DESIGN AND SIMULATION OF NEUROCONTROLLER FOR PROCESS. CONTROL APPLICATIONS

Abstract

The primary objective of controller is to generate the actuating signal to the final control element of a given physical process to yield its desired response. In classical feedback control our main purpose is to increase the robustness for a control system, i.e., increase the degree to which the system performs when there is uncertainty. Adaptive control techniques have been developed for systems that must performs over large range of uncertainties due to large variation in parameter values, environmental conditions and signal inputs. The objective of the design of an intelligent controller is similar to that for an adaptive control system. The object with intelligent control is to design a system with acceptable performance characteristics over a wide range of uncertainties. The term neurocontroller referes to a neural network based controller. The advantages of neural networks are two fold. One is its learning ability and another is its versatile mapping capabilities from input to output. Versatile mapping capabilities should provide a means of controlling complex systems which can not be carried out well with conventional feedback controllers. The learning ability can reduce human effort in designing controllers and it even suggests in discovering better control schemes than presently known. With neural networks two types of control system architectures are possible. The first type is a self tuning configuration, in which the neural network adjusts the parameters of a conventional controller, but here the overall performance of system depends on the conventional controller limitations. The second type is a simpler configuration in which the neural network serves as a direct feedback controller. Here the controller directly provide the actuating signals to the plant, so it can detect and decide, the control strategy itself during training. In this dissertation work the concept of direct feedback controller is implemented for a first order thermal system and a more generalized second order thermal system using modified genetic algorithm based selftrained neurocontroller. Controller is trained for the performance criterion of minimizing the integral square error (ISE). For the training of controller a self learning scheme is used and network is trained by batch learning process with considering 27 patterns for first 3 order system and 81 patterns for second order system. Standard version of Genetic Algorithm is modified to apply it for neural network training. In genetic algorithm the improvement is triggered by crossover, mutation and natural selection. Simulated systems are analyzed for both cases. A three layer neural network is taken as neurocontroller. For the first order system three neurons are considered in the input layer, five neurons in the hidden layer and one neuron in the output layer. Three inputs to the network are setpoint, feedback signal from the plant and the measured value of fluid temperature at inlet. This configuration works as a form of feedforward control. For the second order system five neurons are considered in the input layer to accomodate five inputs, six neurons are in the hidden layer and one neuron in the output layer. Five inputs to the network are setpoint, two feedback signals ( c(k) and c(k-1)) from the plant, the measured value of fluid temperature at inlet and the measured value of atmospheric temperature. This configuration also works as feedforward control for second order thermal system. In both cases change in fluid temperature at inlet is considered as disturbance. Simulation is performed for three cases without disturbance, with step disturbances and with impulse disturbances. Simulated results show that the controller performance is quite satisfactory in both cases and all variations of inlet temperature and setpoint. A complete user interactive menu driven software package is devloped for the training of controller and to study the response behaviour of controller after completion of training. 4