An auto-associative residual based approach for railway point system fault detection and diagnosis. (April 2018)

• Record Type: Journal Article
• Authors: Shi, Zhe; Liu, Zongchang; Lee, Jay
• Is Part Of: Measurement. Volume 119(2018)
• Page Start: 246
• Page End: 258
• Publication Date: 2018-04
• Subjects: Prognostics and health management -- Point machine -- Fault detection and diagnosis -- Auto-associative memory; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2018.01.062
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: An auto-associative residual based approach for FDD is proposed. The KNN classifier performs best together with AAR approach. AAR based approach outperforms the conventional feature based method. The AAR based approach is suitable for downstream system FDD. Abstract: Railway point systems are highly reliable systems the failure of which could lead to significant system delay and have a high chance of causing a fatal accident. It is therefore necessary to develop an online monitoring system to detect incipient failures and prevent faults from happening by applying appropriate maintenance. This paper proposes a novel auto-associative residual (AAR) based approach to evaluate point machine heath condition and diagnose faults from multiple failure modes. The AAR based approach developed in this paper employs auto-associative model to generate residuals from low cost on-board multivariate time series signal, then applies fault detection and diagnosis (FDD) models based on residuals. Commonly used FDD models are applied to evaluate the effectiveness of the proposed approach, including Principal Component Analysis (PCA), Self-organizing Map (SOM), Support Vector Machine (SVM), Naive Bayes Classifier(NBC) and K-Nearest Neighbors (KNN) classifier. Compared with existing approaches, the AAR based approach requires less expert knowledge for model development and minimizes human effort for diagnostic feature extraction. The AAR based approach for FDD achieves more than 97% fault diagnosis accuracy which outperforms existing approaches in the case study.