EAMPI-Net: Evidence-Attributed Multi-Prototype Inference for Interpretable Class-Incremental Bearing Fault Diagnosis

Abstract Long-term condition monitoring of rotating machinery must accommodate new classes as previously unseen faults or damage severities emerge during service. Diagnostic models must incorporate these classes while retaining previously acquired fault knowledge. Reliable deployment also requires explicit and phase-comparable explanations of retained-class decisions, enabling changes in their diagnostic evidence to be traced throughout sequential updates. To address these issues, this study develops an Evidence-Attributed Multi-Prototype Inference Network (EAMPI-Net) that couples incremental fault recognition with evidence-level analysis. The network organizes vibration signals into physical response tokens and envelope modulation tokens with explicit physical semantics, while a Transformer encoder models their contextual dependencies. A multi-prototype memory retains compact local representations of learned classes and supports inference over the expanding label space without storing historical vibration signals. For interpretability, target-class token attribution estimates the contribution of each physically defined token to a target-class score, and class-level profiles enable comparison of this evidence across learning phases. Experiments on two bearing test platforms demonstrate stable accumulated-class recognition and consistent advantages over representative incremental learning baselines. Evidence analyses show that the learned profiles are consistent with expected bearing-fault responses, remain stable across phases, and are relevant to model decisions. The resulting framework integrates incremental fault diagnosis with phase-comparable evidence interpretation for rotating machinery.