Fault diagnosis method integrated on scale-wavelet power spectrum, rough set and neural network

In order to diagnose the faults of complex reciprocating machine, such as internal-combustion engine (ICE), the vibration signals under normal and abnormal patterns were measured by experiments. According to the nonstationarity of ICE signals, the scale-wavelet power spectrum (SWPS) of signals was obtained by continuous wavelet transform (CWT). The wavelet power (WP) distribution on different scales of each pattern is observed to be similar and mainly concentrated in particular scope of 1-32. By analyzing the diversity of SWPS distribution, the WP which is most sensitive to the characteristic of each pattern were extracted by the reduction algorithm of rough set (RS theory) as feature and taken as input to train the back-propagation neural network (BPNN). By the trained BPNN to diagnose the fault signals under detection, the correctness rate is 100%. The fault diagnosis method integrated on SPWS, rS and BPNN demonstrates to be efficient and feasible. It has preferable engineering applicability and referenced value to diagnosis for complex machines.