Integrated method of generalized demodulation and artificial neural network for robust bearing fault recognition

Abstract

Proper functioning of rolling element bearings is critical to ensuring reliable and safe power transmission. The ability to automatically recognize fault-related characteristics is key to enabling intelligent bearing fault recognition. While many techniques have been developed, effective bearing fault recognition under non-stationary conditions remains a challenge. In this paper, a hybrid method that integrates generalized demodulation and artificial neural network is presented that has shown to improve the fault recognition accuracy. Based on the modulation characteristics of bearing vibration signals, a phase function is designed, which allows the mapping of the time-varying modulation rotating frequencies and fault characteristic frequencies into constant frequency components in the demodulation spectrums, thereby eliminating the effect of non-stationarity and facilitating physics-based feature extraction. The features are subsequently classified by an artificial neural network for fault recognition. The physical nature of the features provides the basis for the network to generalize well for unseen non-stationary conditions, and the method has shown to outperform a variety of existing bearing fault recognition techniques in experimental evaluations.