A depthwise separable convolution-based neural network for rolling bearing fault diagnosis

Abstract

A depthwise separable convolution-based neural network (DSCNN) is proposed to achieve deep feature extraction and efficient end-to-end identification for rolling bearing fault diagnosis. DSCNN is mainly stacked by depthwise separable convolution layers and their blocks with residual structures. The depthwise separable convolutions are firstly employed to untie the correlation between spatial dimensions and channel dimensions in the process of convolutional calculation, which enables DSCNN to extract fault features deeply and comprehensively via more depthwise separable convolution layers without efficiency reduction. Furthermore, the residual structures are designed to prevent DSCNN from learning degradation caused by multi-layer network. In addition, other layers are also considered to enhance the capabilities of DSCNN such as the global average pooling for generalization. The experimental results show that the prediction accuracy of the optimized DSCNN can not only reach 96.88–100% higher than the other methods in diagnostic of bearing fault, but also shows better performance in terms of convergence, robustness, sparsity, generalization, etc. Hence, DSCNN is a high-efficient approach to adaptively perform deep feature extraction and fault identification through end-to-end convolutional network, adapting to the development of fault diagnosis techniques in intelligent large-scale rotatory machinery.