Conditional feature disentanglement learning for anomaly detection in machines operating under time-varying conditions. (15th May 2023)

• Record Type: Journal Article
• Authors: Zhou, Haoxuan; Lei, Zihao; Zio, Enrico; Wen, Guangrui; Liu, Zimin; Su, Yu; Chen, Xuefeng
• Is Part Of: Mechanical systems and signal processing. Volume 191(2023)
• Publication Date: 2023-05-15
• Subjects: Time-varying operating conditions -- Anomaly detection -- Machines -- Feature disentanglement learning -- Variational autoencoder -- Bearing; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2023.110139
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: Data-driven AD of machines under time-varying operating conditions is considered in this paper. A feature disentanglement learning is proposed to solve entanglement between OCs and health states. A Conditional VAE network combined with PCA is developed to realize the feature disentanglement and AD. Experiments on accelerated fatigue degradation of bearings verify the effectiveness of the proposed method. Abstract: Anomaly detection(AD) is an important task of machines' condition monitoring(CM). Data-driven policies can be used in a more intelligent way to achieve anomaly detection and effectively avoid the introduction of expert experience, thus having a broader scope of application. However, Machines like wind turbines often work under time-varying operating conditions(TVOCs), and the performance of traditional data-driven AD methods is significantly degraded because TVOCs can lead to "false alarms" and "missed alarms" in the implementation due to the monitoring data shifting caused by variation of operating conditions(OCs). To address this problem, this paper proposes a novel conditional feature disentanglement learning framework to solve the disturbance in AD on account of entanglement between OCs and health states. The proposed approach performs conditional self-supervised AD by utilizing the variational autoencoder(VAE) and OCs information. Then, a feature disentanglement conditional VAE(FDCVAE) network is developed to realize the disentanglement of OCs and health states. Subsequently, An anomaly indicator(ANI) is constructed by the dimension reduction of the disentangled health state-related feature and combined with the statistics anomaly threshold for AD. Experiments on accelerated fatigue degradation of bearings under TVOCs validate the effectiveness of the proposed method and further demonstrate the superiority of the constructed ANI in eliminating TVOC interference, compared with common fault mechanisms-based and data-driven ANI. The proposed method not only achieves anomaly detection under TVOCs but also provides a new way for representation learning under variable working conditions in machine health management applications in the foreseeable future.