A Feature Transferring Fault Diagnosis based on WPDR, FSWT and GoogLeNet

Compared with the traditional bearing fault diagnosis methods, the convolution neural network (CNN) can automatically extract features. However, the construction of CNN model usually needs a large dataset, and it is very timeconsuming to train a CNN model. To address this issue, a feature transferring fault diagnosis method is proposed. Firstly, raw signals are decomposed into sub-signals of different frequencies by wavelet packet decomposition, and the subsignals are refactored into a new signal in order to filter noise. Secondly, 2D time-frequency images are constructed by the frequency slice wavelet transform to enhance signal feature. Finally, the proposed model is trained to identify classification. The effectiveness of proposed method is verified on the famous motor bearing data provided by the Case Western Reserve University.