Multi-view sensor-aware graph contrastive learning for condition monitoring of mechanical transmissions with insufficient labeled data

Recently, self-supervised contrastive learning has attracted increasing attention in machinery health management due to its ability to learn discriminative representations from limited labeled time-series data. With the growing availability of multi-source sensor signals, an important challenge lies in how to jointly model complex temporal dynamics and inter-sensor dependencies in a unified self-supervised framework. Existing methods primarily emphasize temporal consistency while largely overlooking the heterogeneity of individual sensors and the spatial correlations among them. Moreover, complex temporal patterns are often vulnerable to noise and data augmentation, which may lead to unreliable positive pairs and degrade representation quality. To address these issues, we propose a multi-view sensor-aware graph contrastive learning (MSGCL) framework for mechanical fault diagnosis under limited data annotations. The proposed framework learns sensor representations from multiple augmented views and constructs a data-driven sensor graph to capture inter-sensor correlations. By integrating temporal contrastive learning and graph-based contrastive objectives, MSGCL jointly enforces temporal invariance at the sensor level and relational consistency at the graph level. An edge-aware enhancement mechanism is further introduced to emphasize stable and informative sensor dependencies, while a memory-based negative knowledge mining strategy helps mitigate the influence of unreliable positive pairs. Extensive experiments on two transmission system datasets demonstrate that MSGCL consistently outperforms state-of-the-art methods under extremely low-label scenarios. In particular, with only 1% labeled data, the proposed framework achieves up to 5% performance improvement compared with representative baselines, highlighting its effectiveness and robustness for multi-sensor fault diagnosis with limited annotations. © 2026 IOP Publishing Ltd.