Abstract
The increasing complexity of critical infrastructure (CI) systems, ranging from smart power grids to urban water management networks, has necessitated a shift from centralized monitoring to decentralized, real-time anomaly detection frameworks. This study proposes a novel Edge-AI driven architecture designed to handle the inherent heterogeneity of modern sensor networks while ensuring low-latency response times and high detection accuracy. By deploying machine learning models at the network edge, we address the bottlenecks associated with cloud-based processing, such as bandwidth constraints and latency-induced vulnerabilities. Our methodology integrates spatio-temporal correlation algorithms and fault-tolerant intrusion detection systems to identify subtle deviations in sensor data that may signify cyber-physical attacks or systemic failures. We evaluate our framework using the Edge-IIoTset and simulated data from SCADA environments, focusing on protocols such as Modbus, MQTT, and BACnet. The results demonstrate that our Edge-AI approach achieves an F1-score of 0.97 across diverse attack vectors while reducing data transmission overhead by approximately 64% compared to centralized models. Furthermore, the integration of spatio-temporal analysis significantly improves the detection of low-intensity anomalies that often bypass traditional threshold-based systems. This research provides a scalable solution for securing critical infrastructure in the burgeoning era of 6G and digital twins, offering a robust defense mechanism against sophisticated cyber threats. The findings underscore the necessity of local intelligence in heterogeneous environments to maintain the resilience and reliability of essential services in an increasingly digitized global landscape.