MASS: Marginal-Aligned Selective Supervision for Multi-Class Bearing Fault Diagnosis under Label Noise

Abstract Industrial bearing fault diagnosis from vibration data is limited by noisy labels that are asymmetric and class-dependent rather than random. We propose Marginal-Aligned Selective Supervision (MASS) for robust multi-class diagnosis. MASS estimates the class marginal, calibrates predictions against corrupted priors, and selects reliable windows via agreement-based supervision without extra networks. On a composite bogie test bench and the public BJTU-RAO dataset, MASS consistently outperforms uniform cross-entropy and matched selective baselines across seeds, backbones, and noise sweeps. Ablations show marginal distribution alignment drives the gain; selection-precision analysis confirms cleaner sample retention rather than added model capacity. Paired tests on six seeds and a field-proxy weak-label study further clarify when MASS gains hold under controlled versus deterministic corruption.