Dual-Gate Electro-Visual Gating for Cleaning Decisions in Edge-Based Photovoltaic Maintenance

Photovoltaic (PV) systems in arid regions experience substantial energy losses from dust accumulation, but automated cleaning can introduce risk when damaged modules are visually misclassified as dirty. Manual inspection is also impractical at utility scale. This study presents an edge-based cleaning-decision architecture in which vision proposes a cleaning request and independent electrical checks determine whether actuation is permitted. The prototype combines a YOLOv8n detector, INA219 voltage/current sensing, photodiode-based shading context, a Raspberry Pi 4, and an Arduino co-processor. Cleaning is allowed only when three conditions are satisfied: persistent dust detection, measured panel power below a predefined threshold, and the absence of an electrical structural-fault signature based on rolling-window voltage depression and instability. The detector was trained using 950 field-recorded frames containing 2850 annotated panel instances and achieved 77.3% mAP@0.5 on the validation set; across five retraining seeds, mAP@0.5 was 77.1±1.1%. In a retrospective ablation of the recorded test campaign, the two verification gates reduced false or unsafe cleaning activations from five to one, while both unsafe activations involving the tested cracked module were vetoed. A 13.5 h three-day campaign recorded seven persistent visual dust requests; the power gate rejected six requests that did not justify cleaning, and the remaining request triggered a successful cleaning cycle. In that field event, panel power increased from 0.33 W to 7.48 W, corresponding to a 22.9-fold recovery. An assumption-based sensitivity analysis indicates potential water-cost savings from condition-based cleaning. An offline, template-constrained language model is used only for report generation and has no connection to actuation.