Research On Automatic Classification Method Of Solar Cell Defects Based On Model Fusion

With the increasingly acute problem of world energy scarcity,countries have turned their attention to new energy sources,among which solar energy as a renewable energy source has attracted the attention of all countries in the world.However,the solar cell as a power generation carrier often causes various defects on the surface of the cell due to the limitation of production technology.In the classification of the appearance defects of the cell,the traditional visual classification method needs to manually extract the sample characteristics,the classification speed is slow,the classification accuracy is not high,and it cannot meet the production identification needs of the enterprise.Therefore,it is necessary to develop a fast,efficient,and non-destructive method for classifying solar cell defects.This paper takes the electroluminescence(EL)image of the solar cell as the research object,and studies a high-accuracy classification method for solar cell defects based on convolutional neural network model fusion,and applies it to the research and development of an automatic solar cell defect classification system based on EL Provide theoretical and technical basis in the actual production inspection of solar cells.The research contents are as follows;(1)Research and analyze the composition structure and feature extraction principles of three classic convolutional neural networks,which are Le Net-5,Alex Net,and Goog Le Net models.Improve the model network structure,activation function and parameter update algorithm,and solve the exiting problems of the above three models.The solar cell defect data set has problems such as slow training speed,many weight parameters,and low classification accuracy.(2)In-depth study of three model fusion methods based on correlation entropy criterion,feature fusion and stacking,respectively,to achieve the improved Le Net-5,improved Alex Net,and improved Goog Le Net based on the three model fusion methods.The effective integration of the models has further completed the design and implementation of the algorithm flow.(3)Constructed the solar cell defect data set through the data enhancement method,and constructed the cell defect classification software system with the help of the Tensor Flow 2.0 framework,and carried out comparative experiments on the defect classification effects of the three model fusion methods proposed in this paper.The results show that the model fusion classification method based on feature fusion has the highest accuracy,which can reach 97.46%,which meets the accuracy requirements of enterprises for automatic cell classification.