Design of photovoltaic panel intelligent cleaning system based on deep learning and SLAM

Aiming at the problems of traditional photovoltaic panel cleaning systems lacking environmental perception ability, relying on preset paths leading to low cleaning efficiency, multiple blind spots, and insufficient obstacle avoidance ability, a research proposes an intelligent cleaning system that integrates improved YOLOv8 object detection model with semantic synchronization positioning and map construction. The system constructs a photovoltaic panel detection model by introducing a lightweight MobileViT backbone network and MPDIoU loss function, achieving perception of the position and orientation of the photovoltaic panel. Embedding visual recognition results into the SLAM navigation loop, constructing a semantic map, and integrating improved A* global planning and DWA local obstacle avoidance algorithm to form an integrated intelligent closed-loop of "perception decision". The experimental findings denote that the improved YOLOv8 model is suitable for photovoltaic panel detection tasks mAP@0.5 98.2%, with a parameter size of 6.4M and a computational complexity of 18.5 GFLOPs. Under the semantic SLAM navigation strategy, the system's task completion rate reached 100%, the average cleaning time was reduced to 109 seconds, the path length was shortened to 10.6m, the path smoothness was significantly improved, and the success rate of dynamic obstacle avoidance reached 95%. Overall, the system designed by the research institute has significant advantages in perception accuracy, navigation efficiency, and dynamic environmental adaptability, and can provide reliable technical solutions for photovoltaic panel cleaning tasks in complex outdoor environments.