A Methodology for Progressive Physics Integration in Data-Driven Anomaly Detection: Application to Circuit Breaker Monitoring

Abstract

Physics-informed machine learning has emerged as a promising paradigm for industrial health monitoring, yet practical guidance on when and how to integrate domain knowledge into detection pipelines remains limited. This paper proposes a structured methodology for progressive physics integration in unsupervised anomaly detection, organised into three levels of increasing depth. Level 0 refers to purely data-driven models operating on raw signals. Level 1 injects operational covariates such as temperature or equipment subtype through stratification or conditioning. Level 2 integrates physical knowledge about the internal structure of the signal—its segmentation into electromechanical phases of differing diagnostic relevance—through phase-aware representations, scoring, or end-to-end architectural integration. The methodology is applied systematically to three method families—statistical envelopes, isolation forests and variational autoencoders—for monitoring medium-voltage circuit breaker coil currents, where the breaker’s protection-switching function makes condition monitoring critical. At the deepest level, a physics-informed conditional VAE (PicVAE) injects domain knowledge through phase-segmented inputs, FiLM-conditioned architecture, and a phase-weighted reconstruction loss. Validated on real operational data with expert-labelled anomalies, the results reveal two findings: (i) operational conditioning at Level 1 consistently improves detection across all three method families; (ii) structural physics injection at Level 2 has a method-dependent impact, yielding clear benefits for phase-aware representation learning while introducing trade-offs for simpler models. The PicVAE achieves the best overall performance (AUC-ROC =0.951, Youden J = 0.854). The proposed methodology provides a reproducible template for integrating domain knowledge into anomaly detection pipelines.