Sihan Wang1, Dazhi Wang1, Deshan Kong1, Wenhui Li2, Huanjie Wang2, and Michael Pecht3d
1 School of Information Science and Engineering, Northeastern University, Shenyang, 110819, China
2 School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, and the Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China
3 Center for Advanced Life Cycle Engineering, University of Maryland, College Park, MD, 20742, USA
Abstract:
Rotating machinery fault diagnosis based on deep learning has been successfully applied in modern industrial equipment. However, many existing types of research suffer from two significant deficiencies. First, most deep neural networks are based on a single or same kind of similarity measurement method, which cannot fully exploit the data to extract different levels of feature information. Second, most intelligent fault diagnosis methods can only partially solve the data sparsity and domain shift problems caused by small samples, noise, variable working conditions or compound faults. The model's performance will degenerate rapidly when the above problems occur simultaneously. To address this problem, this paper develops a cross-level fusion neural network method that extracts abundant information on features by calculating spatial-level, channel-level, and second-order statistical information and adaptively fusing the three levels to obtain the final relationship score. First, the signal is input into the embedding module through a Fast Fourier Transform to obtain the feature embedding of the one-dimensional sequence signal. Then, the cross-level metrics learning module calculates the similarity of query sets and support sets at different levels. Finally, the similarities of different levels are fused through the adaptive fusion module to output the final relationship score. The bearing fault diagnosis experiments in the compound variable condition scenario show that the proposed method improves at least 78.53% compared to the traditional deep learning method, at least 3.22% and at most 35.52% compared to multiple few-shot learning methods. In addition, the ablation test analyzes the contribution of different level measurement methods to the model, and the maximum difference between them will reach 32.49%. In summary, the cross-level fusion method can effectively alleviate the data sparsity and domain shift problems.
This article is available online here and to CALCE Consortium Members for personal review.
.