguistic rules become
is chosen, after many trials, so that the robot system
If Cj (em ) and Cj (cem ) Then Cj (u) (6) dynamic can follow such model with an applicable
While leaving the width vectors implicitly treated. One control action values. Applying equation (1) with gains
may ask how the fuzzy concept is handled in such p= 1.0 and q= 0.0, the required data for unit step
paradigm. The answer as concluded before is that the change in input with sampling time of 0.1 sec. have
NN inherently possesses some fuzziness, which is been collected in the short-term memory to be used
exhibited in the form of interpolation over new later.
situations. Using these input-output data, 17-extracted rules are
The clustering criterion based on mean-tracking obtained as shown in Table 1, using the following
algorithm using fuzzy number can be done over the clusters:
vectors defined by equations (2, 3, 4) along with their em = 0.0 to 1.0 step 0.15
corresponding control action (u) to get the centers of cem = -0.85 to 0.0 step 0.15
those variables to be learned using Back-Propagation sem = 1.0 to 9.5 step 1.5 (9)
Neural Network (BNN) of a structure shown in Figure
7. An ECS controller mode of clustered training Table 1. ECS clustered training vector of 17 rules used to train the
vectors, conducted from controlling a process using the BNN of Figure 7.
MRAC system are used to learn BNN of Figure 7. Center Center Center Center of Control
of em of cem of sem Action
W i=1,2, 3
ijj=1.... n 1 0 1 6.484656
em tansh 1 -0.25 2.5 6.291598
W 1 -0.55 2.5 6.021482
j
0.9 -0.85 4 5.814579
U 0.75 -0.85 4 5.65887
cem 0.6 -0.85 5.5 4.701158
0.45 -0.7 7 3.688067
0.45 -0.55 7 2.936129
sem tansh
0.3
0.3
-0.55
-0.4
7
7
2.737491
2.235832
n-node 0.3 -0.4 8.5 2.091233
hidden layer 0.15 -0.4 8.5 1.696041
0.15 -0.25 8.5 1.271224
Figure 7. ECS type BNN represents the proposed functional neuro 0.15 -0.1 8.5 0.805575
controller. 0 -0.1 8.5 0.382604
0 -0.1 9.5 0.116563
The functional neuro-fuzzy controller of Figure 7 0 0 9.5 0.090025
has many advantages over that of MRAC. These can It is important to state here that since the integral
be summarized as follows: action has been included by the accumulation in
• The size of the controller architecture decreases control signal, thus the sum of error will be of no use
dramatically to offer less storage memory, less and it is added just to keep the notation of its existence.
computations time and less fuzziness. Thus more Generally NNs can be used to memorize or discover
efficient controller is available also a less cost the control strategy. Since the control law is extracted
hardware controller, if needed, can be easily based on the MRAC scheme shown if Figure 5, thus
achieved. BNN is applied here to memorize such control rules.
• More robust controller is obtained as will be Applying the error back-propagation learning
verified in the simulation results. algorithm to BNN with the following characteristics:
Topology: 3-node input layer, 12-node tansh nonlinear
5. Simulation Results hidden layer, 1-node linear output layer.
Parameters setting: Random initial weights of values
A robot model of an open loop type (0) second order
between -0.5 to 0.5, steepness= 1.0, threshold= 1.0,
transfer function:
learning rate= 0.1, momentum term= 0.0 and a
G(s) = 1/ s2 + 7.5s + 0.09 (7) stopping criterion of 0.01.
is taken to be control using the proposed control The 17 ESC fuzzy clustered rules shown in Table 1
scheme. The closed loop response to a unit step change are used to train the BNN illustrated by Figure 7.
in input shows the sluggish over damped behavior of Convergence has been reached after 1645 iterations to
A Hierarchical Neuro-Fuzzy MRAC of a Robot in Flexible Manufacturing Environment 213

give final weights set which is inherently representing output (t)

the controller behavior. 1.5

Figure 8 illustrates the two controlled responses of 1

that conventional PID and neuro controllers along with with MRAC
0.5
that of uncontrolled one. The superiority of the neuro time (sec.)
controller can be detected directly. 0
0 5 10 15 20
output(t) -0.5
1.4
with PID -1
1.2 input (t)
input(t)
1 -1.5
0.8 Figure 11. Controlled response tracking to a square wave input.
with MRAC
0.6
uncontrolled
0.4 output (t)
2.5
0.2 time (sec.)
0 2
with MRAC
0 5 10 15 20
1.5
Figure 8. Comparisons between responses.
1 input (t)
Many simulation tests have been achieved as well to
verify the proposed controller capabilities as below: 0.5
time (sec.)
Robustness test: Applying a disturbance of 20% of 0
input value at steady state will not drift the controlled 0 5 10 15 20
response into instability but to a slight acceptable Figure 12. Controlled response tracking to a stair case input.
steady state error of 0.037 value, which is within the
tolerance band as shown in Figure 9. While Figure 10
6. Conclusions
illustrates the stand still controlled response if a time
delay of 1 sec is occurred initially. Many concluded points of high importance can be
output (t) declared as follows:
1 • The complexity of MRAC is in choosing the
effect of input (t)
0.8 with MRAC appropriate model, which the underlying controlled
disturbanc
system must follow.
0.6
• Suitable BNN parameters setting and topology are
0.4 of high importance to gain fast convergence.
0.2 Unfortunately there is no specified setting criterion,
time (sec.) thus a trial and error procedure is applied.
0
0 5 10 15 20 • Representation of the input-output data achieved by
Figure 9. Effect of disturbance on the controlled response. using the mean-tracking algorithm is found to
produce a robust functional neuro controller.
Output(t)
output(t) • The number of the extracted centers should be
1
chosen neither large that gives a meaningless use of
input(t) the clustering criterion nor so small that yields a bad
0.8 with MRAC representation of the original data.
0.6 • Generalization feature offered by neural networks
0.4 gives the flexibility and adaptivity to use the
resultant controller in many applications.
0.2
time (sec.) • A hardware form of the neuro controller can be
0 easily achieved since it will be of small size and low
0 5 10 15 20 cost.
Figure10. Effect of time delay on the controlled response.

Tracking ability: Although the controller is extracted References

based on the unit step change in input, the [1] Al-Titinchi A. A. and Al-Aubidy M. K., “A
generalization feature offered by neural networks Hierarchical Manufacturing Route Planner
gives the advantage of the ability to follow another Design Based on Heuristic Algorithm: Design &
input signal successfully. This is clearly shown in Evaluation,” Systems Analysis Modelling
Figures 11 and 12, which illustrate the good tracking Simulation Magazine, vol. 42, no. 7, pp. 1119-
ability to both square, and staircase waves respectively. 1141, July 2002.
214 The International Arab Journal of Information Technology, Vol. 1, No. 2, July 2004

[2] Al-Titinchi A. A. and Al-Aubidy M. K., [17] Voda A. A. and Landau I. D., “A Method for the
“Modeling an Interactive FMS Scheduler Using Auto-Calibration of PID Controllers,”
Colored Petri Nets,” in Proceeding of 2 nd Middle Automatica, vol. 31, no. 1, pp. 41-55, 1995.
East Conference on Simulation and Modeling [18] Warwick K., “New Ideas in Fuzzy Clustering and
(SCS-MESM’2000), Jordan, pp. 54-61, 2000. Fuzzy Automata,” in Proceeding of the
[3] Astrom K. J. and Wittenmark B., Adaptive International ICSC Symposium on Fuzzy Logic,
Control, Addison Wesley, Reading, MA, 1989. Swiss Federal Institute of Technology (ETH),
[4] Brown M. and Harris C., Neuro-Fuzzy Adaptive Zurich, Switzerland, May 26-27, 1995.
Modeling and Control, Prentice-Hall
International Ltd., UK, 1994. Kasim Al-Aubidy received his BSc
[5] Ge S. S., Lee T. H., Gu D. L., and Woon L. C., and MSc degrees in control and
“A One-Step Solution in Robotic Control System computer engineering from the
Design,” IEEE Robotics and Automation, vol. 7, University of Technology, Iraq in
no. 3, pp. 42-54, 2000. 1979 and 1982, respectively, and the
[6] Jamshidi M., Vadiee N., Rose J. T., and Ross T., PhD degree in real-time computing
Hardware Applications of Fuzzy Logic Control, from the University of Liverpool,
In Soft Computing: Fuzzy Logic, Neural England in 1989. He is currently an associative
Networks, and Distributed Artificial Intelligence, professor in the Department of Computer Engineering
in Aminzadeh F. and Jamshidi M. (Eds), Chapter at Philadelphia University, Jordan. His research
3, Prentice-Hall, USA, 1994. interests include fuzzy logic, neural networks, genetic
[7] Kim S. W. and Lee J. J., “Neural Network algorithm and their real-time applications. He is the
Control by Learning the Inverse Dynamics of winner of Philadelphia Award for the best researcher in
Uncertain Robotic System,” Journal of Institute 2000. He is also a member of the editorial board of the
of Control, Automatic System Engineering, vol. Asian Journal of Information Technology. He has
1, no. 2, pp. 88-93, 1995. coauthored 3 books and published 47 papers on topics
[8] Li Q., Poo A. N., Teo C. L., and Lim C. M., related to computer applications.
“Developing a Neurocompensator for the
Adaptive Control of Robots,” in Proceedings of Mohammed Ali received his BSc,
IEE, Control Theory Applications, vol. 142, no. MSc and PhD degrees in computer
6, pp. 562-568, 1995. and control engineering from the
[9] Linkens D. A. and Nie J., “Constructing Rule - University of Technology, Iraq in
Bases for Multivariable Fuzzy Control by Self- 1981, 1993, and 1998, respectively.
Learning- Part 2,” International Journal of He is currently an assistant professor
Systems SCI, vol. 24, no. 1, pp. 129-157, 1993. in the Department of Computer
[10] Liu M., “An Adaptive Control Scheme for Engineering at Philadelphia University, Jordan. His
Robotic Manipulator,” in Proceedings of 5th research interests include fuzzy logic, neural networks,
International Symposium on Industrial Robotics, image processing, and computer interfacing. He has 6
USA, 1985. published papers related to neurofuzzy and real-time
[11] Palm R., “Fuzzy Controller for Sensor Guided computer control applications.
Robot Manipulator,” Fuzzy Sets and Systems,
USA, 1989.
[12] Psalits D., Sideris A., and Yamamura A. A.,
“Neural Controllers,” in Proceedings of IEEE 1st
International Conference on Neural Networks,
San Diego, CA, vol. 4, pp. 551-558, 1987.
[13] Spong W., Robot Dynamics & Control, John
Wiley & Sons, USA, 1989.
[14] Su C. Y. and Leung T. P., “A Sliding Mode
Controller with Bound Estimation for Robot
Manipulators,” IEEE Transactions Robotic s &
Automations, vol. 9, no. 2, pp. 208-214, 1993.
[15] Tsai C. H., Wang C. H. , and Lin W. S., “Robust
Fuzzy Model Following Control of Robot
Manipulators,” IEEE Transactions Fuzzy
Systems, vol. 8, no. 4, pp. 462-469, 2000.
[16] Viswanadhan N. and Narahari Y., “Performance
Modeling of Automated Manufacturing Systems,”
Prentice-Hall, USA, 1994.