Hybrid Statistical–Physics Informed Graph Neural Network for Robust Fault Detection and Uncertainty Quantification in Distributed Sensor Systems

Abstract

Distributed sensor systems are essential for environmental surveillance, industrial automation, and smart-infrastructure, yet they frequently suffer from noise, drift, and data loss, leading to unreliable measurements. Traditional models either use statistical techniques lacking robustness or deep neural architectures lacking physical interpretability. In this paper, a Hybrid Statistical–Physics Informed Graph Neural Network (SP-GNN) is proposed to achieve unsupervised anomaly detection in sensor networks. The framework integrates diffusion-driven physical priors directly into the message-passing mechanism of the GNN, allowing representations to respect the underlying spatial physics while remaining data-driven. Bayesian uncertainty quantification using Monte-Carlo dropout enables reliable confidence estimation. Experiments using the Intel Berkeley Research Lab (IBRL) dataset demonstrate the model’s capability to identify physically inconsistent sensor behavior through reconstruction error and predictive uncertainty. The results are visualized using 2-D error–uncertainty plots and 3-D spatial maps, showing that SP-GNN offers trustworthy, interpretable, and physically coherent anomaly detection for distributed sensor systems.