Adaptive event-triggered anomaly detection in compressed vibration data

• Published in: Mechanical systems and signal processing Vol. 122; pp. 480 - 501
• Main Authors: Zhang, Yang, Hutchinson, Paul, Lieven, Nicholas A.J., Nunez-Yanez, Jose
• Format: Journal Article
• Language: English
• Published: Berlin Elsevier Ltd 01.05.2019; Elsevier BV
• Subjects: Adaptive learning; Algorithms; Anomalies; Data compression; Datasets; Extreme values; Feature extraction; Health status modeling; Machine faulty detection; Machinery condition monitoring; Malfunctions; Rotating machinery; Rotating machines; Signal compression; Vibration monitoring; Adaptive learning; Signal compression; Health status modeling; Machine faulty detection
• ISSN: 0888-3270; 1096-1216; 1096-1216
• DOI: 10.1016/j.ymssp.2018.12.039

•A hybrid algorithm composed of trend estimation and health status modeling for PCM.•Principal Component Analysis and Multivariate Gaussian Distribution Modeling.•Evaluation on how the performance is influenced by compressed/distorted signal. Anomaly detection is a crucial task in Prognostics and Condition Monitoring (PCM) of machinery. In modern remote PCM systems, data compression techniques are regularly used to reduce the need for bandwidth and storage. In these systems the challenge arises of how the compressed (distorted) vibration data affects the condition monitoring algorithms. This paper introduces a novel algorithm that can adaptively establish normal bounds of operation from continuous noisy vibration profiles working with compressed vibration data. The proposed technique is based on four modules, including feature extraction, feature fusion, extreme value vibration modeling and adaptive thresholding for anomaly detection. The proposed method has been validated with experiments using three time-series datasets. The experimental results indicate that the proposed algorithm is able to perform detection of malfunctions in rotating machines effectively without faulty reference data. Moreover, the proposed method is able to produce accurate early warning and alarm indications from both the raw and compressed (distorted) datasets with equal veracity.