Physics-Informed Residual-Based Anomaly Detection and Open-Set Recognition System: A Case Study on Ring Bearings

Le résumé fourni par la source

Industrial anomaly detection often faces a trade-off between the interpretability of physics-based models and the flexibility of data-driven methods. This paper proposes a novel residual-based framework for anomaly detection and recognition in dynamic systems. Our approach employs a Physics-Informed Neural Network trained on healthy data to generate residuals that quantify deviations from expected physical behavior. The approach enables highly effective interpretable anomaly detection by using Extreme Value Theory to define exceedance based on the modeled tail distributions of healthy residual features. A Siamese Neural Network, trained with triplet loss on residual features from known fault types, creates a similarity embedding space that allows open-set recognition of both known and unseen fault groups via K-Nearest Neighbors. The combined approach provides an interpretable pipeline where detection stems from physical model violations and recognition leverages learned residual similarity, enhancing operator understanding and decision-making. We demonstrate the approach on ring bearing data from the MaFaulDa dataset, achieving detection accuracy of 98.5% with 99.27% F1-score, and overall recognition accuracy and F1-score of 82% under open-set conditions.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.