Research On Wafer Map Recognition And Classification Based On Feature Pyramid Fusion And Attention Mechanism

Wafer manufacturing is an important part of the semiconductor manufacturing process.Process problems in the wafer manufacturing process will cause wafer defects,thereby affecting the production yield of the wafer.The distribution of wafer defects can be described by the wafer map,and the analysis of the defect pattern of the wafer map can locate the root cause of the defects,thereby providing useful information for engineers and helping to improve the production yield of wafers.However,due to the angle diversity of wafer defects and the uneven distribution of defect types,the accuracy of existing wafer defect pattern recognition and classification algorithms needs to be further improved.Therefore,this paper proposes a convolutional neural network based on feature pyramid fusion and attention mechanism to identify and classify wafer map defect patterns to improve the accuracy of wafer defect pattern recognition and classification.The main research work is as follows:(1)Aiming at the problem that the deep convolutional neural network only predicts the features of the last layer and ignores some important shallow feature information,resulting in low recognition accuracy of some smaller target features,a convolutional neural network based on feature pyramid fusion is proposed.The 11-layer network DCNN11,the 27-layer residual network DCNN-Res27 and the deeper residual network Resnet50 are used as the benchmark network.Then,the feature pyramid fusion method was incorporated into these three network models,resulting in FPN-DCNN11,FPN-DCNN-Res27,and FPN-Resnet50,respectively.The experimental results show that the method of feature pyramid fusion can effectively improve the classification performance of the model,especially for the recognition and classification of small target features such as scratch.The maximum improvement of the three classification indicators,Precision,Recall,and F1-Score,are12.36%,7.0%,and 9.74%,respectively.After incorporating the feature pyramid fusion method into the three models,their classification accuracy reached 95.80%,95.02%,and95.76%,respectively,which are 1.8%,0.56%,and 0.31% higher than the original models.(2)Aiming at the problems of unbalanced data distribution of wafer map defect patterns,diversity of feature angles,easy confusion of similar types of features,and low accuracy of convolutional neural network for certain types of detection and recognition,a convolutional neural network based on an improved multi-branch attention mechanism is proposed.First,the Convolutional Block Attention Module(CBAM)is improved,and a multi-branch channel attention module is designed.Secondly,Resne Xt50 with a grouped convolution structure is used as the backbone feature extraction network,and the improved attention mechanism(I-CBAM)is integrated into Resne Xt50 to increase the attention to important feature information.Finally,after the features are extracted by the fully connected layer of the feature extraction network I-CBAM-Resne Xt50,the support vector machine(SVM)and Error Correcting Output Codes(ECOC)are combined as the classifier in the final stage.The experimental results show that the method proposed in this paper can assign different attention weights to different types of defect features,and can effectively identify salient and important feature information.Comparing the model proposed in this paper with the classic models VGG16,Resne Xt50 and Resnet50,the method proposed in this paper has achieved better classification results,and the accuracy rates have increased by 0.85%,1.56%,and2.40%,respectively.Compared with other literatures,the method proposed in this paper has a better classification effect,achieving a classification accuracy of 96.96%.(3)In order to further utilize the classification advantages of the four models of FPNDCNN11,FPN-DCNN-Res27,FPN-Resnet50 and I-CBAM-Resne Xt50 for different types of defect wafer maps.The method of convolutional neural network adaptive decision is adopted to improve the recognition and classification accuracy of the wafer map defect patterns.Firstly,the entropy of the output results of each classifier model is calculated,which can be used to determine the uncertainty of the classification results.Then,the four information entropies are weighted to adaptively generate the fusion weight of each classifier,and the classifier with more reliable classification gets a larger weight.Finally,decision fusion is performed on the weighted output results,and ECOC-SVM is used as the final classifier to obtain the final prediction output results.The experimental results show that the method of convolutional neural network adaptive decision fusion can effectively improve the classification performance compared to a single classification model.The classification accuracy of the proposed method can reach 97.10%,which is 0.54% higher than the best model,I-CBAM-Resne Xt50,among the four classification models,and 0.14% higher than I-CBAM-Resne Xt50-ECOC-SVM.