Abstract: To address the problems of dispersed multi-source data, high cloud processing latency, and insufficient on-site intelligent analysis in power inspection, this paper designs an edge-intelligent system for power operation and maintenance based on smart helmets. It constructs an integrated multi-source data collection and three-layer fusion architecture, and completes the optimization of lightweight AI algorithms for four core scenarios. Tests show that the YOLOv8n model achieves a detection accuracy of 98.6% with a single-frame inference time of 90 ms, and the infrared temperature inversion error is controlled within ±0.5 ℃. Field applications indicate that the system improves inspection efficiency by 42%, reduces data upload volume by 90%, and cuts bandwidth consumption by 65%, providing a technical reference for intelligent operation and maintenance transformation in similar industrial scenarios.