Аңдатпа
Dual-branch RGB-thermal infrared (TIR) deep networks are increasingly being utilized for Unmanned Aerial Vehicle (UAV)-based photovoltaic (PV) module fault detection, but most published evidence remains single-site and does not rigorously demonstrate cross-installation robustness under varying panel types, layouts, and environmental conditions. To address this, this paper presents a two-phase research pathway toward a robust multi-site PV fault detection using a dual-branch RGB-TIR framework analyzed across heterogeneous datasets. The phase 1 of this study comprises of assessments conducted on RGB and TIR datasets from public repositories to characterize model behavior, training dynamics, and dataset-driven limitations before deployment on proprietary, higher quality self-acquired site imagery. This study identifies practical data-centric requirements for stable training which involves allocating approximately$\mathbf{4 0 - 6 0 \%}$background/null imagery to reduce false positives and mitigate overfitting towards defect-only images. This enforces distinct and sufficiently represented defect classes while excluding rare or visually confusable categories, and finally, adopting oriented bounding boxes (YOLOv8-OBB) for RGB imagery to accommodate the rotated geometry of PV modules and defect extents in Unmanned Aerial Vehicle (UAV) imagery. Phase 2 of this study will transfer these insights into the assessment of our self-acquired imagery, with the goal of improving cross-site generalization and operational readiness for scalable PV operations and maintenance.