关键词: 机电工程；故障诊断；多维分层残差去噪；自校正混合空洞卷积；残差分裂注意力网络；多重残差结构

Abstract: Aiming at the problems of imbalance in the training samples for bearing failure under the background of industrial noise, difficulty in accurate identification of weak early fault features and difficulty in fully mining multi-dimensional correlation features, a fault diagnosis method of rolling bearing based on four-dimensional hierarchical multiple residual neural networks (FH-ResNeMt) was proposed. Firstly, the multi-dimensional hierarchical denoising module (MHDM) was introduced as a primary residual construction unit, and the multi-scale correlation feature extraction and adaptive filtering were realized by effectively integrating self-calibrated hybrid dilated convolution. Secondly, the feature enhanced architecture (FEA) was constructed with Hourglass network as the second residual structure, which solved the problem of information loss caused by feature compression from high-dimensional space to low-dimensional space when processing spatial information transformation. Finally, residual split-attention network (ResNeSt) was integrated to construct cross-channel multi-dimensional feature correlation and make full use of different dimensional features by introducing a third residual structure. The experimental results show that the average accuracy of FH-ResNeMt on JNU and PU rolling bearing data sets reaches 99.84 % and 99.56 %, respectively. The proposed method shows the advantages of rapid convergence and high accuracy on the bearing fault data set (CME) of Chongqing Changjiang Bearing Co., Ltd., which provides important theoretical support for bearing fault analysis and life cycle management.

Key words: mechatronics engineering; fault diagnosis; multi-dimensional hierarchical residual denoising; self-calibrated hybrid dilated convolution; residual split-attention network; multi-residue structure