A Two-Stage Hybrid Architecture for Unknown Anomaly Detection in Rotating Machinery Bearings

Abstract

Supervised fault classifiers frequently encounter difficulties when dealing with faults that are not present in the training data, which is a prevalent issue in industrial contexts. To mitigate this limitation, we introduce a two-stage hybrid framework for diagnosing bearing fault. In the first stage, a supervised classifier assesses the posterior probability of each Known Fault class and predicts a Known Fault when its confidence exceeds a calibrated threshold $\tau $ . Samples with confidence below $\tau $ proceed to the second stage, where a one-class anomaly detector, trained solely on Normal Operation data, evaluates whether the sample aligns with normal behavior (Normal Operation) or deviates from it (Unknown Anomaly). This confidence-based routing produces three diagnostic outcomes: Normal Operation, Known Fault, and Unknown Anomaly, without necessitating retraining of the anomaly detector when new fault types are identified. Using the MAFAULDA dataset, which includes 41 experimental blocks covering closed-set, partial open-set, and fully open-set conditions, we evaluated supervised classifiers, deep learning architectures, and one-class detectors as potential candidates for each stage. Machine learning classifiers demonstrated superior performance compared to the deep learning architectures assessed, with the Isolation Forest proving to be the most effective one-class detector. Integrating the most robust classifier with the Isolation Forest under a calibrated confidence threshold enhanced the identification of previously unseen fault types compared to a baseline relying solely on supervised methods.