Self-adaptive wafer defect identification method and system in semi-supervised field

Abstract

The invention discloses a self-adaptive wafer defect identification method and system in a semi-supervised field. The method comprises the following steps: acquiring a data set of a wafer defect source domain and a wafer defect target domain; the data set is preprocessed; constructing a cross alignment network, and respectively extracting image features of a source domain and a target domain of the preprocessed data set by using a feature extractor and an encoder; for the extracted image features, using a classifier to obtain an output probability, for unmarked data, determining and updating a false label threshold value of each category through a dynamic threshold value strategy, then using a sample-by-sample cross alignment method, a category-by-category cross alignment method and a consistency regularization method to calculate a loss function, and performing training; obtaining a trained cross alignment network; and the trained cross alignment network is used as a prediction model, and the wafer defect category for identifying the unmarked data of the target domain is output. According to the method, knowledge migration among different data sets can be effectively realized, and relatively high recognition accuracy is realized under the condition that the target domain data set only has a very small number of labels, for example, each category only has one label.

Description

Background Art

[0002] The production of wafers is a complex and lengthy process that requires highly sophisticated manufacturing technologies. Despite significant advances in semiconductor manufacturing in recent decades, wafer map defects are still unavoidable. Typically, dust, slight contamination, or manufacturing errors may cause defects on the wafer map. In most cases, defective chips on the wafer map converge to form specific patterns, such as centers, scratches, and rings. These defect patterns often indicate problems in the underlying manufacturing process. Identifying these defect patterns enables factories to pinpoint and resolve defects, thereby optimizing production workflows and increasing yields. Therefore, it is very important to quickly and accurately identify wafer map defects. This not only improves the accuracy and reliability of vision-based measurement systems, but also helps in the development of advanced instruments and measurement systems.

[0003] Wafer defect identification is a key step in semiconductor manufacturing and measurement, and is of great significance for optimizing production workflows and improving yields. Although current research has made significant progress in wafer defect identification, the lack of data labels remains a key issue in real production environments. A general solution can be to use existing large external labeled datasets to train the model and then test it on real industrial data. However, due to changes in production conditions, data from different sources often have different distributions, causing the model to show significant performance degradation. This problem is called domain shift. In deep learning, domain adaptation aims to alleviate domain shift by adjusting the distribution of the source and target domains.

[0004] Recently, semi-supervised domain adaptation has received more attention due to its practical applicability in real-world scenarios. Semi-supervised domain adaptation allows the use of a limited number of labeled samples from the target domain, which can improve the performance compared to unsupervised methods. However, the label mismatch problem in semi-supervised domain adaptation can also significantly affect the model performance.

[0005] An existing classification method is to preprocess the wafer defect image to be classified, input the preprocessed image into the wafer defect classification model, and then the wafer defect classification model outputs the category of the wafer defect, thereby completing the classification of the wafer defects. The wafer defect classification model is a stacking ensemble model. By establishing a stacked generalization model to improve the classification accuracy, especially the recognition performance of small sample categories, this method can effectively utilize the complementary information of different base learners and improve the adaptability of the overall model to complex data structures. It can not only effectively alleviate the challenges brought by unbalanced data sets to wafer defect detection, but also significantly improve the model's recognition accuracy for small sample defect types.

[0006] Another recognition method is to extract multiple wafer images from each of the nine wafer defect types in the wafer data set to form a wafer defect data set; convert the format and unify the size of the wafer defect data set to obtain a transformed wafer defect data set, and divide the transformed wafer defect data set into a training set and a test set; construct a deep residual network ResNet; input the training set into the deep residual network ResNet for training, and the loss function used in the training is the focal loss function; use the trained deep residual network ResNet to perform defect recognition on the wafer image to be recognized, and obtain the defect type of the wafer image. This method can be used to alleviate the problem of class imbalance in semiconductor wafer defect recognition.

[0007] The above methods are based on a large amount of labeled data, and rarely consider the situation where there are no labels or only very few labels; secondly, the existing technologies are all based on a single data set, the generalization of the methods is weak, and it is difficult to expand, and it is impossible to achieve the migration of models between different data sets. Summary of the invention

[0008] The technical problem to be solved by the present invention is to provide a semi-supervised domain adaptive wafer defect recognition method and system for solving the technical problems of domain offset and label mismatch in view of the deficiencies in the above-mentioned prior art.

[0009] The present invention adopts the following technical solutions: A semi-supervised domain adaptive wafer defect recognition method comprises the following steps: Acquire data sets of wafer defect source domain and target domain; and preprocess the data sets; Construct a cross-alignment network and use a feature extractor and encoder to extract image features of the source domain and target domain of the preprocessed dataset respectively; For the extracted image features, the classifier is used to obtain the output probability. For the unlabeled data, the pseudo-label threshold of each category is determined and updated through the dynamic threshold strategy. Then, the loss function is calculated and trained using the sample-by-sample cross alignment, class-by-class cross alignment and consistency regularization methods to obtain a trained cross alignment network. The trained cross-alignment network is used as a prediction model to output wafer defect categories that identify unlabeled data in the target domain.

[0010] Preferably, the data sets of the wafer defect source domain and the target domain are obtained from the public data sets WM-811K and MixedWM38, and the data sets are preprocessed as follows: Set all images in the public datasets WM-811K and MixedWM38 to 64x64 in size; use the formula: Normalize the image to ensure that the pixel value falls within the range of [-1,1]; then use random flipping and random cropping for preprocessing.

[0011] Preferably, in the cross-aligned network, represents the source domain labeled data, represents the target domain labeled data, Represents unlabeled data in the target domain; Using source domain encoder and the target domain encoder Separate the source domain and the target domain for learning; Source Encoder Map the data to the source domain space, the target domain encoder Map the data to the target domain space, and use the source domain classifier Labeling data in the source domain Perform classification and obtain the predicted probability value; design the target domain classifier and pseudo-label classifier Label the target domain data separately and target domain unlabeled data Perform classification and obtain the predicted probability value.

[0012] Preferably, the sample-by-sample cross alignment is specifically as follows: Source Sample After passing through the feature extractor, the source domain encoder and the target domain encoder obtain the features and ; and Through the classifier get and ; Determine the source domain loss function; For the target domain samples , after the feature extractor and encoder, we get and ; Through the classifier get , Through the classifier and After getting and .

[0013] Preferably, the loss function of the source domain is:

[0014]

[0015] in, is the hard label loss, is the soft label loss, is the mini-batch size, For source sample Tags, for After feature extractor, target domain encoder and classifier The output probability value after for After feature extractor, source domain encoder and classifier The output probability value after ; The loss function of the target domain is:

[0016]

[0017] in, is the indicator function, is the confidence threshold, The target domain sample The output value of the weakly enhanced version of for The corresponding pseudo-labels are for After feature extractor, source domain encoder and classifier The output probability value after is the KL divergence algorithm, for After feature extractor, target domain encoder and classifier The output probability value after .

[0018] Preferably, the cross-alignment by category is specifically as follows: For the source domain sample set , is the source domain sample and its corresponding label; and , is the source domain encoder, is the feature extractor, for The output features after the feature extractor and source domain encoder are: for The output features after the feature extractor and target domain encoder are: is the target domain encoder; exist and Perform comparative learning between them, minimize the distance between positive sample pairs, maximize the distance between negative sample pairs, and determine the source domain loss function; For the target domain sample , whose characteristics are expressed as and , first confirm the pseudo label value, and finally determine the source domain loss function.

[0019] Preferably, the source domain loss function is as follows:

[0020] in, , is the temperature parameter, for The positive sample indicator set; Target domain loss function as follows:

[0021] in, , for The positive sample indicator set, It is used to filter out samples whose probability is less than the confidence threshold. is the batch size, is the indicator function, for The positive sample of For Different samples.

[0022] Preferably, consistency regularization is used to ensure that the outputs of the target domain classifier and the pseudo-label classifier are consistent, as follows:

[0023] in, The target domain sample The output value of the weakly enhanced version of To correspond to the output of the strongly enhanced version, is the indicator function, is the confidence threshold, for The corresponding pseudo labels.

[0024] Preferably, in the dynamic threshold strategy, the final learning effect is applied to the scaling of the fixed threshold as follows:

[0025]

[0026] in, is the maximum value of the threshold, is the minimum value, is the indicator function, is the learning effect of class c after normalization at step t, is the learning effect of class c after considering the current learning state at step t, is the batch size, The target domain sample The output value of the weakly enhanced version of Refers to the category, is the dynamic threshold of the output; The learning effect of a class is defined as the ratio of the number of samples in each category that are predicted to fall into that class and whose probability is greater than the threshold to the number of samples predicted to fall into that class. Class in step length Learning effect for:

[0027] in, for Class in step The threshold value at The target domain sample The output value of the weakly enhanced version of .

[0028] In a second aspect, an embodiment of the present invention provides a semi-supervised domain adaptive wafer defect recognition system, comprising: The data module obtains the data sets of the wafer defect source domain and target domain, and pre-processes the data sets; The extraction module builds a cross-alignment network and uses a feature extractor and an encoder to extract image features of the source domain and target domain of the preprocessed dataset respectively; The recognition module uses a classifier to obtain the output probability for the extracted image features. For unlabeled data, a dynamic threshold strategy is used to determine and update the pseudo-label threshold of each category. Then, the cross-alignment learning and consistency regularization methods are used to calculate the loss function and perform training. The output module uses the trained cross-alignment network as a prediction model to output the wafer defect categories that identify the unlabeled data in the target domain.

[0029] Preferably, in the cross-aligned network, represents the source domain labeled data, represents the target domain labeled data, Represents unlabeled data in the target domain; Using source domain encoder and the target domain encoder Separate the source domain and the target domain for learning; Source Encoder Map the data to the source domain space, the target domain encoder Map the data to the target domain space, and use the source domain classifier Labeling data in the source domain Perform classification and obtain the predicted probability value; design the target domain classifier and pseudo-label classifier Label the target domain data separately and target domain unlabeled data Perform classification and obtain the predicted probability value.

[0030] Preferably, the sample-by-sample cross alignment is specifically as follows: Source Sample After passing through the feature extractor, the source domain encoder and the target domain encoder obtain the features and ; and Through the classifier get and ; Determine the source domain loss function; For the target domain samples , after the feature extractor and encoder, we get and ; Through the classifier get , Through the classifier and After getting and ; The loss function of the source domain is:

[0031]

[0032] in, is the hard label loss, is the soft label loss, is the mini-batch size, For source sample Tags, for After feature extractor, target domain encoder and classifier The output probability value after for After feature extractor, source domain encoder and classifier The output probability value after ; The loss function of the target domain is:

[0033]

[0034] in, is the indicator function, is the confidence threshold, The target domain sample The output value of the weakly enhanced version of for The corresponding pseudo-labels are for After feature extractor, source domain encoder and classifier The output probability value after is the KL divergence algorithm, for After feature extractor, target domain encoder and classifier The output probability value after .

[0035] Preferably, the cross-alignment by category is specifically as follows: For the source domain sample set , is the source domain sample and its corresponding label; and , is the source domain encoder, is the feature extractor, for The output features after the feature extractor and source domain encoder are: for The output features after the feature extractor and target domain encoder are: is the target domain encoder; exist and Perform comparative learning between them, minimize the distance between positive sample pairs, maximize the distance between negative sample pairs, and determine the source domain loss function; For the target domain sample , whose characteristics are expressed as and , first confirm the pseudo label value, and finally determine the source domain loss function; The source domain loss function is as follows:

[0036] in, , is the temperature parameter, for The positive sample indicator set; Target domain loss function as follows:

[0037] in, , for The positive sample indicator set, It is used to filter out samples whose probability is less than the confidence threshold. is the batch size, is the indicator function, for The positive sample of For Different samples.

[0038] Preferably, consistency regularization is used to ensure that the outputs of the target domain classifier and the pseudo-label classifier are consistent, as follows:

[0039] in, The target domain sample The output value of the weakly enhanced version of To correspond to the output of the strongly enhanced version, is the indicator function, is the confidence threshold, for The corresponding pseudo labels.

[0040] Preferably, in the dynamic threshold strategy, the final learning effect is applied to the scaling of the fixed threshold as follows:

[0041]

[0042] in, is the maximum value of the threshold, is the minimum value, is the indicator function, is the learning effect of class c after normalization at step t, is the learning effect of class c after considering the current learning state at step t, is the batch size, The target domain sample The output value of the weakly enhanced version of Refers to the category, is the dynamic threshold of the output; The learning effect of a class is defined as the ratio of the number of samples in each category that are predicted to fall into that class and whose probability is greater than the threshold to the number of samples predicted to fall into that class. Class in step length Learning effect for:

[0043] in, for Class in step The threshold value at The target domain sample The output value of the weakly enhanced version of .

[0044] In a third aspect, a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned semi-supervised field adaptive wafer defect identification method when executing the computer program.

[0045] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned semi-supervised field adaptive wafer defect identification method.

[0046] In a fifth aspect, a chip comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned semi-supervised field adaptive wafer defect identification method when executing the computer program.

[0047] In a sixth aspect, an embodiment of the present invention provides an electronic device, including a computer program, which, when executed by the electronic device, implements the steps of the above-mentioned semi-supervised field adaptive wafer defect identification method.

[0048] Compared with the prior art, the present invention has at least the following beneficial effects: A semi-supervised domain adaptive wafer defect recognition method can effectively realize knowledge transfer between different wafer defect data sets while keeping the framework basically unchanged, solve the label mismatch problem between data sets, and achieve a high recognition accuracy when the target domain data set has only a very small number of labels.

[0049] Furthermore, the dataset is normalized, and random flipping and random cropping are used for data enhancement to improve data diversity.

[0050] Furthermore, a cross-alignment network is proposed. This network avoids the influence of source supervision on target classification by decoupling the feature subspace and label space. In addition, the labeled target data is used to train a debiased target domain classifier, which can generate high-quality pseudo labels for unlabeled target data, thereby improving the overall performance.

[0051] Furthermore, a sample-by-sample cross alignment is proposed to align the classifier output probability value of each sample in the two subspaces, thereby improving the output consistency of the same sample between different encoders.

[0052] Furthermore, we propose a class-by-class cross-alignment method so that the representations of the same class in different domains are pulled together, while the representations of different classes are moved away from each other, thereby promoting feature clustering.

[0053] Furthermore, the weakly enhanced images are used to generate pseudo labels through the consistency regularization method to provide supervision signals for the strongly enhanced images, in order to ensure the output consistency of samples under different views.

[0054] Furthermore, in order to maximize the utility of unlabeled data, a dynamic threshold strategy is proposed. This strategy comprehensively considers the overall learning effect and the current specific learning effect of each class, and accurately reflects the learning status of each class. Compared with the fixed threshold, this dynamic strategy significantly improves the learning efficiency and optimizes the data utilization.

[0055] It can be understood that the beneficial effects of the second to sixth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here.

[0056] In summary, the present invention can effectively realize knowledge transfer between different data sets, and achieve a high recognition accuracy when the target domain data set has only a very small number of labels, such as only one label for each category.

[0057] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0059] Figure 1 This is a schematic diagram of the cross-alignment network; Figure 2 It is a schematic diagram of the process of the present invention; Figure 3 A schematic diagram of a computer device provided by an embodiment of the present invention; Figure 4 The block diagram is a chip provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] In the description of the present invention, it should be understood that the terms “include” and “comprises” indicate the presence of described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0062] It should also be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0063] It should be further understood that the term "and / or" used in the present specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes these combinations. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in the present invention generally indicates that the associated objects are in an "or" relationship.

[0064] It should be understood that, although the terms first, second, third, etc. may be used to describe preset ranges, etc. in the embodiments of the present invention, these preset ranges should not be limited to these terms. These terms are only used to distinguish preset ranges from each other. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0065] The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0066] Various structural schematic diagrams of the embodiments disclosed in the present invention are shown in the accompanying drawings. These figures are not drawn to scale, and some details are magnified and some details may be omitted for the purpose of clear expression. The shapes of various regions and layers shown in the figures and the relative sizes and positional relationships therebetween are only exemplary, and may deviate in practice due to manufacturing tolerances or technical limitations, and those skilled in the art may additionally design regions / layers with different shapes, sizes, and relative positions according to actual needs.

[0067] The present invention provides a semi-supervised domain adaptive wafer defect recognition method, which adopts a cross-alignment network and uses different encoders and classifiers to decouple the feature subspace and label space. In order to promote the knowledge transfer between the source domain and the target domain, the present invention also proposes a cross-alignment learning method; the cross-alignment learning consists of sample cross-alignment and class cross-alignment, which aims to ensure the consistency of the encoder output and promote the formation of robust clusters. In addition, the method also integrates consistency regularization to enforce the consistency of the output between different views of the sample, and also introduces a dynamic threshold strategy to tailor the threshold for each class according to its learning effect, thereby improving the efficiency of the model and data utilization.

[0068] The present invention focuses on domain adaptation scenarios, which can effectively solve the problem of lack of data labels in real industrial production processes and has better generalization of application scenarios; the present invention decouples the label space, thereby avoiding the influence of source supervision on target classification and effectively solving the problem of possible label mismatch; the present invention uses cross-alignment learning to fully promote knowledge interaction and fusion between data sets and improve the stability of the algorithm; the present invention uses a dynamic threshold strategy to accurately reflect the learning status of each class, effectively improves the learning efficiency of the model, and optimizes data utilization.

[0069] See also Figure 2 The present invention provides a semi-supervised field adaptive wafer defect recognition method, comprising the following steps: S1. Obtain the datasets of wafer defect source domain and target domain by self-collection or online download of public datasets; then, preprocess the dataset data, including random horizontal flipping, random cropping, and normalization; The public datasets WM-811K and MixedWM38 are used. These two datasets are the most widely used in the field of wafer image defect recognition. Each dataset contains the same nine categories. Each wafer image consists of three pixel values: 0, 1, and 2, where 0 represents the background, 1 represents a normal working chip, and 2 represents a defective chip, which are displayed in dark purple, green, and yellow in the image, respectively. WM-811K contains 811,457 wafer defect images, of which 172,950 are labeled images for experiments. MixedWM38 is a dataset for identifying mixed defect patterns, including 8 single-type defects and 29 mixed-type defects. Mixed-type defects are a combination of single-type defects. In the present invention, in order to meet the conditions of domain adaptation, the present invention only uses a single type of defect for experiments. Both datasets exhibit class imbalance problems, of which WM-811K is more serious because the "none" class accounts for 85% of the dataset.

[0070] In order to standardize the image size, the present invention sets the size of all images in the two datasets to 64x64; Next, use the formula: Normalize the image to ensure that the pixel values ​​fall within the range of [-1, 1]; Then preprocess the data using random flipping, random cropping, etc.

[0071] In consistency regularization, a powerful data augmentation technique, namely random augmentation RandAugment, is used.

[0072] S2, build a cross-alignment network and use a shared feature extractor and encoder to extract image features from the source domain and the target domain; See also Figure 1 , in the cross-aligned network, represents the source domain labeled data, represents the target domain labeled data, Represents the target domain unlabeled data. This invention uses the source domain encoder and the target domain encoder The source domain and target domain are separated for learning; the source domain encoder maps the data to the source domain space, and the target domain encoder maps the data to the target domain space, and then classifies them; Source domain uses source domain classifier , for the target domain, design the target domain classifier and pseudo-label classifier ; The target domain classifier is optimized only using the target domain labeled data, which can effectively avoid the training bias caused by using pseudo labels. At the same time, the target domain classifier also provides pseudo labels for the pseudo label classifier, and the pseudo labels are the category labels with the maximum probability greater than the confidence threshold. Only the target domain classifier will be used during testing. This design keeps the label space separate, thus effectively solving the label mismatch problem.

[0073] S3. For the extracted image features, the classifier is used to obtain the output probability. For the unlabeled data, the pseudo-label threshold of each category is determined and updated through the dynamic threshold strategy. Then, the loss function is calculated and trained using the cross-alignment learning and consistency regularization methods. Use sample-by-sample and class-by-class cross alignment to ensure the output consistency of the encoders in the source domain and the target domain. Use consistency regularization to ensure the output consistency of the target domain classifier and the pseudo-label classifier. Use a dynamic threshold strategy to determine and update the pseudo-label threshold for each category. The specific steps are as follows: S301, sample-by-sample cross alignment In order to ensure the consistency of sample output on the two encoders, a sample-by-sample cross alignment method is first proposed. The main idea of ​​the method is to align the logical values ​​of the classifier output of each sample in the two subspaces. In order to enhance the knowledge interaction between the two domains, a method combining hard labels and soft labels is used for training. Hard labels are traditional classification labels, usually in the form of one-hot encoding. Unlike hard labels, the values ​​of soft labels are continuous, contain the probability distribution of each category, and contain more information. The combination of the two methods helps the two encoders transfer richer knowledge.

[0074] First, the source sample After passing through the feature extractor, the source domain encoder and the target domain encoder obtain the features and ; after, and Through the classifier get and ; Next, determine the source domain loss function; For the target domain sample , after the feature extractor and encoder, we get and ; Then, Through the classifier get , fut passes through the classifier and After getting and .

[0075] The loss function of the source domain is:

[0076]

[0077] in, is the hard label loss, is the soft label loss. is the mini-batch size. Kl divergence is used here to calculate the similarity of the output logical values.

[0078] The loss function of the target domain is:

[0079]

[0080] in, is the indicator function, is the confidence threshold.

[0081] S302, cross alignment by category In order to make samples from different domains form consistent and stable clusters, a class-by-class cross-alignment is introduced, which means that the representations of the same class in different domains are pulled together, while the representations of different classes are moved away from each other. This approach can promote the consistency of the outputs of the two encoders and enhance feature learning.

[0082] For source domain samples ; get and ; In order to align the two feature distributions, we propose and We do contrastive learning between them, hoping to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs.

[0083] For the target domain sample , whose characteristics are expressed as and ;because They are all unlabeled data, so the pseudo-label values ​​must be confirmed first; Determine the loss function of the source domain and the target domain; The source domain loss function is as follows:

[0084] in, , is the temperature parameter, for The positive sample indicator set.

[0085] The target domain loss function is as follows:

[0086] in, , for The positive sample indicator set, It is used to filter out samples whose probability is less than the confidence threshold.

[0087] S303, the purpose of the consistency regularization method is to ensure the output consistency of the sample under different views. This method generally works on unlabeled data in the target domain and is implemented through pseudo labels. Usually, weakly enhanced images are used to generate pseudo labels to provide supervision signals for strongly enhanced images. The cross entropy loss function is used for learning consistency regularization, as follows:

[0088] in, is the target domain sample The output value of the weakly enhanced version of This corresponds to the output of the strongly enhanced version.

[0089] S304, Dynamic Threshold Strategy Currently, most semi-supervised domain adaptation methods rely on a fixed and high threshold to select pseudo labels, ignoring the learning differences between different categories. All methods require new threshold definitions.

[0090] The ideal situation should be that different categories set different thresholds according to the learning effect. This application defines the learning effect of a class as the ratio of the number of samples in each category that are predicted to fall into that class and have a probability greater than the threshold to the number of samples predicted to fall into that class. The formula is:

[0091] in, for Class in step length The learning effect of for Class in step length The threshold value at . Only a single mini-batch of samples is processed, thus representing the current learning effect of each class.

[0092] In order to comprehensively consider the learning situation of previous samples, the overall learning effect is estimated and normalized, which is expressed as:

[0093]

[0094] After normalization, the class with the best learning state The value is 1. However, the normalized value only considers the relative learning between classes, without considering the learning of the entire batch, which is not conducive to learning. The learning of a batch can be expressed as the ratio of the number of samples in the batch with the maximum probability greater than the threshold to the batch size.

[0095] Taking all factors into consideration, the final learning effect is obtained and applied to the scaling of fixed thresholds:

[0096]

[0097] Among them, two user-defined thresholds are introduced , , represents the maximum value of the threshold, represents the minimum value, Will be scaled based on study status. Used to prevent the threshold from being too small.

[0098] S4. Use the trained cross-alignment network as the prediction model, retain the output result of the target domain classifier as the final value predicted by the cross-alignment network, and output the wafer defect category that identifies the unlabeled data in the target domain.

[0099] It will be appreciated by those skilled in the art that various aspects of the present invention may be implemented as systems, methods or program products. Therefore, various aspects of the present invention may be specifically implemented in the following forms, namely: complete hardware implementation, complete software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, which may be collectively referred to herein as "circuits", "modules" or "platforms".

[0100] In another embodiment of the present invention, a semi-supervised domain adaptive wafer defect recognition system is provided. The system can be used to implement the above-mentioned semi-supervised domain adaptive wafer defect recognition method. Specifically, the semi-supervised domain adaptive wafer defect recognition system includes a data module, an extraction module, a recognition module and an output module.

[0101] Among them, the data module obtains the data sets of the wafer defect source domain and the target domain; and preprocesses the data sets; The public datasets WM-811K and MixedWM38 are used. These two datasets are the most widely used in the field of wafer image defect recognition. Each dataset contains the same nine categories. Each wafer image consists of three pixel values: 0, 1, and 2, where 0 represents the background, 1 represents a normal working chip, and 2 represents a defective chip, which are displayed in dark purple, green, and yellow in the image, respectively. WM-811K contains 811,457 wafer defect images, of which 172,950 are labeled images for experiments. MixedWM38 is a dataset for identifying mixed defect patterns, including 8 single-type defects and 29 mixed-type defects. Mixed-type defects are a combination of single-type defects. In the present invention, in order to meet the conditions of domain adaptation, the present invention only uses a single type of defect for experiments. Both datasets exhibit class imbalance problems, of which WM-811K is more serious because the "none" class accounts for 85% of the dataset.

[0102] In order to standardize the image size, the present invention sets the size of all images in the two datasets to 64x64; Next, use the formula: Normalize the image to ensure that the pixel values ​​fall within the range of [-1, 1]; Then preprocess the data using random flipping, random cropping, etc.

[0103] In consistency regularization, a powerful data augmentation technique, namely random augmentation RandAugment, is used.

[0104] The extraction module builds a cross-alignment network and uses a feature extractor and an encoder to extract image features of the source domain and target domain of the preprocessed dataset respectively; In the cross-aligned network, represents the source domain labeled data, represents the target domain labeled data, Represents unlabeled data in the target domain; Using source domain encoder and the target domain encoder Separate the source domain and the target domain for learning; Source Encoder Map the data to the source domain space, the target domain encoder Map the data to the target domain space, and use the source domain classifier Labeling data in the source domain Perform classification and obtain the predicted probability value; design the target domain classifier and pseudo-label classifier Label the target domain data separately and target domain unlabeled data Perform classification and obtain the predicted probability value.

[0105] The recognition module uses a classifier to obtain the output probability for the extracted image features. For unlabeled data, the pseudo-label threshold of each category is determined and updated through a dynamic threshold strategy. Then, the loss function is calculated and trained using sample-by-sample cross alignment, class-by-class cross alignment, and consistency regularization methods to obtain a trained cross-alignment network. The specific sample-by-sample cross alignment is: Source Sample After passing through the feature extractor, the source domain encoder and the target domain encoder obtain the features and ; and Through the classifier get and ; Determine the source domain loss function; For the target domain samples , after the feature extractor and encoder, we get and ; Through the classifier get , Through the classifier and After getting and .

[0106] The loss function of the source domain is:

[0107]

[0108] in, is the hard label loss, is the soft label loss, is the mini-batch size, For source sample Tags, for After feature extractor, target domain encoder and classifier The output probability value after for After feature extractor, source domain encoder and classifier The output probability value after ; The loss function of the target domain is:

[0109]

[0110] in, is the indicator function, is the confidence threshold, The target domain sample The output value of the weakly enhanced version of for The corresponding pseudo-labels are for After feature extractor, source domain encoder and classifier The output probability value after is the KL divergence algorithm, for After feature extractor, target domain encoder and classifier The output probability value after .

[0111] The specific cross-alignment by category is: For the source domain sample set , is the source domain sample and its corresponding label; and , is the source domain encoder, is the feature extractor, for The output features after the feature extractor and source domain encoder are: for The output features after the feature extractor and target domain encoder are: is the target domain encoder; exist and Perform comparative learning between them, minimize the distance between positive sample pairs, maximize the distance between negative sample pairs, and determine the source domain loss function; For the target domain sample , whose characteristics are expressed as and , first confirm the pseudo label value, and finally determine the source domain loss function.

[0112] The source domain loss function is as follows:

[0113] in, , is the temperature parameter, for The positive sample indicator set; Target domain loss function as follows:

[0114] in, , for The positive sample indicator set, It is used to filter out samples whose probability is less than the confidence threshold. is the batch size, is the indicator function, for The positive sample of For Different samples.

[0115] Use consistency regularization to ensure that the outputs of the target domain classifier and the pseudo-label classifier are consistent, as follows:

[0116] in, The target domain sample The output value of the weakly enhanced version of To correspond to the output of the strongly enhanced version, is the indicator function, is the confidence threshold, for The corresponding pseudo labels.

[0117] In the dynamic threshold strategy, the final learning effect is applied to the scaling of the fixed threshold as follows:

[0118]

[0119] in, is the maximum value of the threshold, is the minimum value, is the indicator function, is the learning effect of class c after normalization at step t, is the learning effect of class c after considering the current learning state at step t, is the batch size, The target domain sample The output value of the weakly enhanced version of Refers to the category, is the dynamic threshold of the output; The learning effect of a class is defined as the ratio of the number of samples in each category that are predicted to fall into that class and whose probability is greater than the threshold to the number of samples predicted to fall into that class. Class in step length Learning effect for:

[0120] in, for Class in step The threshold value at The target domain sample The output value of the weakly enhanced version of .

[0121] The output module uses the trained cross-alignment network as a prediction model to output the wafer defect categories that identify the unlabeled data in the target domain.

[0122] In another embodiment of the present invention, a terminal device is provided, which includes a processor and a memory, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or other general-purpose processors, graphics processors (GPU), tensor processors (TPU), digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, which are suitable for implementing one or more instructions, and are specifically suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding functions; the processor described in the embodiment of the present invention can be used for the operation of the semi-supervised field adaptive wafer defect recognition method, including: Obtain wafer defect source and target domain datasets; preprocess the datasets; construct a cross-alignment network, use a feature extractor and an encoder to extract image features of the source and target domains of the preprocessed datasets respectively; use a classifier to obtain the output probability for the extracted image features, and for unlabeled data, determine and update the pseudo-label threshold of each category through a dynamic threshold strategy, then use sample-by-sample cross-alignment, class-by-class cross-alignment, and consistency regularization methods to calculate the loss function and perform training to obtain a trained cross-alignment network; use the trained cross-alignment network as a prediction model to output and identify the wafer defect category of unlabeled data in the target domain.

[0123] See also Figure 3 , the terminal device is a computer device, and the computer device 60 of this embodiment includes: a processor 61, a memory 62, and a computer program 63 stored in the memory 62 and executable on the processor 61. When the computer program 63 is executed by the processor 61, the semi-supervised domain adaptive wafer defect recognition method in the embodiment is implemented. To avoid repetition, it is not described one by one here. Alternatively, when the computer program 63 is executed by the processor 61, the functions of each model / unit in the semi-supervised domain adaptive wafer defect recognition system in the embodiment are implemented. To avoid repetition, it is not described one by one here.

[0124] The computer device 60 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The computer device 60 may include, but is not limited to, a processor 61 and a memory 62. Those skilled in the art will appreciate that Figure 3 This is only an example of the computer device 60 and does not constitute a limitation of the computer device 60. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the computer device may also include input and output devices, network access devices, buses, etc.

[0125] The processor 61 may be a central processing unit (CPU), or other general-purpose processors, graphics processing units (GPU), tensor processing units (TPU), digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.

[0126] The memory 62 may be an internal storage unit of the computer device 60, such as a hard disk or memory of the computer device 60. The memory 62 may also be an external storage device of the computer device 60, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc., equipped on the computer device 60.

[0127] Furthermore, the memory 62 may include both an internal storage unit of the computer device 60 and an external storage device. The memory 62 is used to store computer programs and other programs and data required by the computer device. The memory 62 may also be used to temporarily store data that has been output or is to be output.

[0128] See also Figure 4, the terminal device is a chip, and the chip 600 of this embodiment includes a processor 622, which may be one or more, and a memory 632 for storing a computer program executable by the processor 622. The computer program stored in the memory 632 may include one or more modules, each corresponding to a set of instructions. In addition, the processor 622 may be configured to execute the computer program to perform the above-mentioned semi-supervised domain adaptive wafer defect recognition method.

[0129] In addition, the chip 600 may further include a power supply component 626 and a communication component 650, wherein the power supply component 626 may be configured to perform power management of the chip 600, and the communication component 650 may be configured to implement communication, for example, wired or wireless communication, of the chip 600. In addition, the chip 600 may further include an input / output interface 658. The chip 600 may operate based on an operating system stored in the memory 632.

[0130] In another embodiment of the present invention, the present invention also provides a storage medium, specifically a computer-readable storage medium, which is a memory device in a terminal device for storing programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and the extended storage medium supported by the terminal device. The computer-readable storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs. It should be noted that the computer-readable storage medium here can be a high-speed RAM memory, or a non-volatile memory (Non-VolatileMemory), such as at least one disk memory.

[0131] One or more instructions stored in a computer-readable storage medium may be loaded and executed by a processor to implement the corresponding steps of the semi-supervised field adaptive wafer defect recognition method in the above embodiment; one or more instructions in the computer-readable storage medium may be loaded and executed by a processor as follows: Obtain wafer defect source and target domain datasets; preprocess the datasets; construct a cross-alignment network, use a feature extractor and an encoder to extract image features of the source and target domains of the preprocessed datasets respectively; use a classifier to obtain the output probability for the extracted image features, and for unlabeled data, determine and update the pseudo-label threshold of each category through a dynamic threshold strategy, then use sample-by-sample cross-alignment, class-by-class cross-alignment, and consistency regularization methods to calculate the loss function and perform training to obtain a trained cross-alignment network; use the trained cross-alignment network as a prediction model to output and identify the wafer defect category of unlabeled data in the target domain.

[0132] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention described and shown in the drawings here can usually be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0133] The feature extraction backbone network used in the present invention can be replaced by networks other than Resnet-34, such as VGG, AlexNet, etc. The vision-based wafer defect information input in the present invention can theoretically be replaced by other types of data, such as visual photography information, etc.

[0134] The core of the cross-aligned network structure is to achieve feature alignment between domains, thereby reducing the distribution difference between the source domain and the target domain. It involves the use of two sub-networks (one for the source domain and the other for the target domain), as well as a joint feature representation. Through the loss function in the optimization process, the feature representations of the two domains are made as close as possible in the same feature space. The cross-aligned network structure shows higher performance in the wafer defect recognition task, which is mainly due to its efficient use of domain adaptive learning. By combining semi-supervised learning and domain adaptive technology, the model can effectively utilize unlabeled data under limited labeled data, thereby improving the generalization ability and recognition accuracy of the model.

[0135] The cross-aligned network structure used by the method of the present invention has been well verified on the WM-811K and MixedWM38 public datasets, and its performance is improved compared with the current mainstream models, proving that the method is feasible.

[0136] In summary, the present invention provides a semi-supervised domain adaptive wafer defect recognition method and system, proposes a cross-alignment network, decouples the label space to achieve semi-supervised domain adaptive learning between two data sets, and achieves high-accuracy recognition when the target data set contains very few labels, such as only one or three labels per class. In addition, the present invention also proposes a cross-alignment learning method to effectively ensure the consistency of the encoder output to promote knowledge transfer between the source domain and the target domain. And also introduces a dynamic threshold strategy to tailor the threshold for each class according to its learning effect, thereby improving learning efficiency and data utilization.

[0137] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0138] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0139] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the present invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0140] In the embodiments provided by the present invention, it should be understood that the disclosed devices / terminals and methods can be implemented in other ways. For example, the device / terminal embodiments described above are only schematic. For example, the division of the modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0141] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0142] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0143] If the integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0144] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices, and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0145] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0146] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0147] The above contents are only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A semi-supervised domain adaptive wafer defect recognition method, characterized in that: The following steps are involved: Acquire data sets of wafer defect source domain and target domain; and preprocess the data sets; Construct a cross-alignment network and use a feature extractor and encoder to extract image features of the source domain and target domain of the preprocessed dataset respectively; For the extracted image features, the classifier is used to obtain the output probability. For the unlabeled data, the pseudo-label threshold of each category is determined and updated through the dynamic threshold strategy. Then, the loss function is calculated and trained using the sample-by-sample cross alignment, class-by-class cross alignment and consistency regularization methods to obtain a trained cross alignment network. The trained cross-alignment network is used as a prediction model to output wafer defect categories that identify unlabeled data in the target domain.

2. The semi-supervised domain adaptive wafer defect recognition method according to claim 1, characterized in that: The datasets of wafer defect source domain and target domain are obtained from the public datasets WM-811K and MixedWM38, and the datasets are preprocessed as follows: Set all images in the public datasets WM-811K and MixedWM38 to 64x64 in size; use the formula: Normalize the image to ensure that the pixel value falls within the range of [-1,1]; then use random flipping and random cropping for preprocessing.

3. The semi-supervised domain adaptive wafer defect recognition method according to claim 1, characterized in that: In the cross-aligned network, represents the source domain labeled data, represents the target domain labeled data, Represents unlabeled data in the target domain; Using source domain encoder and the target domain encoder Separate the source domain and the target domain for learning; Source Domain Encoder Map the data to the source domain space, the target domain encoder Map the data to the target domain space, and use the source domain classifier Labeling data in the source domain Perform classification and obtain the predicted probability value; design the target domain classifier and pseudo-label classifier Label the target domain data separately and target domain unlabeled data Perform classification and obtain the predicted probability value.

4. The semi-supervised domain adaptive wafer defect recognition method according to claim 1, characterized in that: The specific sample-by-sample cross alignment is: Source Sample After passing through the feature extractor, the source domain encoder and the target domain encoder obtain the features and ; and Through the classifier get and ; Determine the source domain loss function; For the target domain samples , after the feature extractor and encoder, we get and ; Through the classifier get , Through the classifier and After getting and .

5. The semi-supervised domain adaptive wafer defect recognition method according to claim 4, characterized in that: The loss function of the source domain is: in, is the hard label loss, is the soft label loss, is the mini-batch size, For source sample Tags, for After feature extractor, target domain encoder and classifier The output probability value after for After feature extractor, source domain encoder and classifier The output probability value after ; The loss function of the target domain is: in, is the indicator function, is the confidence threshold, The target domain sample The output value of the weakly enhanced version of for The corresponding pseudo-labels are for After feature extractor, source domain encoder and classifier The output probability value after is the KL divergence algorithm, for After feature extractor, target domain encoder and classifier The output probability value after .

6. The semi-supervised domain adaptive wafer defect recognition method according to claim 1, characterized in that: The specific cross-alignment by category is: For the source domain sample set , is the source domain sample and its corresponding label; get and , is the source domain encoder, is the feature extractor, for The output features after the feature extractor and source domain encoder are: for The output features after the feature extractor and target domain encoder are: is the target domain encoder; exist and Perform comparative learning between them, minimize the distance between positive sample pairs, maximize the distance between negative sample pairs, and determine the source domain loss function; For the target domain sample , whose characteristics are expressed as and , first confirm the pseudo label value, and finally determine the source domain loss function.

7. The semi-supervised domain adaptive wafer defect recognition method according to claim 6, characterized in that: The source domain loss function is as follows: in, , is the temperature parameter, for The positive sample indicator set; Target domain loss function as follows: in, , for The positive sample indicator set, It is used to filter out samples whose probability is less than the confidence threshold. is the batch size, is the indicator function, for The positive sample of For Different samples.

8. The semi-supervised domain adaptive wafer defect recognition method according to claim 1, characterized in that: Use consistency regularization to ensure that the outputs of the target domain classifier and the pseudo-label classifier are consistent, as follows: in, The target domain sample The output value of the weakly enhanced version of To correspond to the output of the strongly enhanced version, is the indicator function, is the confidence threshold, for The corresponding pseudo labels.

9. The semi-supervised domain adaptive wafer defect recognition method according to claim 1, characterized in that: In the dynamic threshold strategy, the final learning effect is applied to the scaling of the fixed threshold as follows: in, is the maximum value of the threshold, is the minimum value, is the indicator function, is the learning effect of class c after normalization at step t, is the learning effect of class c after considering the current learning state at step t, is the batch size, The target domain sample The output value of the weakly enhanced version of Refers to the category, is the dynamic threshold of the output; The learning effect of a class is defined as the ratio of the number of samples in each category that are predicted to fall into that class and whose probability is greater than the threshold to the number of samples predicted to fall into that class. Class in step length Learning effect for: in, for Class in step The threshold value at The target domain sample The output value of the weakly enhanced version of .

10. A semi-supervised domain adaptive wafer defect recognition system, characterized in that: include: The data module obtains the data sets of the wafer defect source domain and target domain; And preprocess the data set; The extraction module builds a cross-alignment network and uses a feature extractor and an encoder to extract image features of the source domain and target domain of the preprocessed dataset respectively; The recognition module uses a classifier to obtain the output probability for the extracted image features. For unlabeled data, the pseudo-label threshold of each category is determined and updated through a dynamic threshold strategy. Then, the loss function is calculated and trained using sample-by-sample cross alignment, class-by-class cross alignment, and consistency regularization methods to obtain a trained cross-alignment network. The output module uses the trained cross-alignment network as a prediction model to output the wafer defect categories that identify the unlabeled data in the target domain.

Images