Manifold-Aware Temperature-Free Prototypical Networks for Few-Shot Open-Set Cross-Domain Fault Diagnosis

Abstract

Vibration-signal-based fault diagnosis in practical rotating machinery and aeroengine systems is often challenged by limited labeled fault samples, variable working conditions, unseen fault modes, and cross-equipment distribution shifts. These common task-level difficulties make few-shot, open-set, cross-domain diagnosis highly challenging. Existing domain adaptation and prototypical learning methods often rely on shared normalization statistics, static source prototypes, and temperature-scaled bounded similarities, which may cause statistical interference, prototype bias, and weak unknown rejection. To address these issues, this article proposes a manifold-aware temperature-free prototypical network (MATF). In particular, domain-specific batch normalization (DSBN) is introduced to decouple source and target statistics during feature extraction. A manifold-aware, dynamic prototype adaptation (MA-DPA) module with a geometric gate is developed to rectify prototypes toward the target manifold while suppressing target-private outliers. Meanwhile, a temperature-free, prototypical contrastive metric maps bounded similarities into an unbounded decision space to enlarge prototype margins without manual temperature tuning. In addition, manifold-aware, partial optimal transport (MA-POT) is employed to align shared source-target manifolds and reduce negative transfer. Extensive experiments on multiple industrial benchmarks validate the superior diagnostic accuracy, few-shot robustness, open-set recognition ability, and cross-domain generalization capability of MATF.