Classification And Detection Of Surface Defects In Solar Cells Based On Convolutional Neural Network

In the production process of solar cells,the detection of surface defects of solar cells under visible light has become a key process in the manufacturing process.The surface defect detection of polycrystalline silicon solar cells has the characteristics of complex background unevenness,uncontrollable lattice features,and many types of features.The types of defects include paste spot,broken gate,scratches,Color difference,thick lines,dirty cells,etc.Surface defect detection of polycrystalline silicon solar cells is a multi-feature detection in a non-uniform texture background.Most of the current surface inspections of solar cells rely on artificial naked eye detection.Some traditional artificial feature extraction and machine learning algorithms are highly targeted and cannot be adapted to all types of solar cell surface defect detection.To this end,this paper focuses on the visual intelligence detection of surface defects of polycrystalline silicon cells based on deep learning,and gradually realizes the classification and segmentation of various types of defects on the surface of solar cells.In this paper,the optimal design of the Convolutional Neural Network model for solar cells is realized.The Multi-spectral Convolutional Neural Network model is proposed,which significantly improves the accuracy of detection classification and feature extraction.Then an improved multi-scale class activation mapping model is proposed.The detection of the defect position of the surface defect of the solar cells are improved.The specific research contents and contributions of this paper are as follows:(1).In order to solve the similar and indeterminate defect detection problem of solar surface with heterogeneous texture and defects under complex background,the Convolutional Neural Network(CNN)is used for the surface defect detection of solar cells,and the specification is reliable.By adjusting the depth and width of the model,the influence of model depth and kernel size on the recognition result is evaluated,and the best-performing structure was selected.The parameters establish a solar cell Solar-CNN model.(2).In order to improve the classification effect of CNN model,this paper proposes a Multi-spectral Deep Convolutional Neural Network(MS-CNN),which greatly improves the accuracy of classification effect.By analyzing the performance characteristics of defects in multiple spectra,the method of image multi-spectral information feature separation and extraction is used to enhance the model's ability to extract multi-spectral image information features,and a multi-spectral convolutional neural network model is constructed to enhance the model.The ability to distinguish between texture background features and defect features.Then,all the training data are traversed by K-fold cross-validation.The experimental results show that the multi-spectral CNN model can effectively solve the problem of complex background texture,various defect features and random shape of the cell surface,and strengthen the model's characteristics for multiple spectral information.The extraction ability has higher accuracy and stronger stability,and the defect recognition accuracy rate is over 95%.(3).In order to realize the position detection and defect segmentation of surface defects of solar cells,this paper proposes Multiscale Class Activation Mapping(MCAM),which improved the accuracy and adaptability of the Class Activation Map(CAM)for surface defect detection and segmentation in solar cells.In this paper,multi-scale hybridization of low-level features and advanced features of the model is realized.Combined with multispectral CNN,a multi-scale class-activated mapping model is implemented.The improved multi-scale CAM model has more accurate background distinguishing ability and feature recognition ability.