Neuro-Fuzzy System with Increased Accuracy Suitable for Hardware Implementation

Abstract

Fuzzy controllers are easy to design for complex control surfaces but produce rough control surfaces which might lead to unstable operation. On the other hand neural controllers are hard and complex to train but they produce very accurate output control surfaces compared to that of fuzzy controllers.The Neuro-Fuzzy controller proposed in this thesis exploits the fuzzy systems nature of utilizing expert knowledge and also produces smooth surfaces by implementing it in neural networks.

Monotonic sigmoid membership function is used to make the neural implementation an easy task. Levenberg-Marquardt algorithm(LM) which is used for feed forward networks is implemented to train the neurons. A defuzzification with trigonometric approximation algorithm using LUT-Lookup Table is developed to implement in low cost microcontrollers to make the control system highly cost effective. It is shown through extensive simulations that the proposed model produces accurate and smooth control surfaces.