GridAI: Edge AI Smart Grid Fault Detection with Goal-Oriented Communication

Reliable fault detection in electric distribution grids depends on local analysis at the sensing point and the quality of communication between sensing devices and monitoring servers. Existing systems treat these two aspects independently: threshold-based detectors transmit data continuously regardless of its content, while communication-efficient protocols rarely account for the meaning of the data they carry. In this work, we present GridAI, an edge-AI monitoring system that addresses fault detection and communication together. GridAI combines an XGBoost-based two-task fault detection method with semantic communication, running on a Raspberry Pi-class device installed at a distribution transformer terminal. The fault detection method performs fault-normal classification, per-phase fault type identification, and fault-area localization; the communication layer then uses this information to decide what to transmit and when. Training data is generated from a MATLAB/Simulink model of a distribution grid covering 7 fault types across 42 locations. Laboratory experiments on a physical transformer testbed show that GridAI achieves approximately 99% fault detection accuracy with sub-second inference latency on the edge device. By replacing continuous 105-byte measurements (151 KB/day) with 15-byte heartbeats and on-demand fault reports (40 bytes/day), the goal-oriented policy reduces communication overhead by more than three orders of magnitude without sacrificing detection performance on resource-constrained hardware.