ARTFEED — Contemporary Art Intelligence

DPR-GM: Domain-Prior-Regularized Graph Modeling for Anomaly Detection in Cyber-Physical Systems

ai-technology · 2026-07-29

A novel approach known as DPR-GM (Domain-Prior-Regularized Graph Modeling) has been introduced for detecting anomalies in cyber-physical systems (CPS). This technique tackles difficulties faced in small-scale physical systems where there is a lack of labeled anomalies and normal data, and where graph-based models may reveal misleading correlations. By utilizing a large language model (LLM), DPR-GM integrates knowledge of system design into the graph-building process, extracting directed physical couplings between sensor pairs from system documentation. These couplings are represented as a binary domain adjacency matrix that regulates sensor relationships, adjusted by Pearson correlations. This forecasting-oriented method seeks to identify subtle deviations from typical behavior indicative of process disruptions. The study is available on arXiv under ID 2607.23197.

Key facts

  • DPR-GM is a forecasting-based framework for anomaly detection in CPS.
  • It uses a large language model (LLM) to extract directed physical couplings from system documentation.
  • The couplings are encoded as a binary domain adjacency matrix acting as a structural gate.
  • The gate is modulated by Pearson correlations.
  • The method targets small-scale physical systems with scarce labeled anomalies and limited normal data.
  • It aims to reduce spurious correlations and unstable sensor topologies in graph-based models.
  • The research is published on arXiv (ID 2607.23197).
  • The approach incorporates system design knowledge into graph construction.

Entities

Institutions

  • arXiv

Sources