Explainable condition monitoring via probabilistic anomaly detection applied to helicopter transmissions

Condition Monitoring is a delicate task, where one must balance the advantages of data-driven techniques (e.g., high-accuracy predictions) with the reliability of standard methods, such as hand-crafted Indices (e.g., intelligibility, robustness). We see the lack of transparency and need of faulty examples as the main obstacles faced by data-driven methods. Thus, we propose a novel Explainable methodology for Anomaly Detection (AD) exploiting “inliers” only, learning their probability distribution to detect “outliers” at runtime. Our two main contributions lie in probabilistic measures of discrepancy under a Bayesian perspective, leveraging its inherent Quantification of Uncertainties to inform decisions, and an interpretability-oriented model architecture, which is exploited to build ad-hoc explanatory tools. The methodology is validated experimentally for Condition Monitoring against popular AD alternatives on two use cases: a publicly available benchmark, and a real-world Helicopter Transmission dataset collected over multiple years. In both applications, the method proves competitive, as measured by performance indicators on the fault detection task, with respect to state-of-the-art anomaly detection methods, with the added advantage of explainability.