Research On Wafer Surface Defect Detection System Based On Deep Learning

Wafer manufacturing is the upstream core of the semiconductor industry chain and plays an important role in supporting the development of the semiconductor industry.In order to reduce the cost burden caused by the continued processing of non-conforming wafers,it is necessary to perform fast and efficient defect inspection to improve the yield of wafers shipped.Vision-based wafer surface defect detection technology has attracted a lot of attention in recent years in semiconductor defect detection,and has achieved a lot of results in related fields.However,deep learning-based wafer surface defect detection methods still have a large research space in algorithm models and practical engineering applications.In this paper,we improve and study the wafer surface defect detection algorithm based on building a detection platform for wafer surface defect detection AOI system.Combined with the detection difficulties in actual engineering projects,the algorithm model improvement is focused on improving the speed and accuracy of the system defect detection.The main research of this paper is to first analyze the theoretical basis and detection mechanism of the detection algorithm,compare the structural differences of different models,and analyze the model performance with the detection effect of each model under the same data set to provide ideas for subsequent model improvement;then determine the causes and types of wafer defects by analyzing the defects from actual projects,and build the system’s machine vision system according to the defect imaging effect under different vision solutions.Then,we analyze the conventional inspection process of the wafer AOI system and build the mechanical structure model of the inspection system;then,we improve the algorithm to address the problems of the existing model,such as the large size of the existing model is not easy to deploy,and the overall detection accuracy of wafer surface defects is not high.The algorithm is improved by lightening the backbone network,adopting an efficient feature fusion mechanism,adding attention to the features,etc.Finally,a sample dataset is created to verify the effectiveness of the algorithm,and the overfitting phenomenon caused by a small dataset is avoided by sample expansion,and a Py Qt5-based human-computer interface is designed to facilitate the operation.The work in this paper enriches the detection methods in visual inspection of wafer surface defects,and the main results and innovations are as follows:(1)By comparing the performance of existing target detection algorithms,we analyze the possible reasons for the poor detection performance of the model and provide ideas for model improvement.The two-stage target detection algorithm based on region selection and the single-stage target detection algorithm based on regression are analyzed in detail,and the model is trained and tested on the same dataset VOC-2007,and the detection accuracy and speed of the model are comprehensively compared,and the experiments show that the YOLOv5 network model has better performance,so the YOLOv5 network is selected as the basic structure for the improvement of the algorithm in this paper.(2)In order to obtain higher quality inspection samples,the wafer inspection platform is built.We analyze the wafer production process and classify the common wafer surface defects;consider the imaging of four types of wafer surface defects under different light source conditions,wafer size and cost,etc.,and complete the vision solution design and selection of light source,camera and lens,etc.;according to the process requirements of wafer surface defect visual inspection process,complete the mechanical design including Z-axis module,loading three-axis and U-bearing sheet table,etc.According to the process requirements of the wafer surface defect visual inspection process,we will complete the mechanical design including Z-axis module,loading three axes and U-bearing chip table to build a complete inspection platform model.(3)In order to improve the detection accuracy and speed of the model for wafer surface defects,the algorithm improvement for wafer surface defect features is proposed based on a series of achievements of current deep learning in the field of target detection.Firstly,we propose the lightweighting of the model backbone network to address the problems of large model size,which leads to difficulties in field deployment and low detection speed,and the number of parameters of the model is reduced by 30% after improvement;secondly,we propose an efficient feature fusion mechanism to address the problem of poor detection accuracy of the model for small targets,which achieves the fusion of more features without increasing the parameters and improves the expression of small targets;finally,we add to the model Finally,an attention mechanism is added to the model to give different weights to the features from the spatial dimension and the channel dimension respectively to improve the attention of the model to the detected region.The improvement of the feature fusion mechanism results in the reduction of the number of model parameters to 90% of the original,a 9.4% reduction in computation,about 23% improvement in inference speed,92.3% detection accuracy,and better performance for all four defects,with an overall 9.8% improvement.(4)Data set creation and GUI interface design.The existing wafer data samples are expanded by data enhancement and data balancing,and finally a data set with a total data volume of 7510 is obtained,which effectively solves the overfitting problem caused by the small data volume and reduces the impact of sample imbalance on the model accuracy.Based on Py Qt5,the human-computer interface design including login registration,model training and result query is completed.In summary,this paper builds a vision inspection system oriented to the actual demand of visual inspection of defects on the surface of wafers in the manufacturing process,proposes a method for wafer surface defect detection and verifies its effectiveness,which has certain theoretical and engineering significance for improving the automation level of wafer production in China.