A semi-supervised domain adaptive wafer defect recognition method and system

Abstract

The application discloses a kind of semi-supervised field self-adaptive wafer defect identification method and system, obtain the data set of wafer defect source field and target field;And the data set is preprocessed;Cross alignment network is constructed, and the image features of source field and target field of the data set after pre-processing are respectively extracted using feature extractor and encoder;For the extracted image feature, output probability is obtained using classifier, for unlabelled data, determine and update the pseudo-label threshold of each class by dynamic threshold strategy, then use sample-by-sample cross alignment, class-by-class cross alignment and consistency regularization method to calculate loss function and train, obtain the trained cross alignment network;Using the trained cross alignment network as prediction model, the wafer defect class of the unlabelled data of target field is output to be identified.The present application can effectively realize the knowledge transfer between different data sets, and achieve high recognition accuracy in the case that the target domain data set has only a small amount of labels, such as only one label per class.

Description

Technical Field

[0001] This invention belongs to the field of wafer defect identification technology, specifically relating to a semi-supervised adaptive wafer defect identification method and system. Background Technology

[0002] Wafer fabrication is a complex and lengthy process requiring highly sophisticated manufacturing techniques. Despite significant advancements in semiconductor manufacturing over the past few decades, wafer pattern defects remain unavoidable. Dust, minor contamination, or manufacturing errors can typically lead to defects on wafer patterns. In most cases, defective chips on the wafer pattern converge to form specific patterns such as centers, scratches, and rings. These defect patterns often indicate underlying problems in the manufacturing process. Identifying these defect patterns allows factories to pinpoint and resolve defects, thereby optimizing production workflows and increasing throughput. Therefore, the rapid and accurate identification of wafer pattern defects is crucial. This not only improves the accuracy and reliability of vision-based measurement systems but also contributes to the development of advanced instruments and measurement systems.

[0003] Wafer defect identification is a critical step in semiconductor manufacturing and measurement, significantly impacting production workflows and increasing yield. While current research has made substantial progress in wafer defect identification, the lack of data labels remains a critical issue in real-world production environments. A common solution is to train a model using existing large-scale exogenous labeled datasets and then test it on real-world industrial data. However, due to varying production conditions, data from different sources often exhibit different distributions, leading to significant performance degradation in the model. This problem is known as domain shift. In deep learning, domain adaptation aims to mitigate domain shift by adjusting the distributions of the source and target domains.

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

[0005] One existing classification method involves preprocessing the image of the wafer defect to be classified, then inputting the preprocessed image into a wafer defect classification model. The model then outputs the category of the wafer defect, thus completing the classification. This wafer defect classification model is a stacking ensemble model. By establishing a stacked generalization model, classification accuracy is improved, especially for identifying small sample categories. This method effectively utilizes the complementary information of different base learners, improving the overall model's adaptability to complex data structures. It not only effectively alleviates the challenges posed by imbalanced datasets to wafer defect detection but also significantly improves the model's accuracy in identifying small sample defect types.

[0006] Another identification method involves extracting multiple wafer images from each of the nine wafer defect types in a wafer dataset to form a wafer defect dataset. This dataset is then format-converted and resized to obtain a transformed wafer defect dataset, which is divided into training and testing sets. A deep residual network (ResNet) is constructed, and the training set is input into the ResNet for training using the focal loss function. The trained ResNet is then used to identify defects in the wafer images to determine the defect types. This method can alleviate the class imbalance problem in semiconductor wafer defect identification.

[0007] The above methods rely on a large amount of labeled data and rarely consider the case of no labels or very few labels. Secondly, existing technologies are all based on a single dataset, which results in weak generalization, difficulty in expansion, and inability to transfer the model between different datasets. 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 identification method and system to address the shortcomings of the prior art, thereby solving the technical problems of domain offset and label mismatch.

[0009] The present invention adopts the following technical solution:

[0010] A semi-supervised domain adaptive wafer defect identification method includes the following steps:

[0011] Obtain datasets of wafer defect source and target domains; and preprocess the datasets.

[0012] Construct a cross-alignment network and use a feature extractor and an encoder to extract image features from the source and target domains of the preprocessed dataset, respectively;

[0013] For the extracted image features, a classifier is used to obtain the output probability. For unlabeled data, a dynamic thresholding strategy is used to determine and update the pseudo-label threshold for each category. 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 the trained cross-alignment network.

[0014] The trained cross-alignment network is used as a prediction model to output the wafer defect category that identifies unlabeled data in the target domain.

[0015] Preferably, the datasets of wafer defect source and target domains are obtained from the publicly available datasets WM-811K and MixedWM38, and the datasets are preprocessed as follows:

[0016] Set all images in the public datasets WM-811K and MixedWM38 to 64x64; use the formula: The image is normalized to ensure that the pixel values ​​fall within the range of [-1, 1]; then, random flipping and random cropping are used for preprocessing.

[0017] Preferably, in a cross-aligned network, Represents source domain labeled data, Represents the target domain labeled data. This represents unlabeled data in the target domain;

[0018] Use source domain encoder and target domain encoder Learn the source and target domains separately;

[0019] Source domain encoder Mapping data to the source domain space, target domain encoder Map the data to the target domain space, and use the source domain classifier in the source domain. Label data for source domain Perform classification to obtain predicted probability values; design a target domain classifier. and pseudo-label classifier Label the data in the target domain respectively and unlabeled data in the target domain Classify the data to obtain predicted probability values.

[0020] Preferably, sample-by-sample cross-alignment specifically involves:

[0021] Source Sample The features are obtained after passing through a feature extractor, source domain encoder, and target domain encoder. and ; and After classifier get and Determine the source domain loss function; for target domain samples After passing through the feature extractor and encoder, the result is... and ; After classifier get , After classifier and Later obtained and .

[0022] Preferably, the loss function of the source domain is:

[0023]

[0024]

[0025] in, It is hard label loss. For soft label loss, It is mini-batch size. For source samples The tag, for After feature extractor, target domain encoder and classifier The output probability value after that, for After feature extractor, source domain encoder and classifier The output probability value;

[0026] The loss function for the target domain is:

[0027]

[0028]

[0029] in, For indicator functions, The confidence threshold is... For target domain samples The output value of the weakly enhanced version, for The corresponding pseudo-tags, for After feature extractor, source domain encoder and classifier The output probability value after that, The KL divergence algorithm is used. for After feature extractor, target domain encoder and classifier The output probability value.

[0030] Preferably, the cross-alignment by category specifically refers to:

[0031] For the source domain sample set , For the source domain samples and their corresponding labels; obtain and , For source domain encoder, For feature extractor, for The output features after passing through the feature extractor and source domain encoder for The output features after passing through the feature extractor and the target domain encoder For target domain encoder;

[0032] exist and By performing comparative learning between positive sample pairs, the distance between positive sample pairs is minimized while the distance between negative sample pairs is maximized, and the source domain loss function is determined.

[0033] For target domain samples Its characteristics are represented as and First, confirm the pseudo-label values, and then determine the source domain loss function.

[0034] Preferably, the source domain loss function is as follows:

[0035]

[0036] in, , For temperature parameters, for The set of positive sample indicators;

[0037] target domain loss function as follows:

[0038]

[0039] in, , for The positive sample index set, Samples with a probability less than the confidence threshold are filtered out. The size of the batch. For indicator functions, for Positive samples, To and Different samples.

[0040] Preferably, consistency regularization is used to ensure that the outputs of the target domain classifier and the pseudo-label classifier are consistent, as follows:

[0041]

[0042] in, For target domain samples The output value of the weakly enhanced version, To correspond with the significantly enhanced output, For indicator functions, The confidence threshold is... for The corresponding pseudo-tag.

[0043] Preferably, in the dynamic threshold strategy, the final learning effect is applied to the scaling of the fixed threshold as follows:

[0044]

[0045]

[0046] in, The maximum value of the threshold. To be the minimum value, For indicator functions, The learning effect of class c after normalization at step t. For class c, consider the learning effect after the current learning state at step t. The size of the batch. For target domain samples The output value of the weakly enhanced version, As a reference to a category, The output dynamic threshold;

[0047] The learning effect of a class is defined as the ratio of the number of samples predicted to fall into that class with a probability greater than a threshold to the total number of samples predicted to fall into that class. Class in step size Learning effect at the place for:

[0048]

[0049] in, for Class in steps The threshold at that point For target domain samples The output value of the weakly enhanced version.

[0050] Secondly, embodiments of the present invention provide a semi-supervised domain adaptive wafer defect identification system, comprising:

[0051] The data module acquires datasets of wafer defect source and target domains and preprocesses these datasets.

[0052] The extraction module constructs a cross-alignment network and uses a feature extractor and encoder to extract image features from the source and target domains of the preprocessed dataset, respectively.

[0053] The recognition module uses a classifier to obtain the output probability for the extracted image features. For unlabeled data, it determines and updates the pseudo-label threshold for each category through a dynamic thresholding strategy. Then, it uses cross-alignment learning and consistency regularization methods to calculate the loss function and train the module.

[0054] The output module uses a trained cross-alignment network as a prediction model to output the wafer defect category that identifies the unlabeled data in the target domain.

[0055] Preferably, in a cross-aligned network, Represents source domain labeled data, Represents the target domain labeled data. This represents unlabeled data in the target domain;

[0056] Use source domain encoder and target domain encoder Learn the source and target domains separately;

[0057] Source domain encoder Mapping data to the source domain space, target domain encoder Map the data to the target domain space, and use the source domain classifier in the source domain. Label data for source domain Perform classification to obtain predicted probability values; design a target domain classifier. and pseudo-label classifier Label the data in the target domain respectively and unlabeled data in the target domain Classify the data to obtain predicted probability values.

[0058] Preferably, sample-by-sample cross-alignment specifically involves:

[0059] Source Sample The features are obtained after passing through a feature extractor, source domain encoder, and target domain encoder. and ; and After classifier get and Determine the source domain loss function; for target domain samples After passing through the feature extractor and encoder, the result is... and ; After classifier get , After classifier and Later obtained and The loss function for the source domain is:

[0060]

[0061]

[0062] in, It is hard label loss. For soft label loss, It is mini-batch size. For source samples The tag, for After feature extractor, target domain encoder and classifier The output probability value after that, for After feature extractor, source domain encoder and classifier The output probability value;

[0063] The loss function for the target domain is:

[0064]

[0065]

[0066] in, For indicator functions, The confidence threshold is... For target domain samples The output value of the weakly enhanced version, for The corresponding pseudo-tags, for After feature extractor, source domain encoder and classifier The output probability value after that, The KL divergence algorithm is used. for After feature extractor, target domain encoder and classifier The output probability value.

[0067] Preferably, the cross-alignment by category specifically refers to:

[0068] For the source domain sample set , For the source domain samples and their corresponding labels; obtain and , For source domain encoder, For feature extractor, for The output features after passing through the feature extractor and source domain encoder for The output features after passing through the feature extractor and the target domain encoder For target domain encoder;

[0069] exist and By performing comparative learning between positive sample pairs, the distance between positive sample pairs is minimized while the distance between negative sample pairs is maximized, and the source domain loss function is determined.

[0070] For target domain samples Its characteristics are represented as and First, confirm the pseudo-label values, and then determine the source domain loss function;

[0071] The source domain loss function is as follows:

[0072]

[0073] in, , For temperature parameters, for The set of positive sample indicators;

[0074] target domain loss function as follows:

[0075]

[0076] in, , for The positive sample index set, Samples with a probability less than the confidence threshold are filtered out. The size of the batch. For indicator functions, for Positive samples, To and Different samples.

[0077] Preferably, consistency regularization is used to ensure that the outputs of the target domain classifier and the pseudo-label classifier are consistent, as follows:

[0078]

[0079] in, For target domain samples The output value of the weakly enhanced version, To correspond with the significantly enhanced output, For indicator functions, The confidence threshold is... for The corresponding pseudo-tag.

[0080] Preferably, in the dynamic threshold strategy, the final learning effect is applied to the scaling of the fixed threshold as follows:

[0081]

[0082]

[0083] in, The maximum value of the threshold. To be the minimum value, For indicator functions, The learning effect of class c after normalization at step t. For class c, consider the learning effect after the current learning state at step t. The size of the batch. For target domain samples The output value of the weakly enhanced version, As a reference to a category, The output dynamic threshold;

[0084] The learning effect of a class is defined as the ratio of the number of samples predicted to fall into that class with a probability greater than a threshold to the total number of samples predicted to fall into that class. Class in step size Learning effect at the place for:

[0085]

[0086] in, for Class in steps The threshold at that point For target domain samples The output value of the weakly enhanced version.

[0087] Thirdly, a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the semi-supervised domain adaptive wafer defect identification method described above.

[0088] Fourthly, embodiments of the present invention provide a computer-readable storage medium including a computer program, which, when executed by a processor, implements the steps of the above-described semi-supervised domain adaptive wafer defect identification method.

[0089] Fifthly, a chip includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the semi-supervised domain adaptive wafer defect identification method described above.

[0090] In a sixth aspect, embodiments of the present invention provide an electronic device including a computer program, which, when executed by the electronic device, implements the steps of the above-described semi-supervised adaptive wafer defect identification method.

[0091] Compared with the prior art, the present invention has at least the following beneficial effects:

[0092] A semi-supervised adaptive wafer defect identification method can effectively achieve knowledge transfer between different wafer defect datasets while keeping the framework basically unchanged, solve the label mismatch problem between datasets, and achieve high identification accuracy even when the target domain dataset has only a very small number of labels.

[0093] Furthermore, the dataset is normalized, and data augmentation is performed using random flipping and random pruning to improve data diversity.

[0094] 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, a debiased target domain classifier is trained using labeled target data, which can generate high-quality pseudo-labels for unlabeled target data, thereby improving overall performance.

[0095] Furthermore, a sample-by-sample cross-alignment is proposed to align the classifier output probability values ​​of each sample in two subspaces, thereby improving the output consistency of the same sample across different encoders.

[0096] Furthermore, a class-wise cross-alignment is proposed to bring the representations of the same class in different domains closer together, while different classes move further apart, thereby promoting feature clustering.

[0097] Furthermore, a consistency regularization method is used to generate pseudo-labels from weakly enhanced images to provide supervision signals for strongly enhanced images, with the aim of ensuring the consistency of the output of samples under different views.

[0098] Furthermore, to maximize the utility of unlabeled data, a dynamic thresholding strategy is proposed. This strategy comprehensively considers both the overall learning performance of each class and the current specific learning performance, accurately reflecting the learning status of each class. Compared with a fixed threshold, this dynamic strategy significantly improves learning efficiency and optimizes data utilization.

[0099] It is understood that the beneficial effects of the second to sixth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0100] In summary, this invention can effectively achieve knowledge transfer between different datasets and achieve high recognition accuracy even when the target domain dataset has only a very small number of labels, such as only one label per category.

[0101] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0102] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0103] Figure 1 This is a schematic diagram of a cross-aligned network;

[0104] Figure 2 This is a schematic diagram of the process of the present invention;

[0105] Figure 3 A schematic diagram of a computer device provided in an embodiment of the present invention;

[0106] Figure 4 This is a block diagram of a chip provided according to an embodiment of the present invention. Detailed Implementation

[0107] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0108] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0109] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0110] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this invention generally indicates that the preceding and following objects have an "or" relationship.

[0111] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. 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.

[0112] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0113] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0114] This invention provides a semi-supervised, domain-adaptive wafer defect identification method. It employs a cross-alignment network, using different encoders and classifiers to decouple the feature subspace and label space. To facilitate knowledge transfer between the source and target domains, this invention also proposes a cross-alignment learning method; this method consists of sample cross-alignment and class cross-alignment, aiming to ensure the consistency of encoder outputs and promote the formation of robust clusters. Furthermore, this method integrates consistency regularization to enforce output consistency across different views of samples, and introduces a dynamic thresholding strategy that tailors the threshold based on the learning performance of each class, thereby improving model efficiency and data utilization.

[0115] This invention focuses on domain-adaptive scenarios, effectively addressing the lack of data labels in real industrial production processes and exhibiting better generalization in application scenarios. It decouples the label space, thus avoiding the influence of source supervision on target classification and effectively resolving potential label mismatch issues. The invention utilizes cross-alignment learning to fully promote knowledge interaction and fusion between datasets, enhancing algorithm stability. Furthermore, the invention employs a dynamic threshold strategy to accurately reflect the learning status of each class, effectively improving model learning efficiency and optimizing data utilization.

[0116] Please see Figure 2 This invention discloses a semi-supervised adaptive wafer defect identification method, comprising the following steps:

[0117] S1. Obtain the wafer defect source and target domain datasets by independently collecting or downloading publicly available datasets online; subsequently, preprocess the dataset data, including random horizontal flipping, random cropping, and normalization.

[0118] The publicly available datasets WM-811K and MixedWM38 are used, both of which are widely used in wafer image defect identification. 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 working chip, and 2 represents a defective chip, displayed as dark purple, green, and yellow in the image, respectively. WM-811K contains 811,457 wafer defect images, of which 172,950 are labeled images used for experiments. MixedWM38 is a dataset for identifying mixed defect patterns, including 8 single-type defects and 29 mixed-type defects. Mixed defects are combinations of single-type defects. In this invention, to satisfy the domain adaptation condition, only single-type defects were used for experiments. Both datasets exhibit class imbalance, with WM-811K being more severe because the "none" class accounts for 85% of the dataset.

[0119] To standardize image sizes, this invention sets the size of all images in both datasets to 64x64.

[0120] Next, use the formula: Normalize the image to ensure that pixel values ​​fall within the range of [-1, 1].

[0121] Then, the data is preprocessed using methods such as random flipping and random cropping.

[0122] In consistent regularization, a powerful data augmentation technique is employed, namely random augmentation RandAugment.

[0123] S2. Construct a cross-alignment network and use a shared feature extractor and encoder to extract image features from the source and target domains;

[0124] Please see Figure 1 In cross-aligned networks, Represents source domain labeled data, Represents the target domain labeled data. This represents unlabeled data in the target domain. This invention uses a source domain encoder. and target domain encoder The source domain and target domain are learned separately; 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 classification is performed;

[0125] Source domain uses source domain classifier For the target domain, design a target domain classifier. and pseudo-label classifier ;

[0126] The target domain classifier is optimized using only 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 must also provide pseudo-labels for the pseudo-label classifier, and the pseudo-labels are the class labels with the highest probability greater than the confidence threshold;

[0127] During testing, only the target domain classifier will be used. This design maintains the separation of the label space, thus effectively solving the label mismatch problem.

[0128] S3. For the extracted image features, a classifier is used to obtain the output probability. For unlabeled data, a dynamic threshold strategy is used to determine and update the pseudo-label threshold for each category. Then, the loss function is calculated and trained using cross-alignment learning and consistency regularization methods.

[0129] We use sample-by-sample and class-by-class cross-alignment to ensure the consistency of the encoder outputs in the source and target domains, and use consistency regularization to ensure the consistency of the outputs of the target domain classifier and the pseudo-label classifier. We use a dynamic thresholding strategy to determine and update the pseudo-label threshold for each class. The specific steps are as follows:

[0130] S301, Sample-by-sample cross-alignment

[0131] To ensure consistency in the outputs of samples across the two encoders, a sample-by-sample cross-alignment method is first proposed. The main idea is to align the logical values ​​of each sample's classifier outputs in the two subspaces. To enhance knowledge interaction between the two domains, a combination of hard and soft labels is used for training. Hard labels are traditional classification labels, typically in the form of one-hot encoding. Unlike hard labels, soft labels have continuous values, containing the probability distribution of each class and thus more information. The combination of these two methods helps the two encoders transfer richer knowledge.

[0132] First, the source sample The features are obtained after passing through a feature extractor, source domain encoder, and target domain encoder. and ;

[0133] after, and After classifier get and ;

[0134] Next, determine the source domain loss function;

[0135] For target domain samples After passing through the feature extractor and encoder, the result is... and ;

[0136] Then, After classifier get fut passes through a classifier and Later obtained and .

[0137] The loss function for the source domain is:

[0138]

[0139]

[0140] in, It is hard label loss. This is the soft label loss. This is the mini-batch size. Kl divergence is used here to calculate the similarity of the output logic values.

[0141] The loss function for the target domain is:

[0142]

[0143]

[0144] in, For indicator functions, This is the confidence threshold.

[0145] S302, Cross-alignment by category

[0146] To ensure consistent and stable clustering of samples across different domains, class-wise cross-alignment is introduced. This approach brings together representations of the same class within the same domain across different domains, while distancing different classes. This promotes consistency in the outputs of the two encoders and enhances feature learning.

[0147] For source domain samples ;

[0148] get and ;

[0149] To align the two feature distributions, a proposal is made in and We perform comparative learning between positive and negative sample pairs, aiming to minimize the distance between positive sample pairs while maximizing the distance between negative sample pairs.

[0150] For target domain samples Its characteristics are represented as and ;because All of this data is unlabeled, so we need to confirm the pseudo-label values ​​first.

[0151] Determine the loss functions for the source and target domains;

[0152] The source domain loss function is as follows:

[0153]

[0154] in, , For temperature parameters, for The positive sample index set.

[0155] The target domain loss function is as follows:

[0156]

[0157] in, , for The positive sample index set, Samples with a probability less than the confidence threshold are filtered out.

[0158] S303. The purpose of consistency regularization is to ensure the consistency of output across different views. This method is generally applied to unlabeled data in the target domain and is implemented using pseudo-labels. Typically, weakly augmented images are used to generate pseudo-labels to provide supervision signals for strongly augmented images. The cross-entropy loss function is used for learning consistency regularization, as follows:

[0159]

[0160] in, It is a target domain sample The output value of the weakly enhanced version, This corresponds to the output of the heavily enhanced version.

[0161] S304, Dynamic Threshold Strategy

[0162] 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 a new threshold definition.

[0163] Ideally, different thresholds should be set for different categories based on learning performance. This application defines the learning performance of a class as the ratio of the number of samples predicted to fall into that class with a probability greater than the threshold to the total number of samples predicted to fall into that class, using the following formula:

[0164]

[0165] in, for Class in step size Learning outcomes at the location for Class in step size The threshold at that location. It only processes samples from a single mini-batch, thus representing the current learning performance for each class.

[0166] To comprehensively consider the learning outcomes of previous samples, the overall learning effect is estimated and normalized, as follows:

[0167]

[0168]

[0169] After normalization, the class with the best learning state The value is 1. However, the normalized value only considers the relative learning between classes, not the learning situation of the entire batch, which is not conducive to learning. The learning situation of a batch can be expressed as the ratio of the number of samples with the highest probability greater than a threshold within the batch to the batch size.

[0170] Taking all factors into account, the final learning result is obtained and applied to scaling with a fixed threshold:

[0171]

[0172]

[0173] This involves introducing two user-predefined thresholds. , , Represents the maximum threshold value. Represents the minimum value. It will scale according to the learning status. This is used to prevent the threshold from being too small.

[0174] S4. Using the trained cross-alignment network as the prediction model, retain the output of the target domain classifier as the final value predicted by the cross-alignment network, and output the wafer defect category of the unlabeled data in the target domain.

[0175] Those skilled in the art will understand that various aspects of the present invention can be implemented as systems, methods, or program products. Therefore, various aspects of the present invention can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software aspects, collectively referred to herein as a "circuit," "module," or "platform."

[0176] In another embodiment of the present invention, a semi-supervised domain adaptive wafer defect identification system is provided. This system can be used to implement the above-mentioned semi-supervised domain adaptive wafer defect identification method. Specifically, the semi-supervised domain adaptive wafer defect identification system includes a data module, an extraction module, an identification module, and an output module.

[0177] The data module acquires datasets of wafer defect source and target domains and preprocesses these datasets.

[0178] The publicly available datasets WM-811K and MixedWM38 are used, both of which are widely used in wafer image defect identification. 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 working chip, and 2 represents a defective chip, displayed as dark purple, green, and yellow in the image, respectively. WM-811K contains 811,457 wafer defect images, of which 172,950 are labeled images used for experiments. MixedWM38 is a dataset for identifying mixed defect patterns, including 8 single-type defects and 29 mixed-type defects. Mixed defects are combinations of single-type defects. In this invention, to satisfy the domain adaptation condition, only single-type defects were used for experiments. Both datasets exhibit class imbalance, with WM-811K being more severe because the "none" class accounts for 85% of the dataset.

[0179] To standardize image sizes, this invention sets the size of all images in both datasets to 64x64.

[0180] Next, use the formula: Normalize the image to ensure that pixel values ​​fall within the range of [-1, 1].

[0181] Then, the data is preprocessed using methods such as random flipping and random cropping.

[0182] In consistent regularization, a powerful data augmentation technique is employed, namely random augmentation RandAugment.

[0183] The extraction module constructs a cross-alignment network and uses a feature extractor and encoder to extract image features from the source and target domains of the preprocessed dataset, respectively.

[0184] In cross-aligned networks, Represents source domain labeled data, Represents the target domain labeled data. This represents unlabeled data in the target domain;

[0185] Use source domain encoder and target domain encoder Learn the source and target domains separately;

[0186] Source domain encoder Mapping data to the source domain space, target domain encoder Map the data to the target domain space, and use the source domain classifier in the source domain. Label data for source domain Perform classification to obtain predicted probability values; design a target domain classifier. and pseudo-label classifier Label the data in the target domain respectively and unlabeled data in the target domain Classify the data to obtain predicted probability values.

[0187] The recognition module uses a classifier to obtain the output probability for the extracted image features. For unlabeled data, it determines and updates the pseudo-label threshold for each category through a dynamic thresholding strategy. Then, it uses sample-by-sample cross-alignment, class-by-class cross-alignment, and consistency regularization methods to calculate the loss function and train the network to obtain the trained cross-alignment network.

[0188] The specific steps of sample-by-sample cross-alignment are:

[0189] Source Sample The features are obtained after passing through a feature extractor, source domain encoder, and target domain encoder. and ; and After classifier get and Determine the source domain loss function; for target domain samples After passing through the feature extractor and encoder, the result is... and ; After classifier get , After classifier and Later obtained and .

[0190] The loss function for the source domain is:

[0191]

[0192]

[0193] in, It is hard label loss. For soft label loss, It is mini-batch size. For source samples The tag, for After feature extractor, target domain encoder and classifier The output probability value after that, for After feature extractor, source domain encoder and classifier The output probability value;

[0194] The loss function for the target domain is:

[0195]

[0196]

[0197] in, For indicator functions, The confidence threshold is... For target domain samples The output value of the weakly enhanced version, for The corresponding pseudo-tags, for After feature extractor, source domain encoder and classifier The output probability value after that, The KL divergence algorithm is used. for After feature extractor, target domain encoder and classifier The output probability value.

[0198] Cross-alignment by category specifically refers to:

[0199] For the source domain sample set , For the source domain samples and their corresponding labels; obtain and , For source domain encoder, For feature extractor, for The output features after passing through the feature extractor and source domain encoder for The output features after passing through the feature extractor and the target domain encoder For target domain encoder;

[0200] exist and By performing comparative learning between positive sample pairs, the distance between positive sample pairs is minimized while the distance between negative sample pairs is maximized, and the source domain loss function is determined.

[0201] For target domain samples Its characteristics are represented as and First, confirm the pseudo-label values, and then determine the source domain loss function.

[0202] The source domain loss function is as follows:

[0203]

[0204] in, , For temperature parameters, for The set of positive sample indicators;

[0205] target domain loss function as follows:

[0206]

[0207] in, , for The positive sample index set, Samples with a probability less than the confidence threshold are filtered out. The size of the batch. For indicator functions, for Positive samples, To and Different samples.

[0208] Consistency regularization is used to ensure that the outputs of the target domain classifier and the pseudo-label classifier are consistent, as follows:

[0209]

[0210] in, For target domain samples The output value of the weakly enhanced version, To correspond with the significantly enhanced output, For indicator functions, The confidence threshold is... for The corresponding pseudo-tag.

[0211] In the dynamic thresholding strategy, the final learning effect is applied to scaling of a fixed threshold as follows:

[0212]

[0213]

[0214] in, The maximum value of the threshold. To be the minimum value, For indicator functions, The learning effect of class c after normalization at step t. For class c, consider the learning effect after the current learning state at step t. The size of the batch. For target domain samples The output value of the weakly enhanced version, As a reference to a category, The output dynamic threshold;

[0215] The learning effect of a class is defined as the ratio of the number of samples predicted to fall into that class with a probability greater than a threshold to the total number of samples predicted to fall into that class. Class in step size Learning effect at the place for:

[0216]

[0217] in, for Class in steps The threshold at that point For target domain samples The output value of the weakly enhanced version.

[0218] The output module uses a trained cross-alignment network as a prediction model to output the wafer defect category that identifies the unlabeled data in the target domain.

[0219] In another embodiment of the present invention, a terminal device is provided, comprising a processor and a memory. The memory stores a computer program, the computer program including program instructions, and the processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, graphics processing units (GPUs), tensor processing units (TPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used in the operation of a semi-supervised adaptive wafer defect identification method, including:

[0220] Acquire datasets of wafer defect source and target domains; preprocess the datasets; construct a cross-alignment network, and use a feature extractor and encoder to extract image features of the source and target domains of the preprocessed dataset, respectively; for the extracted image features, use a classifier to obtain the output probability; for unlabeled data, determine and update the pseudo-label threshold for each category through a dynamic thresholding strategy; then use sample-by-sample cross-alignment, class-by-class cross-alignment, and consistency regularization methods to calculate the loss function and train the network to obtain a trained cross-alignment network; use the trained cross-alignment network as a prediction model to output the wafer defect category of the unlabeled data in the target domain.

[0221] Please see Figure 3 The terminal device is a computer device. In this embodiment, the computer device 60 includes a processor 61, a memory 62, and a computer program 63 stored in the memory 62 and executable on the processor 61. When executed by the processor 61, the computer program 63 implements the semi-supervised domain adaptive wafer defect identification method of this embodiment. To avoid repetition, details are omitted here. Alternatively, when executed by the processor 61, the computer program 63 implements the functions of each model / unit in the semi-supervised domain adaptive wafer defect identification system of this embodiment. To avoid repetition, details are omitted here.

[0222] Computer device 60 can be a desktop computer, laptop, handheld computer, cloud server, or other computing device. Computer device 60 may include, but is not limited to, a processor 61 and a memory 62. Those skilled in the art will understand that... Figure 3 This is merely an example of computer device 60 and does not constitute a limitation on computer device 60. It may include more or fewer components than shown, or combine certain components, or different components. For example, computer device may also include input / output devices, network access devices, buses, etc.

[0223] The processor 61 may be a Central Processing Unit (CPU), or other general-purpose processors, graphics processing units (GPUs), tensor processing units (TPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0224] The memory 62 can be an internal storage unit of the computer device 60, such as a hard disk or RAM of the computer device 60. The memory 62 can also be an external storage device of the computer device 60, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on the computer device 60.

[0225] Furthermore, the memory 62 may include both internal storage units of the computer device 60 and external storage devices. The memory 62 is used to store computer programs and other programs and data required by the computer device. The memory 62 can also be used to temporarily store data that has been output or will be output.

[0226] Please see Figure 4 The terminal device is a chip. In this embodiment, the chip 600 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. Furthermore, the processor 622 may be configured to execute the computer program to perform the semi-supervised domain adaptive wafer defect identification method described above.

[0227] Additionally, chip 600 may also include a power supply component 626 and a communication component 650. The power supply component 626 can be configured to perform power management of chip 600, and the communication component 650 can be configured to enable communication of chip 600, such as wired or wireless communication. Furthermore, chip 600 may also include an input / output interface 658. Chip 600 can operate on an operating system stored in memory 632.

[0228] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium, which is a memory device in a terminal device for storing programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and extended storage media supported by the terminal device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor; these instructions can be one or more computer programs. It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device.

[0229] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the semi-supervised adaptive wafer defect identification method in the above embodiments; one or more instructions in the computer-readable storage medium are loaded and executed by the processor to perform the following steps:

[0230] Acquire datasets of wafer defect source and target domains; preprocess the datasets; construct a cross-alignment network, and use a feature extractor and encoder to extract image features of the source and target domains of the preprocessed dataset, respectively; for the extracted image features, use a classifier to obtain the output probability; for unlabeled data, determine and update the pseudo-label threshold for each category through a dynamic thresholding strategy; then use sample-by-sample cross-alignment, class-by-class cross-alignment, and consistency regularization methods to calculate the loss function and train the network to obtain a trained cross-alignment network; use the trained cross-alignment network as a prediction model to output the wafer defect category of the unlabeled data in the target domain.

[0231] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0232] The feature extraction backbone network used in this invention can be replaced by networks other than ResNet-34, such as VGG and AlexNet. The visual-based wafer defect information input in this invention can theoretically be replaced by other types of data, such as visual image information.

[0233] The core of the cross-aligned network structure lies in achieving feature alignment between domains, thereby reducing the distributional differences between the source and target domains. This involves using two sub-networks (one for the source domain and one for the target domain) and a joint feature representation. By optimizing the loss function during the process, the feature representations of the two domains are made as close as possible to each other in the same feature space. The cross-aligned network structure exhibits higher performance on wafer defect recognition tasks, mainly due to its efficient utilization of domain adaptive learning. By combining semi-supervised learning and domain adaptation techniques, the model can effectively utilize unlabeled data with limited labeled data, thereby improving the model's generalization ability and recognition accuracy.

[0234] The cross-aligned network structure used in this invention has been well validated on the WM-811K and MixedWM38 public datasets, and its performance is improved compared to the current mainstream models, proving that the method is feasible.

[0235] In summary, this invention presents a semi-supervised domain-adaptive wafer defect identification method and system. It proposes a cross-alignment network to achieve semi-supervised domain-adaptive learning between two datasets by decoupling the label space, achieving high accuracy even when the target dataset contains very few labels, such as only one or three labels per class. Furthermore, this invention proposes a cross-alignment learning method to effectively ensure the consistency of encoder output, thereby promoting knowledge transfer between the source and target domains. A dynamic threshold strategy is also introduced, tailoring the threshold according to the learning performance of each class, thus improving learning efficiency and data utilization.

[0236] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to 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 embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0237] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0238] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0239] In the embodiments provided by this 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 merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0240] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0241] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0242] If the integrated module / unit is implemented as 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, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random-access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed 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 electrical carrier signals and telecommunication signals.

[0243] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0244] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0245] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0246] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A semi-supervised adaptive wafer defect identification method, characterized in that, Includes the following steps: Obtain datasets of wafer defect source and target domains; and preprocess the datasets. A cross-alignment network is constructed, using a feature extractor and an encoder to extract image features from the source and target domains of the preprocessed dataset, respectively. In the cross-alignment network... Represents source domain labeled data, Represents the target domain labeled data. This represents unlabeled data in the target domain; Use source domain encoder and target domain encoder Learn the source and target domains separately; Source domain encoder Mapping data to the source domain space, target domain encoder Map the data to the target domain space, and use the source domain classifier in the source domain. Label data for source domain Perform classification to obtain predicted probability values; design a target domain classifier. and pseudo-label classifier Label the data in the target domain respectively and unlabeled data in the target domain Classify the data to obtain predicted probability values; For the extracted image features, a classifier is used to obtain the output probability. For unlabeled data, a dynamic thresholding strategy is used to determine and update the pseudo-label threshold for each class. Then, the loss function is calculated and trained using per-sample cross-alignment, per-class cross-alignment, and consistency regularization methods to obtain the trained cross-alignment network. The per-sample cross-alignment is specifically as follows: Source Sample The features are obtained after passing through a feature extractor, source domain encoder, and target domain encoder. and ; and After classifier get and Determine the source domain loss function; for target domain samples After passing through the feature extractor and encoder, the result is... and ; After classifier get , After classifier and Later obtained and ; Cross-alignment by category specifically refers to: For the source domain sample set , For the source domain samples and their corresponding labels; obtain and , For source domain encoder, For feature extractor, for The output features after passing through the feature extractor and source domain encoder for The output features after passing through the feature extractor and the target domain encoder For target domain encoder; exist and By performing comparative learning between positive sample pairs, the distance between positive sample pairs is minimized while the distance between negative sample pairs is maximized, and the source domain loss function is determined. For target domain samples Its characteristics are represented as and First, confirm the pseudo-label values, and then determine the source domain loss function; In the dynamic thresholding strategy, the final learning effect is applied to scaling of a fixed threshold as follows: in, The maximum value of the threshold. To be the minimum value, For indicator functions, The learning effect of class c after normalization at step t. For class c, consider the learning effect after the current learning state at step t. The size of the batch. For target domain samples The output value of the weakly enhanced version, As a reference to a category, The output dynamic threshold; The learning effect of a class is defined as the ratio of the number of samples predicted to fall into that class with a probability greater than a threshold to the total number of samples predicted to fall into that class. Class in step size Learning effect at the place for: in, for Class in steps The threshold at that point For target domain samples The output value of the weakly enhanced version; The trained cross-alignment network is used as a prediction model to output the wafer defect category that identifies unlabeled data in the target domain.

2. The semi-supervised adaptive wafer defect identification method according to claim 1, characterized in that, Data sets of wafer defect source and target domains were obtained from the publicly available datasets WM-811K and MixedWM38. The datasets were then preprocessed as follows: Set all images in the public datasets WM-811K and MixedWM38 to 64x64; use the formula: The image is normalized to ensure that the pixel values ​​fall within the range of [-1, 1]; then, random flipping and random cropping are used for preprocessing.

3. The semi-supervised adaptive wafer defect identification method according to claim 1, characterized in that, The loss function for the source domain is: in, It is hard label loss. For soft label loss, It is mini-batch size. For source samples The tag, for After feature extractor, target domain encoder and classifier The output probability value after that, for After feature extractor, source domain encoder and classifier The output probability value; The loss function for the target domain is: in, For indicator functions, The confidence threshold is... For target domain samples The output value of the weakly enhanced version, for The corresponding pseudo-tags, for After feature extractor, source domain encoder and classifier The output probability value after that, The KL divergence algorithm is used. for After feature extractor, target domain encoder and classifier The output probability value.

4. The semi-supervised adaptive wafer defect identification method according to claim 1, characterized in that, The source domain loss function is as follows: in, , For temperature parameters, for The set of positive sample indicators; target domain loss function as follows: in, , for The positive sample index set, Samples with a probability less than the confidence threshold are filtered out. The size of the batch. For indicator functions, for Positive samples, To and Different samples.

5. The semi-supervised adaptive wafer defect identification method according to claim 1, characterized in that, Consistency regularization is used to ensure that the outputs of the target domain classifier and the pseudo-label classifier are consistent, as follows: in, For target domain samples The output value of the weakly enhanced version, To correspond with the significantly enhanced output, For indicator functions, The confidence threshold is... for The corresponding pseudo-tag.

6. A semi-supervised adaptive wafer defect identification system, characterized in that, include: The data module acquires datasets of wafer defect source and target domains; The dataset is then preprocessed. The extraction module constructs a cross-alignment network, using a feature extractor and an encoder to extract image features from the source and target domains of the preprocessed dataset, respectively. In the cross-alignment network... Represents source domain labeled data, Represents the target domain labeled data. This represents unlabeled data in the target domain; Use source domain encoder and target domain encoder Learn the source and target domains separately; Source domain encoder Mapping data to the source domain space, target domain encoder Map the data to the target domain space, and use the source domain classifier in the source domain. Label data for source domain Perform classification to obtain predicted probability values; design a target domain classifier. and pseudo-label classifier Label the data in the target domain respectively and unlabeled data in the target domain Classify the data to obtain predicted probability values; The recognition module uses a classifier to obtain output probabilities for extracted image features. For unlabeled data, it determines and updates the pseudo-label threshold for each category using a dynamic thresholding strategy. Then, it calculates the loss function and trains the network using sample-by-sample cross-alignment, class-by-class cross-alignment, and consistency regularization methods to obtain a trained cross-alignment network. The sample-by-sample cross-alignment is specifically as follows: Source Sample The features are obtained after passing through a feature extractor, source domain encoder, and target domain encoder. and ; and After classifier get and Determine the source domain loss function; for target domain samples After passing through the feature extractor and encoder, the result is... and ; After classifier get , After classifier and Later obtained and ; Cross-alignment by category specifically refers to: For the source domain sample set , For the source domain samples and their corresponding labels; obtain and , For source domain encoder, For feature extractor, for The output features after passing through the feature extractor and source domain encoder for The output features after passing through the feature extractor and the target domain encoder For target domain encoder; exist and By performing comparative learning between positive sample pairs, the distance between positive sample pairs is minimized while the distance between negative sample pairs is maximized, and the source domain loss function is determined. For target domain samples Its characteristics are represented as and First, confirm the pseudo-label values, and then determine the source domain loss function; In the dynamic thresholding strategy, the final learning effect is applied to scaling of a fixed threshold as follows: in, The maximum value of the threshold. To be the minimum value, For indicator functions, The learning effect of class c after normalization at step t. For class c, consider the learning effect after the current learning state at step t. The size of the batch. For target domain samples The output value of the weakly enhanced version, As a reference to a category, The output dynamic threshold; The learning effect of a class is defined as the ratio of the number of samples predicted to fall into that class with a probability greater than a threshold to the total number of samples predicted to fall into that class. Class in step size Learning effect at the place for: in, for Class in steps The threshold at that point For target domain samples The output value of the weakly enhanced version; The output module uses a trained cross-alignment network as a prediction model to output the wafer defect category that identifies the unlabeled data in the target domain.

Images