太阳能电池片隐裂检测的深度目标网络算法研究

Research on Deep Target Network Algorithm for Hidden Crack Detection of Solar Cells

太阳能电池片作为光电转换的重要部件,生产过程中容易产生表面缺陷和内部隐性缺陷(隐性裂纹,简称隐裂),严重影响了太阳能电池片的发电效率和成品合格率。目前基于图像处理的视觉检测方法针对隐裂的检测效果较差,无法满足工业现场需求。拟利用深度学习的目标检测网络对电池片进行隐裂检测研究。分析深度学习的目标检测网络模型算法,对选取的网络模型进行优化,解决了隐裂检测精度低的问题。利用采集的隐裂样本数据集,对比分析YOLOv5s、SSD、Faster-RCNN三种目标检测算法的隐裂检测效果,发现YOLOv5s模型综合性能较优。同时对YOLOv5s网络结构及功能模块进行优化,提高了隐裂检测速度和检测精度。结果表明,生成对抗样本构成的数据集中隐裂数据样本采用优化后的YOLOv5s进行隐裂检测,准确率可达到96%以上,单张图像检测时间大约为0.06 s。因此,结合生成对抗网络和目标检测网络可以实现隐裂的快速、高精度检测。

As an important part of a photoelectric conversion, the production process is easy to produce surface defects and internal hidden defects(hidden crack), serious impact on the solar cell power generation efficiency and product quality. At present, the visual detection method based on image processing has poor detection effect on hidden cracks and can not meet the needs of industrial field. The target detection network of deep learning was used to study the hidden crack detection of battery cells. The deep learning target detection network model algorithm was analyzed, and the selected network model was optimized to solve the problem of low precision of hidden crack detection. Using the collected crack sample data set, the crack detection effect of the three target detection algorithms of YOLOv5s, SSD and Faster-RCNN were compared and analyzed, and it was found that the YOLOv5s model had better comprehensive performance. The results show that the YOLOv5s model has better comprehensive performance. At the same time, the structure and function module of YOLOv5s network were optimized to improve the detection speed and accuracy of hidden cracks. The results show that the accuracy of P-YOLOv5s can reach 96% and the detection time of single image is about 0.06 s. Therefore, the combination of generative adversarial network and target detection network can realize the fast and high precision detection of hidden cracks.