Multi-Granularity Proxy--Prototype Membership Fusion for Out-of-Distribution Fault Diagnosis
Abstract
While intelligent fault diagnosis performs well on known faults, its in-distribution (ID) assumption can lead to overconfident misclassification of out-of-distribution (OOD) samples, especially when unknown faults share partial structures with known categories. Existing detectors using class prototypes or subclass proxies may fail to balance global category continuity and local fault-mode specificity. We propose multi-granularity proxy–prototype membership fusion (MG-PPMF), which adaptively fuses local subclass-proxy and global class-prototype evidence into a calibrated proxy–prototype membership field whose class-wise values quantify ID support for threshold-based OOD rejection. Experiments on two same-system near-OOD bearing settings, an independent bearing–gearbox dataset, and sensor-noise, speed, and cross-load conditions show competitive performance across independent-dataset near- and far-OOD conditions and delimit robustness under operating and measurement shifts, while retaining an interpretable and lightweight design for open-world fault diagnosis.