Fault Diagnosis of Rotating Machinery Base on Wavelet Packet Energy Moment and HMM

A feature extraction method is presented for the fault diagnosis of large rotating machinery to improve the performance of on-line monitoring. According to the characteristics of fault vibration signals, wavelet packet decomposition, a well-known tool to multi-scale analysis, is applied to extracting frequency band energy features; and Hidden Markova Model (HMM) is used to classify. The final feature array is composed of time-domain, amplitude-domain features and wavelet packet energy moment features which reflect the inherent energy distribution characteristics of different faults. The continuous density hidden Markov model (CDHMM) is adopted to recognize the state in on-line monitoring, and the diagnosis success rates are more than 91% to six typical faults on different rotation speeds. The experimental results show the fault diagnosis system is valid and robust, particularly the method of feature extraction.