Introduction: The accuracy and real-time performance of drone-based defect detection on solar panels are critical to the operation and maintenance (O&M) efficiency, fault early-warning capability, power generation reliability, and emergency decision-making responsiveness of photovoltaic (PV) power plants. As solar farms expand in scale and complexity, automated inspection solutions have surged. Traditional manual or semi-automated methods are increasingly inadequate due to high labor costs, inconsistent diagnostics, and limited coverage. Efficient and robust defect identification thus becomes a cornerstone of intelligent PV system management, enabling proactive maintenance and maximizing energy yield over the plant’s lifetime.
Methods: This study adopts YOLOX—an advanced deep learning-based object detection algorithm—to automatically identify defects in PV modules from large-scale UAV aerial imagery. The model is specifically trained to detect five core defect types: Broken Glass, Diode Failure, Hot Spots, Obscured panels, and Potential-Induced Degradation (PID) Effect. By integrating baseline PV module databases, complex scene adaptation models, and morphology-differentiated recognition algorithms, YOLOX effectively captures both intrinsic fault mechanisms and context-specific distribution patterns. The model leverages its powerful feature extraction and high-precision localization capabilities to suppress common disturbances such as image noise, inter-module shadows, uneven illumination, and strong glare, ensuring reliable detection across diverse environments.
Results: Experimental evaluations confirm that YOLOX maintains high detection accuracy under challenging real-world conditions—including overcast skies, rainy weather, low-light nighttime scenarios, and large-scale PV arrays. The algorithm provides precise spatial coordinates of defects. Comparative tests demonstrate that YOLOX outperforms models like YOLOv5 and Faster R-CNN in both detection precision and inference speed, particularly in scenes with partial occlusion or reflective glare. Its ability to adapt to varying installation types further enhances its practical applicability.
Conclusions: The desgined approach significantly enhances the automation, precision, and intelligence of PV plant O&M by enabling rapid fault localization, dynamic repair prioritization, and accurate energy loss quantification. By providing structured outputs that link defect location and severity, YOLOX supports data-driven decision-making in routine operations and emergency inspections. This study confirms the feasibility, robustness, and practical applicability of YOLOX for real-world solar farm inspection, offering a scalable and efficient solution for next-generation intelligent photovoltaic management systems.
