Unsupervised Feature Learning of Gearbox Fault Using Stacked Wavelet Auto-encoder

Unsupervised feature learning and automatic fault diagnosis of gearbox is still a major challenge. For this purpose, a novel method called stacked wavelet auto-encoder (SWAE) is proposed in this paper. Firstly, wavelet function is adopted as the nonlinear activation function to design wavelet auto-encoder, which can enhance the feature learning ability from the raw vibration signal of gearbox. Secondly, SWAE is constructed with a series of pre-trained wavelet auto-encoders for learning the deep features. Finally, the learned deep features are fed into the softmax classifier for automatic and accurate fault diagnosis. The proposed method is applied to analyze the experimental gearbox signals, and the results show that the proposed method is more effective and robust than the traditional methods.