Fault detection and identification using real-time wavelet feature extraction

Development of real-time fault detection and identification technologies will allow a migration, in the respective theater of operation, from expensive scheduled based maintenance to the more efficient, less costly alternative of condition based maintenance. This paper presents successful initial results applying continuous wavelet transforms coupled with conventional neural networks to the development of a real-time fault detection and classification systems. The approach taken results in a general methodology which is shown to work equally well on fault-seeded, helicopter gear-box data and operational data from Navy shipboard pumps. The family of wavelet basis functions are specifically engineered to allow for real-time implementation. The wavelet basis functions have a time-scale decomposition mathematically inspired from biological systems and provides a clustering in feature space which allows for the development of simplified neural network classifiers. Application to various classes of fault data (helicopter and shipboard pump data) resulted in perfect detection, no false alarms with only modest deferral rates