Hybrid feature selection in fault diagnosis

Abstract

To reduce complexity in design of fault diagnosis system for large scale equipments, a hybrid feature selection algorithm is put forth. By introduction of Markov Blanket, reluctant features can be effectively eliminated to decrease the feature space for input parameters of diagnosis system design. An improved Chl-Square method with introduction of frequency, distribution and concentration is adopted to improve the relevance evaluation performance of the Markov Blanket. The hybrid feature selection algorithm showed high performance in design and implementation of an aero-engine automatic fault diagnosis system based on both neural network and decision tree.