Induction motor fault diagnosis based on deep neural network of sparse auto-encoder

To overcome the drawback of using supervised learning to extract fault features for classification in most of current induction motor fault diagnosis approaches, a deep neural network algorithm is presented, which is realized by the sparse auto-encoder combined with the denoising auto-encoder, to achieve unsupervised feature learning for fault diagnosis of induction motors. Sparse auto-encoder can learn the inherent features and extract the succinct expressions from complex data automatically. In addition, the method of denoising auto-encoder can increase the robustness of feature expression, thus improving the performance of the sparse auto-encoder. The extracted features can then be used to train a neural network classifier and complete the deep neural network construction. The back-propagation algorithm is used for fine-tuning the deep neural network with the purpose of improving the accuracy of fault classification. The "dropout" technique is also introduced into to the entire training process to reduce the prediction error caused by "overfitting". Experimental results have shown that, compared with the traditional back propagation (BP) neural network, the presented deep neural network can realize induction motor fault diagnosis more effectively.