使用深度学习分类缺陷晶片地图上

dataDir = fullfile (tempdir,“WaferDefects”);downloadWaferMapData (dataDir)

dataMatFile = fullfile (dataDir,“MIR-WM811K”,“MATLAB”,“WM811K.mat”);waferData =负载(dataMatFile);waferData = waferData.data;

disp (waferData (1))

unlabeledImages = 0(大小(waferData),“逻辑”);为idx = 1:尺寸(unlabeledImages, 1) unlabeledImages (idx) = isempty (waferData (idx) .trainTestLabel);结束waferData (unlabeledImages) = [];

fieldsToRemove = [“dieSize”,“lotName”,“waferIndex”,“trainTestLabel”];waferData = rmfield (waferData fieldsToRemove);

defectClasses = [“中心”,“甜甜圈”,“Edge-Loc”,“环”,“疯狂的”,”“准””,“随机”,“抓”,“没有”];numClasses =元素个数(defectClasses);

waferData = struct2table (waferData);waferData.Properties。VariableNames = [“WaferImage”,“FailureType”];waferData。FailureType =分类(waferData.FailureType defectClasses);

displaySampleWaferMaps (waferData)

总结(waferData.FailureType)

waferData = oversampleWaferDefectClasses (waferData);

总结(waferData.FailureType)

labelIdx = splitlabels (waferData (0.9 0.05 0.05),“随机”TableVariable =“FailureType”);trainingData = waferData (labelIdx {1}:);validationData = waferData (labelIdx {2}:);testingData = waferData (labelIdx {3}:);

8月= imageDataAugmenter (FillValue = 0, RandXReflection = true, RandYReflection = true, RandRotation = 360年[0]);

inputSize = [48 48];dsTrain = augmentedImageDatastore (inputSize trainingData,“FailureType”DataAugmentation = 8月);

dsVal = augmentedImageDatastore (inputSize validationData,“FailureType”);dsVal。MiniBatchSize = 64;dst = augmentedImageDatastore (inputSize testingData,“FailureType”);

层= [imageInputLayer inputSize [1],…归一化=“rescale-zero-one”、最小值= 0,Max = 2);convolution2dLayer(3 8填充=“相同”)batchNormalizationLayer reluLayer maxPooling2dLayer(2步= 2)convolution2dLayer(填充= 3,16日“相同”)batchNormalizationLayer reluLayer maxPooling2dLayer(2步= 2)convolution2dLayer(3、32、填充=“相同”)batchNormalizationLayer reluLayer maxPooling2dLayer(2步= 2)convolution2dLayer(64,填充=“相同”)batchNormalizationLayer reluLayer dropoutLayer fullyConnectedLayer (numClasses) softmaxLayer classificationLayer];

选择= trainingOptions (“亚当”,…ResetInputNormalization = true,…MaxEpochs = 30,…InitialLearnRate = 0.001,…L2Regularization = 0.001,…MiniBatchSize = 64,…洗牌=“every-epoch”,…Verbose = false,…情节=“训练进步”,…ValidationData = dsVal,…ValidationFrequency = 20);

doTraining =假;如果doTraining trainedNet = trainNetwork (dsTrain层,选项);modelDateTime =字符串(datetime (“现在”格式=“yyyy-MM-dd-HH-mm-ss”));保存(fullfile (dataDir“trained-WM811K——”+ modelDateTime +“.mat”),“trainedNet”);其他的downloadTrainedWaferNet (dataDir);trainedNet =负载(fullfile (dataDir“CNN-WM811K.mat”));trainedNet = trainedNet.preTrainedNetwork;结束

defectPredicted =分类(trainedNet, dst);

defectTruth = testingData.FailureType;cmt = confusionmat (defectTruth defectPredicted);图confusionchart (cmt,类别(defectTruth)标准化=“row-normalized”,…Title =“测试数据混淆矩阵”);

prTable =表(大小= [numClasses 3], VariableTypes = [“细胞”,“细胞”,“替身”),…VariableNames = [“回忆”起,“精度”,“F1”),RowNames = defectClasses);为idx = 1: numClasses numTP = cmt (idx idx);numFP =总和(cmt (:, idx)) - numTP;numFN =总和(cmt (idx:), 2)——numTP;精度= numTP / (numTP + numFP);回忆= numTP / (numTP + numFN);defectClass = defectClasses (idx);prTable。回忆{defectClass} =回忆;prTable。精密{defectClass} =精度; prTable.F1(defectClass) = 2*precision*recall/(precision + recall);结束

prTable

defectProbabilities =预测(trainedNet, dst);

中华民国= rocmetrics (defectTruth defectProbabilities、defectClasses AdditionalMetrics =”前的“);图绘制(中华民国,XAxisMetric =“reca”YAxisMetric =”前的“);包含(“回忆”起)ylabel (“精度”网格)在标题(“所有类Precision-Recall曲线”)

prAUC = 0 (numClasses, 1);为idx = 1: numClasses defectClass = defectClasses (idx);currClassIdx = strcmpi (roc.Metrics。ClassName, defectClass);reca = roc.Metrics.TruePositiveRate (currClassIdx);prec = roc.Metrics.PositivePredictiveValue (currClassIdx);prAUC (idx) = trapz (reca(2:结束),prec(2:结束);% prec(1)总是NaN结束prTable。AUC = prAUC;prTable

defectClass =defectClasses (2);

idxTrue =找到(testingData。FailureType==defectClass); idxPred = find(defectPredicted == defectClass);

idxCorrect =相交(idxTrue idxPred);idxToEvaluateCorrect =1;imCorrect = testingData.WaferImage {idxCorrect (idxToEvaluateCorrect)};

idxIncorrect = setdiff (idxTrue idxPred);idxToEvaluateIncorrect =1;imIncorrect = testingData.WaferImage {idxIncorrect (idxToEvaluateIncorrect)};labelIncorrect = defectPredicted (idxIncorrect (idxToEvaluateIncorrect));

imCorrect = imresize (imCorrect inputSize);imIncorrect = imresize (imIncorrect inputSize);

scoreCorrect = gradCAM (trainedNet imCorrect defectClass);scoreIncorrect = gradCAM (trainedNet imIncorrect labelIncorrect);

t = nexttile图tiledlayout (1、2);displayWaferScoreMap (imCorrect scoreCorrect t)标题(“正确的分类(“+ defectClass +“)”)t = nexttile;displayWaferScoreMap (imIncorrect scoreIncorrect t)标题(“误分类(“字符串(labelIncorrect) + +“)”)