[attachment=8500]

ABSTRACT
Aim of this project is to implement a PowerPC in process control application. The
process to be controlled can be a simple process like a flow control loop or a complex
process like the boiler feed water control. PowerPC embedded processor will make use
of Vxworks real time OS that is ported into the processor and an application( Process
control application) to control the process loop. The algorithm that is being used to
control the process is called CMNN( cerebral model neural network) algorithm which
makes use of neural networks to implement the intelligent controller.PowerPC and the
process are connected using the serial port that is available for the trainer kit.

CHAPTER 1
Introduction
An important capability of an intelligent system is the ability to improve its future
performance based on the experience within its environment. the concept of learning
is usually used to describe the process by which this capability is achieved. The contol
system can be viewed as a mapping from plant outputs to actuation commands so as
to achive certain control objectives,with learning as the process of modifying this map-
ping to improve future closed-looped system performance.The information required for
learning, that is , the information that is required to correctly generate the desired control
system mapping,is obtained through direct interaction with the plant( and its environ-
ment).thus learning can be used to compensate for the lack of priori design information
by exploiting empirical information that is gained experimentally.
CMNN( cerebral model neural network ) algorithm is being implemented as
an application in Vxworks OS and is being ported into MPC8260 PowerPC embedded
processor.This controller is interfaced into the process loop through the serial port avail-
able in the trainer kit.

CHAPTER 2
Literature Review
2.1 Cerebral Model Neural Network(CMNN)
Conventional feedback strategies can provide adequate performance only when the ef-
fects of nonlinearities are feeble. However most of the industrial process exhibit highly
nonlinear behavior. They may be operated over a wide range of conditioned due to
disturbances and changes in set point. In such situations, conventional PID controllers
must be tuned conservatively in order to provide stable behavior over the entire range
of operating conditions. This may result in serious degradation in performance.
Recently, there has been a resurgence of interest in developing model based control
system. In this method, the process model is explicitly used to predict future behaviour;
it can be implicitly used (inverted) to evaluate the control action in such a way as to
satisfy the controller’s design specifications. The most expensive part of realization
of model based schemes is the development of an appropriate mathematical models.
In many cases it is even impossible to obtain a suitable physically founded process
model due to complexity of the underlying process, or the lack of knowledge of critical
parameters.
An effective alternative to overcome these problems is to use neural networks. Neu-
ral network models are derived from measured input-output data of the plant. They can
approximate quite adequately the relationship between systems inputs and outputs, by
learning without recourse to any prior mathematical formulation. The different neural
network architectures include multi layer back propogation, radial basis functions, Hop-
field neural networks, self organizing maps, cerebellar model neural networks (CMNN),
recurrent networks etc. Among them CerebellarModel Neural Networks (CMNN) have
many attractive features suitable for modeling and control of non-linear system.
In 1975 J.S Albus proposed a new approach to manipulator control: The Cerebellar Model Articulation Controller (CMAC). It is an adaptive system by which control func-
tions for many degrees of freedom operating simultaneously can be computed by refer-
ring to a table rather than by mathematical solution of simultaneous equations. CMAC
or CMNN is a single layer neural network and has many advantages over other neural
network architectures. Computational requirements are very simple and it can repre-
sent any complex functional behaviour. CMNN can be successfully used for nonlinear
systems for which conventional controller design is difficult. In 1991 Filson H.Glanz
and W.Thomas Miller described the application of CMAC in pattern recognition, robot
control and signal processing . Miller proposed to combine CMAC and traditional
controller and have reported very good results in robotic control. According to their
scheme, the control is mainly contributed by a constant gain controller in the early
stage of the control process. As CMAC gradually learns the dynamics of the plant, the
control is shifted from the constant gain controller to CMAC.

2.2 Need for Neural Control
Control of non-linear processes such as the distillation column using conventional
controllers will not yield desirable performance. This is due to the following reasons.
It is very difficult to model a non-linear system. Then to design a controller, we have
to first linearise the system. This linear approximation may not produce good results.
Even though the conventional controller is tuned once, it will not perform properly since
the process is non-linear. At a particular direction of disturbance (say) the controller is
well tuned. But at a different situation the controller parameters may not be suitable ( or
will not be well tuned). In the case of non-linear processes what actually needed is that
at each time the controller is to be tuned and parameters are to be adjusted. Since it is
not practical, conventional controllers will not yield satisfactory performance in terms
of settling time, overshoot, oscillations etc. In order to overcome the above drawbacks
Neural Control is incorporated for controlling non-linear processes efficiently.
The three steps in neural computational process are:
 Development of neural models motivated by biological neurons.
 Models of synaptic connections and structures (ie, network topologies and weights)
 The learning rules (ie, the method of adjusting weights or internal connection