Contrastive Learning with Time Series Imaging for Fault Diagnosis under Limited Labels

Time series analysis is a widely applied and important tool across various fields, including predictive analytics, anomaly detection, and healthcare. However, in the field of fault diagnosis, despite the richness of information in the data, it often exhibits high complexity and lacks clear labels, posing significant challenges for building efficient models. To address this, this paper proposes a fault diagnosis method based on contrastive learning with time series imaging. The imaging process preserves the spatiotemporal characteristics of the original data, enabling more comprehensive feature extraction. By leveraging contrastive learning, we utilize the imaging process as an augmentation strategy to enhance the complexity of positive and negative sample modalities. This approach effectively addresses the limitations of traditional models, which rely on learning from the single modality of raw time series data. Experimental results on the CWRU dataset demonstrate that our method achieves an accuracy of 99.6%, significantly outperforming traditional methods, particularly under availability of labels(AOL) as low as 0.1. These findings validate the effectiveness of integrating time series imaging and contrastive learning in fault diagnosis tasks.