Metric-guided graph contrastive learning: an unsupervised approach for few-shot gearbox fault diagnosis

Abstract Planetary gearboxes are critical mechanical components widely deployed in industrial applications such as wind turbines, helicopters, and hybrid vehicles, where their reliable operation directly impacts system performance and safety. Traditional fault diagnosis approaches using graph neural networks and graph contrastive learning (GCL) face significant challenges, including prohibitive costs in fault sample acquisition, ineffective feature extraction from limited data, and semantic distortions in node embedding space that compromise diagnostic accuracy. Furthermore, existing methods struggle with insufficient supervision for complex fault classification and show vulnerability to distribution shifts in new environments. To address these limitations, this research proposes the metric-guided GCL (MGCL) framework, featuring three innovative components: a feature-decoupled pre-training mechanism with graph data augmentation, a sophisticated cosine-Euclidean hybrid distance metric, and a two-stage training paradigm combining unsupervised pre-training with weakly supervised fine-tuning. MGCL significantly advances the field by effectively handling sample scarcity and annotation limitations while enhancing model robustness against domain shifts in real-world industrial applications, ultimately providing a more reliable and practical solution for industrial fault diagnosis.