An Anomaly Detection Method for Metro Signal and Control Systems. Issue 10 (2022)

• Record Type: Journal Article
• Authors: Huang, Yibin; Li, Ken Yat Hung; Pei, Cheng; Wang, Lei; Laura Ming, Wai Lau; Cheung, Kevin Man Sing; Lau, KM Kwok Ming; Chan, CP Chin Pang; Ma, Zeya; Peng, Zhijin; Chen, Jingliang
• Is Part Of: IFAC-PapersOnLine. Volume 55:Issue 10(2022)
• Page Start: 1645
• Page End: 1650
• Publication Date: 2022
• Subjects: equipment health management -- equipment logs -- anomaly detection -- semi-supervised learning -- abnormal pattern mining -- rule generation; Automatic control -- Periodicals; 629.805
• DOI: 10.1016/j.ifacol.2022.09.633
• Languages: English
• ISSNs: 2405-8963

Abstract

Abstract: Abnormal pattern detection is the base of automatic failure diagnosis and prediction in heavy equipment health management. Due to large-volume real-time data, high feature dimension, data uncertainty, and heavy dependence on prior knowledge, it is very demanding for traditional models to detect abnormal patterns in logs generated in metro signal and control systems. This paper aims to improve anomaly detection through inducing expert knowledge and to build a bridge between data-driven anomaly detection and rule-based fault detection. Therefore, a novel semi-supervised anomaly detection method is proposed. The method contains two main steps: 1) abnormal pattern mining, and 2) abnormal pattern refinement. Comparing to traditional anomaly detection models, the proposed method possesses three advantages: 1) features are generated through data mining algorithms rather than pre-defined ones; 2) a novel classification model is adapted from one-class support vector machine; 3) human experts can interact with the method by providing new abnormal templates and evaluating the generated rules and cases. The proposed method is implemented in a practical metro signal and control system and its performance is compared with traditional methods. The results demonstrate that the method is practically useful and outperforms the traditional ones in prediction accuracy and domain knowledge coherency.