Research on multi-defects classification detection method for solar cells based on deep learning.

Area of Science:

• Materials Science
• Electrical Engineering
• Computer Science

Background:

• Solar cells are critical components in aerospace technology.
• Surface defects in solar cell manufacturing can compromise performance and reliability.
• Accurate and efficient defect detection is essential for quality control.

Purpose of the Study:

• To develop and validate a deep learning-based system for detecting and classifying surface defects in solar cells.
• To address the challenges posed by different types of defects, including mismatch, bubbles, cracks, and upside-down glass.

Main Methods:

• Utilized YOLOv5s model with K-means clustering for re-clustering anchor boxes to address mismatch defects.
• Improved YOLOv5s for general defects through image preprocessing, anchor box optimization, and detection head replacement.
• Employed MobileNetV2, a lightweight classification network, for efficient detection of easy-to-detect defects like upside-down glass.

Main Results:

• Achieved high detection precision: 95.64% for mismatch, 91.8% for bubbles, 93.1% for glass cracks, and 98.0% for cell cracks.
• MobileNetV2 achieved 100% average classification accuracy for upside-down glass defects at 13.29 frames per second.
• Demonstrated significant improvements in detection accuracy and speed compared to the original models.

Conclusions:

• The proposed multi-model fusion and optimized deep learning approach effectively enhances solar cell defect detection.
• The system offers a robust solution for quality control in solar cell manufacturing, particularly for aerospace applications.
• The combination of optimized models and lightweight networks provides a balance between accuracy and speed.