Dynamic Multi-Modal Fusion in Photovoltaic Defect Detection: A Lightweight Edge-Cloud Architecture with Three-Stage Collaboration

This paper proposes a three-stage collaborative defect detection framework for PV power plant operation and maintenance, effectively addressing the contradiction between detection accuracy and computational resources in dynamic scenarios. By integrating template matching, color space analysis, and visual language models (VLM), an edge-cloud primary-backup collaborative detection architecture is constructed, enabling real-time detection of module missing, stains, and cracks. Experiments demonstrate that the system achieves millisecond-level processing speed on embedded platforms, with a comprehensive miss rate of 2.68%, representing a 3.2-fold improvement in energy efficiency over traditional methods. The dynamic template update mechanism reduces the false positive rate to 1.43%, and the fusion detection strategy of VLM and HSV maintains high accuracy even under poor network conditions.