Optimized YOLO based model for photovoltaic defect detection in electroluminescence images

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence in Engineering
• Materials Science

Background:

• Photovoltaic (PV) system reliability is crucial for sustained energy production.
• Automated defect detection in PV panels is essential for maintenance and performance.
• Deep learning object detection models offer promising solutions for PV defect identification.

Purpose of the Study:

• To introduce PV-YOLOv12n, an optimized YOLOv12n variant for detecting defects in PV panel electroluminescence (EL) images.
• To enhance feature extraction for improved detection of critical defects such as large cracks, dislocations, and material inconsistencies.
• To evaluate the performance and efficiency of PV-YOLOv12n against existing models on benchmark datasets.

Main Methods:

• Development of PV-YOLOv12n by integrating an A2C2f module at the P5 scale into the YOLOv12n architecture.
• Utilizing electroluminescence (EL) images for defect detection.
• Experimental validation using the PVEL-AD and Roboflow datasets.

Main Results:

• PV-YOLOv12n achieved a mean Average Precision (mAP@50) of 0.91 on both PVEL-AD and Roboflow datasets.
• Outperformed baseline YOLOv12n, demonstrating improved precision and recall for critical PV defects.
• Showcased enhanced generalization with mAP@50-95 scores of 0.58 (PVEL-AD) and 0.75 (Roboflow).
• Maintained efficient inference speeds (4.24 ms and 4.43 ms), suitable for real-time applications.

Conclusions:

• PV-YOLOv12n effectively detects critical defects in PV panels, enhancing reliability.
• The optimized model supports efficient, large-scale solar farm inspections.
• The integration of the A2C2f module significantly improves defect detection capabilities.