Abstract: The vibration of reciprocating compressor mainly contains multi-source nonlinear pulse signal, it is difficult to extract fault characteristics from the signal with traditional methods. A novel fault diagnosis approach based on nonlinear complexity measure for reciprocating compressor is proposed. The gas valve signals of reciprocating compressor in four different states including normal valve sheets, gap valve sheets, fractured valve sheets and bad spring are used as the experimental data. The signals are denoised with threshold-based wavelet so as to reduce the noise interference. The normalized Lempel-Ziv complexity (LZC) indexes are calculated by using mean symbolization method. The LZC characteristics interval for each state is estimated, and the characteristics of effective value, power spectrum energy and LZC for reciprocating compressor are trained and detected by artificial neural network. The results show that the LZC method can extract the different faults states of reciprocating compressor accurately, which supplies a measure of fault diagnosis and maintenance strategy for reciprocating compressor.

Key words: Artificial neural network, Fault diagnosis, Lempel-Ziv complexity, Reciprocating compressor