Manifold Sparse Auto-Encoder for Machine Fault Diagnosis

Although the use of deep learning algorithms to find effective features for fault diagnosis has somewhat enhanced of fault classification accuracy, the lack of guidelines and the parameters such as layers of the deep learning architecture and dimension of each hidden-level has limited further improvement. Based on manifold mapping eigenvalues, an optimized deep learning model, called manifold sparse auto-encoder (MSAE) neural network, is constructed to diagnose the machine faults. Two main contributions of this paper can be summarized as (1) every encoding and decoding process is taken as a module to decline the vanishing gradient problem, and (2) the dimension of each hidden layer is determined by the manifold mapping eigenvalues of hidden neurones, whereas the layers of the deep learning architecture are determined by the clustering distribution of features. Gearbox datasets demonstrated that the proposed MSAE can extract better discriminative high-level features and has higher accuracy in machinery fault diagnosis compared with other machine learning methods.