Real-Time Object Detection of Photovoltaic Defects via Aerial Thermography: Evaluating Convolutional Neural Network and Transformer Architectures

الملخص

The rapid expansion of large-scale photovoltaic plants requires efficient predictive maintenance strategies to ensure optimal energy generation. Traditional manual thermal inspections are time-consuming and subjective, driving the need for automated solutions via Unmanned Aerial Vehicle thermography. This paper presents a comparative study of state-of-the-art Deep Learning architectures for the automated object detection of thermal anomalies. The study focuses on critical defects, e.g., hotspots, open-circuit, and string open-circuit utilizing a specialized, manually curated dataset of aerial thermal images from 50 MWp operational solar plants. Three distinct architectures were evaluated to determine the optimal trade-off between diagnostic precision and computational cost: a Convolutional Neural Network (YOLOv11), a vision-language hybrid (YOLO-World), and an end-to-end real-time Vision Transformer. Experimental results demonstrate that Transformer achieves superior boundary precision and recall (mAP50-95 of 0.80), effectively mitigating false negatives for diffuse micro-anomalies. Conversely, YOLOv11 delivers unrivaled inference agility (4.43 ms per image), making it the optimal choice for massive, high-speed inspections constrained by edge-computing hardware. YOLO-World serves as a highly capable intermediate alternative, offering strong small-object detection with lower computational overhead than pure Transformers. These findings validate the practical applicability of these models for industrial UAV inspections, enabling operators to deploy specific architectures tailored to their unique hardware constraints and operational requirements.