Deep residual shrinkage relation network for anomaly detection of rotating machines. (October 2022)

• Record Type: Journal Article
• Authors: Chen, Zuoyi; Li, Zirui; Wu, Jun; Deng, Chao; Dai, Wei
• Is Part Of: Journal of manufacturing systems. Volume 65(2022)
• Page Start: 579
• Page End: 590
• Publication Date: 2022-10
• Subjects: Anomaly detection -- Rotating machines -- Residual shrinkage network -- Relation network; Manufacturing processes -- Periodicals; Production engineering -- Data processing -- Periodicals; Robots, Industrial -- Periodicals; Production, Technique de la -- Informatique -- Périodiques; Robots industriels -- Périodiques; Electronic journals; 670.42
• DOI: 10.1016/j.jmsy.2022.10.014
• Languages: English
• ISSNs: 0278-6125

Abstract

Abstract: Anomaly detection is an effective method to guarantee the health of rotating machines during their long-term service time. In this paper, a new deep residual shrinkage relation network (DRSRN) is proposed to detect abnormal states of the rotating machines. The proposed method mainly includes residual shrinkage module (RSM) and relation module (RM). The RSM adopts soft threshold to eliminate unimportant specials, so as to generate representations of the input samples. And, the RM is constructed to calculate the relation score between the new sample and healthy sample in order to determine their types. An auxiliary sample library is built to assist the DRSRN to extract more representative and firm health features. Experiments are conducted on real turbine gearbox dataset and run-to-failure bearing dataset. Results demonstrate the effectiveness of the DRSRN method in early anomaly detection; meanwhile, the superiority of the DRSRN method are verified by the comparison with these existing methods. Highlights: A new deep residual shrinkage relation network (DRSRN) is proposed for anomaly detection. A relation module is constructed to reveal the similarity relations between the sample pairs. A residual shrinkage module is designed to effectively eliminate noise-related features. The built auxiliary sample library can assist the DRSRN to mine intra-health category uniqueness.