Rolling bearing fault diagnosis based on correlated channel attention-optimized convolutional neural networks

• Published in: Measurement science & technology Vol. 35; no. 9; p. 96137
• Main Authors: Jing, Zhu, Ou, Li, Minghui, Chen, Lili, Xing
• Format: Journal Article
• Language: English
• Published: 01.09.2024
• ISSN: 0957-0233; 1361-6501
• DOI: 10.1088/1361-6501/ad5a2e

Abstract In the field of intelligent fault diagnosis, traditional convolutional neural network (CNN)-based models for rolling bearing fault diagnosis are effective in extracting signal features but fall short in identifying subtle fault features in noisy environments. To address this challenge, this paper introduces a correlated channel attention-optimized deep convolutional neural network (CAOCNN) for fault diagnosis. The main innovations of this study include: firstly, the expansion of the convolutional kernel width through dilated convolution and optimized network parameter settings, which broadens the receptive field for feature extraction and effectively suppresses high-frequency noise; secondly, the relevant channel attention mechanism was constructed., which not only considers the channel weights post-global average pooling but also analyzes the correlations between channel features and the global feature center, dynamically adjusting channel weights to enhance model focus on critical features; additionally, the use of the Nesterov momentum optimization algorithm to optimize network parameters, reducing oscillations and increasing efficiency during training. Experimental results demonstrate that the CAOCNN achieved accuracies of 99.71% and 100% on the Case Western Reserve University and Xi’an Jiaotong University rolling bearing datasets, respectively, improving by 2.91% and 7.6% over traditional CNN models. In noisy conditions, T-SNE visual analysis further confirmed the excellent robustness and feature classification capability of the CAOCNN. These achievements validate the effectiveness of the CAOCNN in diagnosing rolling bearing faults in complex noise environments, contributing valuable advancements to the technology of intelligent fault diagnosis.