An effective approach to improving photovoltaic defect detection using the new DCD-YOLOv8s model

Area of Science:

• Renewable Energy Systems
• Computer Vision
• Artificial Intelligence

Background:

• Photovoltaic (PV) systems are crucial for renewable energy, but surface defects reduce efficiency.
• Current inspection methods are inefficient, and deep learning models lack adaptability and performance.

Purpose of the Study:

• To develop an enhanced YOLOv8 algorithm for accurate and efficient PV surface defect detection.
• To improve feature extraction, localization, and detection accuracy while minimizing computational load.

Main Methods:

• Proposed DCD-YOLOv8s algorithm integrating deformable convolutional networks (DCNv3), coordinate attention (CA), and dynamic head (DyHead) modules.
• Created a custom dataset combining public and field-collected PV panel defect images.
• Utilized data augmentation and a self-training strategy to expand the dataset.

Main Results:

• DCD-YOLOv8s achieved an F1-score of 92.8%, mAP@50 of 95.0%, and mAP@50-95 of 82.3%.
• Demonstrated superior performance compared to multiple existing YOLO and RT-DETR models.
• Maintained a high inference speed of 45.9 FPS.

Conclusions:

• DCD-YOLOv8s shows significant potential for real-time UAV-based inspection systems.
• The algorithm contributes to cost-effective and reliable PV system maintenance.
• Further development could enable direct UAV deployment for enhanced operational efficiency.