An unsupervised learning method for bearing fault diagnosis based on sparse feature extraction

Data-driven based methods have shown their strong feature extraction ability in intelligent fault diagnosis. But most methods always need sample labels in network training process, which is undoubtedly increasing the cost of network training. So developing an unsupervised feature learning method without training labels is desirable. In our work, sparse filtering, an effective unsupervised learning method, is presented for bearing fault diagnosis. The fault features are first extracted from frequency spectra of vibration signals by sparse filtering automatically, and eLU is adopted as activation function for further sparse processing of the features. Then the learned features are input to softmax regression to identify the corresponding fault types. The classification result shows that our method can extract much sparser features and is better than the comparison method.