Research On Defect Detection Technology Of Mono-like Silicon Wafer Based On Photoluminescence

With the rapid development of society and the increasing requirements for environmental protection,people’s demand for clean and green energy is gradually expanding.Relying on the advantages of sustainable development,solar energy is sought after by countries all over the world.Solar cell is the core component of photovoltaic power generation system,and its carrier is solar silicon wafer.However,in the production process,solar silicon wafers will inevitably produce many types of defects.Although the automation and intelligence of solar silicon wafer manufacturing are already very high,the detection and removal of silicon wafer defects are mostly done manually,and the detection effect is too dependent on the experience ability of the inspector,which leads to low detection efficiency,reliability and Robustness is poor.This paper takes Mono-like silicon(ML-Si)silicon wafers as the research object,uses Photoluminescence(PL)imaging technology and image processing methods to detect silicon wafer defects,and uses deep learning technology to complete the silicon wafer Quality prediction.The main work and innovations of this paper are as follows:(1)Aiming at the situation that commonly used denoising algorithms will cause the loss of target edge pixel information,this paper proposes an image denoising algorithm based on improved mean filtering.First,design the "工"-shaped template,which is rotated in a clockwise direction with the center of the template as the origin,and the angle of each rotation is 45°.Then,the gray value distribution of the pixels covered by the template is described by the gray variance.Finally,the template position with the smallest gray-scale variance is taken as the best position for mean filtering.Experiments show that the edge pixel information of the image processed by this algorithm is better preserved.(2)Regarding the uneven illumination in the collected images,the conventional algorithm takes a long time to process.This paper proposes an image enhancement algorithm based on partial overlap of regions.The brightness difference value is constructed based on the original image and the template.In order to make the adjacent brightness differences have cohesion,40% of the template length is taken as the template sliding step length.The obtained brightness difference points are expanded to the size of the original image to obtain the difference image.By subtracting the original image and the difference image,the purpose of eliminating uneven illumination and reducing the burden for the subsequent segmentation algorithm is achieved.(3)A method of dislocation extraction based on Hessian matrix is proposed.Due to the slip phenomenon that occurs during the formation of dislocations,the gray value of the dislocation branch pixels is small,and it is difficult to extract all of them with traditional segmentation methods.To solve this problem,first obtain the main direction of the dislocation through the Hessian matrix,and then design the "Tian" template to traverse the original image to ensure that the template is aligned with the main direction of the dislocation,and finally calculate the total cosine similarity of the two,and then through the linear mapping.The method obtains the dislocation extraction image.(4)Propose a silicon wafer quality prediction algorithm based on deep learning to further quantify the relationship between dislocations and silicon wafer quality.Based on the VGG and Resnet network structure,the ADAM and SGDM optimization algorithms are used to design the silicon wafer quality prediction model,and the prediction results and reasoning performance are analyzed.The experiment proves that the VGG-19 network model using the ADAM optimization algorithm has the best performance.Subsequently,a silicon wafer quality prediction algorithm based on the improved VGG-19 was proposed to adjust the network structure of the VGG-19 model.After experimental comparison,the prediction error of the improved network model is reduced by 0.1.The inference speed reaches 4FPS per second.