Unsupervised Fault Detection Method via Time-Series Segmentation and Contrastive Masking Learning

The reliability and stability of rotating machinery are critical determinants of industrial productivity and safety. Early fault detection poses significant challenges due to subtle fault signatures that can be masked by base vibrations. Conventional fault detection methods rely on time- and frequency-domain analyses, which often prove inadequate due to their dependence on expert knowledge and the complexity of signal patterns. Emerging deep learning methods are limited by scarce data and a lack of labeled samples, reducing their accuracy. To address these limitations, this study proposes a novel unsupervised framework integrating time-series segmentation with masked contrastive learning. The approach begins with time-series segmentation to preserve health-related features while reducing computational complexity. A transformer-based architecture is designed to implement contrastive mask learning, enabling simultaneous capture of local temporal semantics and global contextual dependencies without labeled data supervision. This combination effectively extracts meaningful features from raw sensor signals without the necessity of labeled data. The efficacy of our method is evaluated on datasets pertaining to rotating machinery. The results demonstrate superior performance compared to conventional reconstruction error-based methods, particularly in early fault detection scenarios with limited fault data, yielding enhanced accuracy and robustness. The framework’s ability to autonomously extract discriminative features from raw signals significantly reduces reliance on manual feature engineering, offering a robust and scalable solution for industrial condition monitoring applications.