Anomaly Detection Methods for Power Equipment Based on Lightweight Deep Learning and Edge Computing

To address the conflict between high-precision deep learning demands and the resource-constrained nature of power grid edge nodes, this paper proposes FED-RepNet, an end-to-end lightweight anomaly detection framework specifically engineered for industrial edge intelligence. The architecture integrates a Time-Domain Morphological Stream and a Frequency-Domain Spectral Gating Stream to capture both transient spikes and subtle harmonic distortions. The architecture integrates a Time-Domain Morphological Stream and a Frequency-Domain Spectral Gating Stream to capture both transient spikes and subtle harmonic distortions. We develop a dual-stream coupling mechanism that synergizes time-domain morphology with frequency-domain spectral gating for high-sensitivity fault isolation. We also implement a structural re-parameterization strategy that collapses complex training topologies into a streamlined single-path inference structure. Experimental results across three heterogeneous datasets-vibration, visual, and multi-sensor-demonstrate that FED-RepNet achieves a superior balance between diagnostic accuracy (95.17% F1-score) and real-time efficiency (3.25 ms latency), establishing a new Pareto frontier for industrial edge intelligence and providing a robust, deployment-ready solution for real-time monitoring in modern smart grids.