An improved YOLOX autonomous UAV-based system for monitoring photovoltaic panel failures

Addressing the prevalent difficulty in detecting anomalies in photovoltaic panels, such as hot spots or incidental surface contamination, this research propounds an autonomous UAV monitoring system. This system synergizes cutting-edge UAV inspection methodologies with an augmented YOLOX model. The foremost stage of the research accentuates the evolution of a sophisticated detection model, innovating beyond the conventional YOLOX framework. This progression incorporates the PVT-v2 network, lauded for its superior feature extraction. Further precision in the model’s diagnostic ability is achieved by embedding the CBAM, refining the network’s analytical focus on pertinent data. The revision in the label allocation strategy fortifies the network training, making it more precise and streamlined. The incorporation of the SIoU loss function accelerates the network’s convergence rate, effectively counteracting complications of misdiagnoses and recognition inaccuracies. The subsequent segment of the research pivots towards the enhancement of a segmentation algorithm via an advanced iteration of DeeplabV3+. This adaptation leverages the MobileNetV2 architecture, adjusting the downsampling multiplier to 8, thereby fine-tuning the algorithm’s performance. The strategy of employing the Dice loss function significantly reduces segmentation errors, especially those triggered by photovoltaic panel reflections. Experimental assessments reveal a substantial improvement in the mAP of the customized detection models, outstripping the benchmarks of not just the original YOLOX models but also other established frameworks like YOLOv5 series, Deformable DETR-R50, DAB-DETR-R50, and EfficientDet series. For segmentation, the enhanced “E” models surpass traditional accuracy thresholds of models like DeeplabV3+ and PSPnet. These empirical results confirm the robustness, reliability, and practicality of the proposed UAV system fortified with the improved YOLOX model for meticulous photovoltaic panel fault monitoring.