Adaptive neuro-fuzzy controller for passive nonlinear systems

The primary idea behind fuzzy control which has proven to be a very successful method, is to build a model of a human control expert that is capable of controlling the plant without the knowledge of the mathematical model of the system. Fuzzy control has a feature by which the control is capable of incorporating expert's knowledge or control rules using linguistic description of the rules. A fuzzy logic controller maps complex nonlinear relations of the control function by a set of IF--THEN rules with associated fuzzy variables described by its membership functions. There are two difficulties with the current fuzzy control methodology. First, a universal fuzzy controller does not exist for all control applications, implying that a seperate set of rule base and membership functions are needed for each individual application. Second, once the rule base and membership functions are developed and implemented for a particular application there is no means of automatically modifying them to changing environment and operating conditions. This means fuzzy logic controller lacks a learning function. Neural network, on the other hand, self-organizes the mapping relationship by learning. So by integrating neural networks and fuzzy logic and with a suitable rule generation mechanism it is possible to overcome these problems. A novel technique which uses two neural networks and a rule generation mechanism which adapts a fuzzy controller is developed here. The adaptation is based on the previous temporal response of the system. One neural network has the ability to identify the pattern of the response and other is used to map the nonlinearity of the system. The rule generation mechanism uses the temporal data to create new fuzzy rules. By combining these an intelligent controller is derived. This algorithm is simulated on several non-linear systems such as desalination process, two-link manipulator and AdeptTwo industrial robot, to evaluate the controller performance. The results show that the adaptive neuro-fuzzy controller successfully self-organizes to improve the dynamic response of the system under consideration. Also a real-time control of a direct drive motor is implemented using a digital signal processor chip and a 486 PC. The result show real-time adaptation in the control architecture. The work has also created, software environment called Dynamic-Fuzzy{dollar}spcopyright{dollar} for self-modification of controller's structure once it has been designed. Other software for multi-layer perception neural network and real time interface have also been developed.