Rolling bearing fault diagnosis based on efficient time channel attention optimized deep multi-scale convolutional neural networks

Abstract

In rolling bearing fault diagnosis, the collected vibration signal has nonlinear and non-Gaussian characteristics, which makes the signal feature extraction incomplete during the feature extraction process, leading to reduced fault diagnosis accuracy. This article proposes a model based on Efficient Time Channel Attention Depth Multi-Scale Convolutional Neural Network (EMCNN) to solve the above problems. This method designs a multi-scale hierarchical expansion strategy in the Multi-Scale Convolutional Neural Network (MSCNN), which can effectively extract different ranges of information from the signal. In addition, the Efficient Time Channel Attention module (E-TCAM) is designed and embedded into the MSCNN to enhance the attention to the important features in both channel and time dimensions, and also to avoid the problem of feature redundancy. Adamax optimization algorithm is used as the optimizer, which realizes the automatic adjustment and optimization of the learning rate and greatly improves the model training efficiency and performance performance. The effectiveness of the method was verified using the datasets from Case Western Reserve University and Xi'an Jiaotong University. By comparing with other diagnostic models, it was verified that the method had a high diagnostic rate and good generalization performance under nonlinear and non-Gaussian complex characteristics.