Comparison between a neural fuzzy system- and a backpropagation-based fault classifiers in a power controller

Abstract

A real-time neural fuzzy (NF) power control system is developed and compared with a backpropagation neural network (BNN) system. The objective is to develop computation hardware and software in order to implement the fault classification of a three-phase motor in real-time response. With online training capability, the NF system can be adaptive to the particular characteristics of a particular motor and can be easily modified for the customer's needs in the future. The preprocessing of a BNN-based fault classifier normalizes the magnitude between [-1,1] and transforms the number of samples to 32 for a cycle of waveform. The trained BNN is used to classify faults from the input waveforms. Real-time response is achieved through the use of a parallel processing system and the partition of the computation into parallel processing tasks. Compared with a four-processor BNN system, the NF system requires smaller cost (three processors) and recognizes waveforms faster. Moreover, with the appropriate feature extraction, the NF system can recognize temporally variant spike and chop occurring within a sin waveform.<>