Hierarchical multi-granularity prototype-contrastive learning with Sinkhorn Knopp for open-set fault diagnosis in planetary gearboxes

Deep learning has shown strong potential for diagnosing faults in planetary gearboxes, which are critical components in wind and gas turbines, electric drives, and hybrid vehicle powertrains. However, industrial deployment remains constrained by reliance on initial labels, single-step contrastive objectives, limited modelling of cluster structure, and difficulty detecting previously unseen faults. To address these issues, we propose a stagewise framework, Hierarchical Multi-Granularity Prototype-Contrastive Learning (HMPCL), which maximises the value of limited known-class data by organising representation learning across instance-, subcluster-, and class-level granularities. The framework adopts a multi-stage objective (Stages 0–3), progressing from coarse instance discrimination to refined prototype-guided alignment, while jointly improving intra-cluster compactness and inter-cluster separability under both over-clustered and standard clustering settings. A prototype bank with Sinkhorn-Knopp balanced assignments is introduced to stabilise representation learning, prevent collapse, and strengthen novel fault detection. This hierarchical prototype-based strategy accelerates convergence, improves generalisation, and yields a robust classification system for gearbox fault diagnosis. Experimental results on planetary gearbox datasets across two case studies demonstrate consistently strong diagnostic performance and practical effectiveness.