Detecção e diagnóstico não supervisionado de falhas mecânicas em máquinas rotativas com explicabilidade (SHAP)

Abstract

In industrial settings, rotating machines are critical assets, and their failures cause unplanned downtime and high costs. Given the scarcity of labeled data, unsupervised anomaly detection is a viable alternative. Even so, it is still necessary to understand why a sample was flagged as anomalous. The proposed approach applies anomaly detection to vibration features and adds explainability for root-cause analysis, as an experimental adaptation of an explainable artificial intelligence framework. Interpretable features were extracted from the signal, and eleven algorithms were trained without fault labels and evaluated using multiple metrics. SHAP and Local-DIFFI were applied to the Isolation Forest. The evaluation used two datasets. On the Mechanical Fault Dataset, the Local Outlier Factor and Feature Bagging achieved the highest F1-scores (0.9600), though close to that of a baseline that classifies all samples as faulty, owing to the strong class imbalance of the test set. The Isolation Forest was therefore adopted as the main model, with greater score separability and SHAP explanations consistent with the underlying physics. On VBL-VA001, the Isolation Forest was the best detector (F1-score of 0.9908 and recall of 1.0). We conclude that unsupervised detection based on harmonic features, coupled with explainability, is promising for fault detection and for supporting root-cause analysis, without replacing expert validation.