An Unsupervised Framework Based on Dual-Domain Contrastive Learning and Tri-Indicator Joint Alarm Strategy for Early Fault Detection of Bearing

In the realm of industrial maintenance, early fault detection (EFD) in bearing components is pivotal for the seamless operation and longevity of machinery. The complexity of accurately identifying early faults lies in the dynamic, heterogeneous nature of sensor data and the scarcity of high-quality labeled datasets. This scarcity poses a significant challenge, as it hampers the ability to implement effective online fault detection mechanisms, leading to potential downtimes and costly maintenance. To address these challenges, this article introduces an unsupervised framework grounded in dual-domain contrastive learning (DDCL) and tri-indicator joint alarm strategy (TJAS) for the EFD of bearings. First, building upon momentum contrastive learning, we propose a DDCL approach that effectively harnesses both time-domain information and frequency-domain information from a vast amount of unlabeled bearing data to train a feature extractor. Subsequently, a tri-indicator alarm strategy, which jointly utilizes anomaly scores (ASs), permutation entropy (PE), and root mean square values, is designed to process the features outputted by the extractor and identify anomalous time steps. Extensive experiments are conducted on the Prognostics and Health Management (PHM) Challenge 2012 bearing dataset and the XJTU-SY dataset using the proposed framework. Ablation studies indicate that the DDCL effectively captures the differences between normal and anomalous data points, and the TJAS significantly reduces false alarms. Comparative experiments further demonstrate that our method detected early stage faults on average 425 samples ahead of existing approaches and reduced false alarms by an average of 30.