A multi-channel CNN fault diagnosis method based on squeeze-and-convolution attention for rotating machinery

Abstract

Deep learning methods are increasingly being applied in fault diagnosis because of the capability of representing internal correlations through network structures and extracting hidden features from original data. Convolutional neural networks (CNNs) can perceive local features through convolutional kernels and obtain more advanced features through iteration, which significantly improved the accuracy of fault diagnosis. However, the randomness of feature extraction will cause the insufficient use of the input information and also limiting the network training for specific faults. In this paper, a multi-channel CNN based on Squeeze-and-Convolution attention is introduced to enhance the efficiency of fault diagnosis for rotating machinery. A multi-channel CNN is used to enhance the input information which reduces the input loss, then a new channel attention mechanism based on squeeze-and-excitation module is proposed to reduce the computational complexity of network while focusing on key features. To combine the multi-channel CNN and SC module, the proposed ISC structure enables basic CNN to extract more comprehensive and important features. The fault diagnosis results, which based on two bearing databases named MFPT datasets and Paderborn University datasets, have demonstrated that the ISC-CNN proposed in this study outperforms most commonly used methods. The method exhibits strong performance on both conventional datasets and small sample datasets, which validates its effectiveness.