Research On Intelligent Detection Algorithms For Internal Defects Of Solar Cells

To deal with the energy crisis of the 21 st century,many countries are vigorously developing photovoltaic power generation.The solar cell is the smallest unit of the photovoltaic system and its internal defects reduce the photoelectric conversion efficiency of photovoltaic system.Although there are some researches on the internal defect detection of solar cells,they suffer from some problems such as poor correlation of detection steps,single detection items and low degree of intelligence.To solve the aforementioned problems,this thesis proposes an image processing technology-based internal defect automatic detection algorithm of monocrystalline silicon solar cell,and improves the defect segmentation algorithm to enhance the versatility of the algorithm for the internal defect detection of solar cells in different production processes.In order to further strengthen the universality and intelligence of the internal defect detection algorithm of solar cells,an intelligent labeling algorithm based on a few-shot learning model is designed to realize the intelligent labeling of large-scale defect samples and reduce the labor cost of manual labeling.Furthermore,a data-driven and end-to-end learning intelligent quantification algorithm for internal defects of solar cells is designed to improve defect segmentation and classification capabilities.The main research contents of this thesis are as follows:(1)Aiming at the poor correlation of the detection steps of the existing internal defect detection algorithms of monocrystalline silicon solar cells,based on the object detection framework of machine vision,the image features are analyzed,and taking into account the correlation between steps,a defect detection algorithm is designed for the internal defect detection of monocrystalline silicon solar cells.For cell positioning,defect segmentation and defect classification,a Sigmoid Hough transform-based geometric segmentation algorithm,a self-comparison algorithm and a defect diagnosis rule based on bagging algorithm are proposed,respectively.The experimental results show that the intersection over union ratio of the proposed cell positioning algorithm is 0.998,the F measure of the defect segmentation algorithm is 0.164 higher than other comparison algorithms,and the precision of the defect classification algorithm is greater than 0.971 and the recall is greater than 0.980.(2)Aiming at the problem that the defect detection algorithm of solar cell is specific to the type of cells or defects,a component-of-interest superposition graph algorithm is designed.Based on the aforementioned original algorithm framework,this algorithm replaces the original defect segmentation algorithm,so that it can be applied to internal defect detection of solar cells in different production processes.This defect detection process is regarded as an anomaly extraction task in complex background.Singular value decomposition is introduced into Fourier transform to construct component graph,so that images with different gray dynamic ranges are mapped to a unified benchmark.The normalized energy margin gain is used to weighted average the different component graphs to obtain the saliency map,so that a global threshold can be used to complete the defect segmentation.The experimental results show that the proposed detection algorithm achieves good results in five solar cell datasets of different production processes,and the F measure values of defect level in PLMP,PLMC,ELMC and ELPC datasets are greater than 0.702,which is universal.(3)Aiming at the problem of high labor intensity of data labeling,an intelligent labeling algorithm based on a few-shot learning model is proposed.The algorithm mines common features among data based upon public large-scale natural image datasets.Then these common features are used to assist a very small number of manually annotated support samples in the task to guide the query samples for defect labeling.Finally,large-scale internal defect labeled samples of solar cells are obtained for the data-driven intelligent algorithm learning.Experimental results show that the recall rate of defect labeling on ELPC dataset is above 0.898,which improves the performance of intelligent labeling.(4)Aiming at the cumbersome design process of image processing-based algorithms and high requirements for domain knowledge of algorithm designers,a data-driven and end-to-end learning defect intelligent quantization algorithm is proposed.By automatically learning the features of large-scale samples,the algorithm realizes the online segmentation and classification for internal defects of solar cells.Considering the loss of the feature information and location information of recessive cracks and small-size defects as the network deepens,a multi-scale connection and coder-decoder network structure is designed to make full use of the shallow and deep feature information.In addition,convolution block and improved Swim Transformer block are also adopted to speed up the network’s understanding of local information and global information.Finally,the Focal and Dice losses of the binary classification are extended to the multi-classification level,which solves the problems of complex sample learning difficulty and unbalanced sample number.The experimental results show that the accuracy of the proposed algorithm is 0.105 higher than that of the generic semantic segmentation network used in computer fields in the task of binary classification.In the defect segmentation task,the F measure of the proposed algorithm is 0.108 higher than that of other networks.Finally,concentration and academic future of intelligent defect detection algorithm is speculated based on the summarization of the whole thesis.