Global–local isolation-based evidence fusion for unsupervised gearbox fault detection

Abstract

Gearbox fault detection under practical operating conditions is challenged by scarce fault labels, heterogeneous sensor measurements, and condition-dependent anomaly patterns. This paper proposes a healthy-only unsupervised framework that combines Isolation Forest (iForest), Isolation-based Nearest Neighbor Ensemble (iNNE), and Dempster–Shafer (D-S) evidence theory. iForest describes global distributional deviations, whereas iNNE captures local neighborhood changes. Their anomaly scores are mapped to basic probability assignments over normal, fault, and ignorance using a cosine S-shaped function. A sample-wise reliability coefficient derived from committed mass and normalized Deng entropy is then used to discount each evidence source before direct D-S combination. The final decision is obtained from Pignistic probabilities and does not require an externally selected anomaly-score threshold for the fused method. Experiments on the Aalto Shim Dataset and the MCC5-THU gearbox dataset cover multiple operating conditions, reduced-channel settings, simple fusion baselines, and natural evidence conflict. The results show consistent benefits from global–local fusion and demonstrate that uncertainty discounting reduces overconfident decisions when the two detectors disagree.