Graph Signal Processing for Fault Localization in Industrial IoT Systems

Abstract

The rapid expansion of the Industrial Internet of Things has revolutionized manufacturing and infrastructure management by enabling unprecedented levels of monitoring and control. However, the increasing complexity and scale of these interconnected networks pose significant challenges for maintaining system reliability and minimizing operational downtime. Fault localization, the process of pinpointing the exact origin of anomalies within a vast network of sensors and actuators, remains a critical bottleneck. Traditional anomaly detection techniques often treat sensor data as independent time series, ignoring the underlying spatial and topological dependencies that characterize physical industrial systems. This paper presents a comprehensive comparative analysis of fault localization methodologies leveraging Graph Signal Processing. By representing the Industrial Internet of Things network as a graph where nodes denote physical devices and edges denote communication or functional relationships, Graph Signal Processing provides a robust framework for analyzing data within its native non-Euclidean domain. We systematically compare traditional time-domain machine learning algorithms with advanced graph-spectral filters and graph neural networks across multiple simulated and real-world industrial datasets. The study evaluates these models based on precision, recall, latency, and robustness to signal noise and topological alterations. The findings demonstrate that methodologies integrating graph spectral features significantly outperform topology-agnostic models in highly correlated fault scenarios, thereby offering a highly scalable and reliable solution for next-generation automated industrial maintenance.