GridScope: A Cloud-Edge Collaborative Framework for Robust and Efficient Power-Grid Inspection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Power-grid inspection requires both high detection reliability and low-latency deployment, yet existing solutions often face a fundamental trade-off between lightweight edge models with limited semantic capability and large multimodal models with high inference cost. To address this issue, we propose GridScope, a cloud-edge collaborative object detection framework for power-grid inspection. GridScope integrates an SLA-aware confidence routing to adaptively coordinate local inference, cloud verification, and dual-channel response according to task priority and detection confidence. To improve communication efficiency, we further design a split-VLM collaborative inference pipeline, in which compact visual features are extracted at the edge and only feature-level representations are transmitted to the cloud for semantic verification. In addition, cloud-side refined predictions are used to periodically optimize the edge detector, forming a continual feedback loop that improves local detection quality over time. Experiments on wildfire and foreign-object intrusion benchmarks demonstrate that GridScope consistently achieves a more favorable accuracy-efficiency trade-off than both lightweight detectors and full VLM baselines.