A novel lightweight relation network for cross-domain few-shot fault diagnosis. (31st May 2023)

• Record Type: Journal Article
• Authors: Tang, Tang; Qiu, Chuanhang; Yang, Tianyuan; Wang, Jingwei; Zhao, Jun; Chen, Ming; Wu, Jie; Wang, Liang
• Is Part Of: Measurement. Volume 213(2023)
• Publication Date: 2023-05-31
• Subjects: Fault diagnosis -- Lightweight network -- Relation network -- Cross-domain -- Few-shot learning; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2023.112697
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: We propose a modified Ghost block, which has better feature extraction capability while reducing the number of model parameters. Then, a lightweight encoder module built from the Ghost block is proposed to improve the relation network. A calibration method based on semi-supervised learning is proposed, which can utilize unlabeled data to alleviate the prototype deviation problem caused by limited data and domain shift. We propose simulating realistic fault diagnosis scenarios by means of few-shot across-domain diagnostic tasks and we also explore ways to better construct meta-tasks in the field of fault diagnosis. Abstract: Recently, the progress of intelligent fault diagnosis shows deep learning-based methods with large data have achieved great success. Nevertheless, in engineering practice, limited labelled data, and various working conditions seriously hinder the widespread application of most deep learning-based fault diagnosis methods. Besides, increasingly complex networks for obtaining powerful feature representation are difficult to deploy in the industry, due to the problem of model efficiency. To address these problems, a novel lightweight relation network (NLRN) is proposed in this paper. The lightweight encoder module in NLRN achieves a strong feature extraction capability with fewer parameters, which means higher model efficiency. Furthermore, a calibration method based on semi-supervised learning is designed to alleviate domain shift due to cross-domain problems, as well as to improve the unreliability of relation networks in few-shot problems. We choose rolling bearings as the research object and three bearing datasets are utilized to demonstrate the effectiveness of the proposed models. The results of our experiment indicate that NLRN has an aptitude to deal with cross-domain few-shot problems. In comparison with other approaches, the proposed method is superior for fault diagnosis under various working conditions with few samples.