DPR-GM Enhances Anomaly Detection in Cyber-Physical Systems

A novel framework called DPR-GM (Domain-Prior-Regularized Graph Modeling) has been introduced to enhance anomaly detection in multivariate sensor time series, particularly for industrial monitoring of cyber-physical systems (CPS). Traditional graph-based approaches often struggle in small-scale physical systems due to limited labeled anomalies and normal data, leading to spurious correlations and unstable sensor topologies. DPR-GM addresses these limitations by integrating system design knowledge directly into its graph construction process. The framework utilizes a large language model (LLM) to extract directed physical couplings between sensor pairs from system documentation. This information is then encoded into a binary domain adjacency matrix, which acts as a structural gate over the sensor relationships. This gate is further modulated by Pearson correlations derived from normal training data. Additionally, the anomaly score is weighted by sensor-level reliability, calculated from the coefficient of variation. All these graph and weighting components are fixed before training, meaning they introduce no new learnable parameters. Evaluations on the SKAB benchmark demonstrate that DPR-GM significantly outperforms various baselines, including other graph-based, statistical, and deep learning methods, across key metrics like F1, AUROC, and AUPRC. This success highlights that incorporating domain-structured graph priors offers a practical and effective alternative to fully learned topologies, especially in data-scarce CPS environments where robust anomaly detection is critical for preventing process disruptions.