| Detection of defects is the key process of solar cell production. Every subtle sectors will likely lead to defects in production process. If diagnose categories of solar cell defects quickly and accurately, and analyze generation mechanism, it can adjust production process promptly and improve passing rate. Traditional methods of detect detection contain artificial visual detection, I-V feature detection and visible image detection. While these methods are vulnerable to environmental interference, so that result is inaccuracy. Because diagnosis technology of infrared image has advantages of non-contact. Crack that the naked eye is difficult to see can be detected. So this paper adopts detection technology of electroluminescent infrared image for diagnosis of solar cell defects.Firstly, this paper tests six kinds of defects that is appears easily in production process of Yingli Energy(China) Co. are summarized. They are debris, crack, off-grid, open-weld,black chip and shading chip. Meanwhile the article gives generation reasons of each defect.Then device of infrared image are described. Thus a variety of infrared images of defects are obtained. Aiming at the characteristics of infrared images of defects, the paper carried out remove main line, median filtering, hat transform, edge detection, threshold segmentation.According to divided parts of defect image, HU moment invariants are applied for feature extraction. This method ensure extracted features have properties of translation, rotation and scaling simultaneously. This paper uses BP neural network as classifier for recognition. 4800 electroluminescence infrared images are used to train the network. The network of different training sample proportion, verify sample proportion, identified sample proportion and hidden layer neurons are trained repeatedly. The network of 97.5% of correct rate is saved. 480 images of six kinds of solar cell defects are recognized. Finally the recognition rate of each kind defect reached above 95%. |
PDF Full Text Request