Cross-domain detection method, equipment and medium for wafer surface defects of known and unknown categories

Abstract

The present invention provides a cross-domain detection method, device and medium for wafer surface defects of known unknown categories. Based on the traditional surface defect detection method, transfer learning is introduced to learn the domain-invariant features of source domain and target domain data for surface defect recognition and classification, which effectively solves the problem of data cross-domain caused by equipment aging in the actual production process. At the same time, a data pseudo-label learning module is constructed to identify and classify the target domain unlabeled data with the help of source domain data labels and target domain data domain-invariant features. In addition, for new types of unknown faults that may occur in the actual production process, an unknown category defect detection method is designed, and the transfer learning pseudo-label module is combined to make a comprehensive judgment on the fault type of the sample at the same time, so as to achieve high-precision detection and recognition of wafer surface defects.

Description

Technical Field

[0001] The present invention relates to the technical field of wafer surface defect detection, and in particular to a cross-domain detection method, device and storage medium for wafer surface defects of known and unknown categories. Background Art

[0002] In the semiconductor processing industry such as chips and panels, wafer processing is a crucial link. A large number of integrated circuits are etched on the surface of the wafer, which makes the processing process extremely complex and dynamic. At the same time, due to its high complexity, various faults are inevitable in the wafer processing process. Since such faults generally appear on the surface of the wafer, they are also called surface defects of the wafer. Depending on the type of fault in the processing process, the specific presentation of the surface defects is also different. By detecting the type of surface defects that occur in the actual processing process and identifying the correct type of fault, it can help locate the specific root cause of the fault, assist processing personnel in adjusting equipment parameters, replacing aging equipment, and improving production paths, thereby optimizing the processing process, improving raw material utilization, and improving product yield. Therefore, wafer surface defect detection is a crucial research topic for the semiconductor processing industry.

[0003] With the rapid development of the semiconductor industry, the surface defect detection of wafers is also facing increasing challenges. The tiny circuit structure and complex circuit composition make the detection of wafer surface defects particularly difficult. With the improvement of processing efficiency and output, it has become impractical to manually detect wafers under massive data. Major manufacturers have already moved the detection method closer to artificial intelligence, but artificial intelligence and deep learning methods for surface defect detection still face many problems. The first is the problem of data labeling. The data used for deep learning are all wafer images collected during the actual production process. These data are all unlabeled data, that is, the data does not carry its fault type. Deep learning fault detection is generally based on supervised learning, and the amount of data is too large, making manual labeling impractical. In addition, since the deep learning surface defect detection algorithm is trained based on given data, it can only identify the fault types that have already occurred. For new fault types, the algorithm cannot accurately identify them, but in the actual processing process, new faults are very likely to occur, which also makes the surface defect detection algorithm face new difficulties. In addition, due to the long-term characteristics of the factory processing process, various parts of the processing equipment will age and wear over time. Specifically, a high-yield production line will have a significant decrease in yield over time. For production equipment with different degrees of wear, the same processing parameter settings will have different yields. The detection problem in this scenario is generally classified as a cross-domain learning problem. The existence of cross-domain problems means that a set of detection algorithms cannot simultaneously reflect high accuracy in both, and the use of multiple detection models will make the algorithm too large, which is not conducive to improving detection efficiency.

[0004] At present, there are a large number of related studies on wafer surface defect detection at home and abroad, and its detection methods generally include the following categories:

[0005] 1) Surface defect detection method based on computer vision

[0006] With the rapid development of computer vision, more and more computer vision methods are being used in other fields, and they have achieved good results. Surface defect detection of wafers is essentially the detection of photos taken on the surface of wafers. Different surface defects will show different shapes and features in the wafer images. Based on this, a large number of studies have introduced computer vision target detection and instance segmentation methods into wafer surface defect detection, such as YOLO, MaskR-CNN, FasterR-CNN, etc. With the superiority of these algorithms, the detection accuracy of wafer surface defects has risen to a new level, but the detection dilemma mentioned above still exists, and the problem of unlabeled data and cross-domain still has a great impact on the detection algorithm.

[0007] 2) Detection methods based on data enhancement methods such as generative adversarial networks

[0008] In response to the data problem mentioned above, some studies have used various data enhancement methods to improve detection performance, including generative adversarial networks (GANs), rotation and replication data enhancement methods, etc. Since the training process of deep learning algorithms relies on the amount of data, generative adversarial networks are used to synthesize data, synthesizing large batches of labeled artificial data based on small batches of labeled data; the rotation and replication data enhancement method is used to extract known faults and synthesize them with fault-free samples to obtain new samples. This method solves the data problem to a certain extent, but it still cannot identify faults of unknown categories.

[0009] In general, with the rapid development of the semiconductor industry, the traditional wafer image surface defect detection method can no longer meet the product yield requirements in the current environment. Although the surface defect detection methods proposed in recent years have made significant breakthroughs with the help of computer vision algorithms, there are still many problems. Since real-time data does not have fault labels, supervised deep learning algorithms cannot make good use of data, and manual labeling has become impractical. Secondly, for new types of faults that appear in the production process, the existing algorithms cannot make accurate judgments, which will have a certain impact on the processing of subsequent products. In addition, as the processing time goes by, the equipment will age to varying degrees, and the cross-domain problems it brings will cause the algorithm to have different accuracies on different devices, which greatly affects the optimization of the product processing process and the improvement of the yield. Summary of the invention

[0010] The present invention proposes a cross-domain detection method, device and storage medium for wafer surface defects of known and unknown categories. To address the above three problems, the method introduces cross-domain transfer learning to solve data labeling problems and data cross-domain problems caused by equipment aging, and at the same time builds an unknown defect detection model to identify surface defects of unknown types, effectively solving the above three detection problems and improving the accuracy and efficiency of wafer surface defect detection.

[0011] To achieve the above object, the present invention adopts the following technical solutions:

[0012] A cross-domain detection method for wafer surface defects of known and unknown categories includes the following processes:

[0013] S1, using the generative network to extract source domain wafer image features for subsequent detection;

[0014] S2, using the known characteristics of the source domain wafer image labels, the source domain features and labels extracted in S1 are used to train the classification network model, so that it can accurately identify the fault category of the wafer image according to the source domain features;

[0015] S3, transplanting the generation network and classification network used for source domain data processing to the target domain to extract the features of the target domain wafer image and generate pseudo labels;

[0016] S4. The obtained source domain features, source domain labels, target domain features, and target domain pseudo labels are sent as input to the discriminant module. First, the discriminant network will filter the data to eliminate the negative transfer effect. Then, the discriminant network will judge the features of the source domain and the target domain to identify whether they come from the source domain or the target domain, so as to extract domain-invariant features.

[0017] S5, finally discriminate the network output result, perform back propagation based on the obtained recognition error, and update the network parameters of the entire model;

[0018] S6. For the recognition of unknown categories, the decoding network is first trained based on the data features obtained by the generating network. The decoding network is used to restore the features to data under the guidance of the data label category;

[0019] S7. During the detection, the data features are restored according to the results of the classification network and the output of the generation network, and the restored error is judged based on the error threshold to identify whether it is a known category fault or an unknown category fault.

[0020] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0021] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0022] It can be seen from the above technical solution that the present invention introduces a cross-domain transfer learning module to address the defects of traditional wafer surface defect detection methods, and proposes a cross-domain detection method for wafer surface defects of known unknown categories. On the basis of traditional surface defect detection methods, transfer learning is introduced to learn the domain-invariant features of source domain and target domain data for surface defect recognition and classification, which effectively solves the problem of cross-domain data caused by equipment aging in the actual production process. At the same time, a data pseudo-label learning module is constructed to identify and classify the target domain unlabeled data with the help of source domain data labels and target domain data domain-invariant features. In addition, a method for detecting unknown category defects is designed for new types of unknown faults that may occur in the actual production process, and the transfer learning pseudo-label module is combined to make a comprehensive judgment on the fault type of the sample at the same time, so as to achieve high-precision detection and identification of wafer surface defects.

[0023] Specifically, the advantages of the present invention are as follows:

[0024] The present invention proposes a cross-domain detection method for wafer surface defects of known and unknown categories. Compared with the existing detection means that are completely based on computer vision algorithms, the present invention adopts a surface defect detection method based on domain-invariant features, introduces transfer learning, and realizes cross-domain detection of data while ensuring the wafer defect recognition effect; at the same time, with the help of transfer learning, a pseudo-label generation method for unlabeled target domain data is constructed. Compared with the existing data enhancement means using generative adversarial networks, the method adopted by the present invention has small distortion and higher efficiency, and the pseudo-label generation part is embedded in the process of surface defect detection, and a more streamlined surface defect recognition process is established for unlabeled target domain data; in addition, the present invention designs surface defect detection of known and unknown categories to ensure that surface defect types that have never appeared in the data set will not be recognized as existing defect categories by the algorithm, thereby reducing the misjudgment rate of the algorithm under special circumstances, and the overall algorithm is simple and efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flowchart of the present invention;

[0026] Figure 2 is a schematic diagram of a feature reconstruction module according to an embodiment of the present invention;

[0027] Figure 3 It is a schematic diagram of network construction according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] 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 only part of the embodiments of the present invention, not all of the embodiments.

[0029] In view of the defects of the traditional wafer surface defect detection method, the embodiment of the present invention introduces a cross-domain transfer learning module and proposes a cross-domain detection method for wafer surface defects of known unknown categories. On the basis of the traditional surface defect detection method, transfer learning is introduced to learn the domain-invariant features of the source domain and target domain data for surface defect recognition and classification, which effectively solves the problem of cross-domain data caused by equipment aging in the actual production process. At the same time, a data pseudo-label learning module is constructed to identify and classify the target domain unlabeled data with the help of the source domain data labels and the domain-invariant features of the target domain data. In addition, in view of the new types of unknown faults that may appear in the actual production process, an unknown category defect detection method is designed, and the transfer learning pseudo-label module is combined to make a comprehensive judgment on the fault type of the sample at the same time, so as to achieve high-precision detection and identification of wafer surface defects.

[0030] Specifically, a cross-domain detection method for wafer surface defects of known and unknown categories according to an embodiment of the present invention comprises the following steps:

[0031] S1, using the generative network to extract source domain wafer image features for subsequent detection;

[0032] S2, using the known characteristics of the source domain wafer image labels, the source domain features and labels extracted in S1 are used to train the classification network model, so that it can accurately identify the fault category of the wafer image according to the source domain features;

[0033] S3, transplanting the generation network and classification network used for source domain data processing to the target domain to extract the features of the target domain wafer image and generate pseudo labels;

[0034] S4, the previously obtained source domain features, source domain labels, target domain features, and target domain pseudo labels are sent as input to the discriminant module. First, the D0 discriminant network will filter the data to eliminate the negative transfer effect; then the discriminant network will judge the features of the source domain and the target domain, and identify whether they come from the source domain or the target domain, so as to extract domain invariant features;

[0035] S5, finally discriminate the network output result, perform back propagation based on the obtained recognition error, and update the network parameters of the entire model;

[0036] S6. For the recognition of unknown categories, the decoding network is first trained based on the data features obtained by the generating network. The decoding network is used to restore the features to data under the guidance of the data label category;

[0037] S7. During the detection, the data features are restored according to the results of the classification network and the output of the generation network, and the restored error is judged based on the error threshold to identify whether it is a known category fault or an unknown category fault.

[0038] Among them, the cross-domain transfer learning module is as follows Figure 1 As shown in the figure, for the labeled source domain data, the wafer image data is extracted through the generative network G, and the extracted features z s It is used to train the classification network C. At the same time, the unlabeled target domain data is also processed above to obtain the feature z t , and the pseudo-label is obtained through the pseudo-label network. The discriminant network is used to ensure that the feature z extracted by the generative network belongs to the domain invariant feature. Figure 2 The figure shows a surface defect detection method with known and unknown categories. The feature z of the target domain data is output by the generated network G. t After that, The purpose of training the decoding network as input is to use the feature z t Restore x t , and according to the reconstructed With x t The error is used to determine the label Whether it is a true label, the final output is the predicted surface defect type.

[0039] The following are specific instructions:

[0040] 1) Generate network feature extraction

[0041] The input of this step is the source domain data x, and the output is the feature z corresponding to the source domain data.

[0042] The generating network G is used to extract the features of the data for subsequent detection. Because of the differences in the feature space of the source domain and the target domain data, the same model often cannot be used to detect both at the same time. The key part of transfer learning is to extract the domain-invariant features of the source domain and the target domain. After obtaining the domain-invariant features, a model can be used to simultaneously identify and process the data of the source domain and the target domain. The generating network G can use a relatively simple CNN feature extraction network. A typical CNN model consists of an input layer, a convolution layer, a pooling layer, and a fully connected layer. CNN extracts features through convolution layers and reduces the number of network parameters by sharing convolution weights and pooling operations. The update of its parameters will use the subsequent discriminant network to extract the domain-invariant features of the source domain and the target domain data. Because the essence of transfer learning is to find domain-invariant features, compared with previous methods, the present invention starts with features to find breakthroughs, rather than being limited to data and labels.

[0043] 2) Source domain data classification model training

[0044] This step follows the previous generative network, uses the output z of the generative network as input, and outputs the label prediction value y for z.

[0045] Due to the heterogeneity of data features in the source domain and the target domain, the original data cannot be directly used to train the classification network. The present invention trains the classification network based on the domain-invariant features extracted by the generative network, such as Figure 1 The purpose of the present invention is to obtain a classification network that performs well in the target domain. Therefore, a classification network for the source domain is first trained in a supervised manner using source domain data and labels on the source domain. The source domain data features z extracted in the previous step are s As input, the classification network outputs the corresponding predicted label:

[0046] f c (x):R d →R c

[0047] Its loss function is as follows:

[0048]

[0049] Among them, x s and s Represents source domain data and labels respectively, n s represents the number of samples, and L(·,·) represents the cross entropy loss function, as shown below:

[0050]

[0051] in, Represents the indicator function.

[0052] In the subsequent processing of target domain data, the classification network is used to extract pseudo labels of the target domain data to facilitate the domain invariant feature determination of the subsequent discriminant network.

[0053] 3) Domain-invariant feature extraction based on domain discrimination

[0054] The input of this step is the feature z extracted by the generation network and the label y given by the classification network, which is used to determine whether each input belongs to the source domain or the target domain.

[0055] In general, generative adversarial networks (GANs) have difficulty converging during training. This is because GANs do not have an effective metric to measure the generated images, and it is difficult to ensure that the generator is well synchronized with the discriminator during training, which causes the model parameters to oscillate, become unstable, and never converge. Compared to GANs, deep transfer learning models based on adversarial training produce domain-invariant features and generally exhibit better stability.

[0056] As mentioned above, the function of the generative network is to extract domain-invariant features of the source domain and the target domain. Therefore, in the processing of target domain data, the first step is to extract features with the help of the generative network G. In the source domain, since the data has a label y s Therefore, the generated network processes the input x s Get feature z s After that, we can s ,y s ) is input into the discriminant network. However, the target domain data has no labels, and the learning of the discriminant network requires that the labels of the data are known, so that the loss function of the discriminant network can be calculated and the network parameters can be updated. Therefore, a pseudo-label module is needed to give a temporary label to the target domain data. Existing deep transfer networks based on adversarial training make full use of the advantages of GAN and improve the invariance of features in different domains. However, these existing algorithms do not pay enough attention to the class discrimination of domain-invariant features. Therefore, the present invention proposes a pseudo-label learning block to enhance the class discrimination of features in the target domain and improve the recognition performance of the model. In transfer learning based on domain adaptation, there is no class label information in the target domain. Therefore, in semi-supervised learning, it is difficult to determine the source domain data P s (x) and target domain data P t (x). In order to ensure that the generated features have domain invariance and class discrimination, a pseudo-label learning block is embedded, and the pseudo-labels of the target domain data are generated based on the domain classifier. In this invention, the source domain classification network C is used to generate pseudo-labels for the target domain data. After obtaining the target domain pseudo-labels, the target domain Enter the discriminant network.

[0057] The discriminant network determines the source of the input feature z and identifies whether it comes from the source domain or the target domain. For the generative network G, if it does not have a good domain-invariant feature extraction function, the discriminant network can successfully identify whether z comes from the source domain or the target domain. On the contrary, if it has a good domain-invariant feature extraction function, the discriminant network will find it difficult to determine the source of the input. At this time, it is considered that the generative network already has the ability to extract domain-invariant features.

[0058] exist Figure 1 In the above, the discriminant network is divided into two parts, D and D0. The above mentioned is the function of the discriminant network D, while the discriminant network D0 is established to solve the negative transfer problem. In transfer learning, when the source domain label space is larger than the target domain label space, negative transfer will occur, which will reduce the transfer learning performance and is not conducive to the convergence of the model. The phenomenon of negative transfer is specifically manifested in the presence of a part of special data in the source domain. This part of the data does not have the characteristics of the target domain at all, that is, no matter what method is used, the domain invariant features between this part of the data and the target domain data cannot be extracted. The existence of such data will greatly interfere with the training of the discriminant network. The present invention introduces the discriminant network D0 here, and its function is to screen the input features before the discriminant network D makes a judgment, and eliminate those features z that are too different from the features of the target domain, that is, the z that is too different may come from those label samples in the source domain that do not exist in the target domain, which solves the impact of negative transfer on model performance to a certain extent.

[0059] In the present invention, D0 is a binary classification network used to distinguish whether the input belongs to the negative transfer case. In the present invention, the D0 output layer maps the weight ω to 0 or 1, indicating whether it is a negative transfer sample. The specific method of weight calculation is:

[0060]

[0061] in, is the optimal value of the network, and its calculation formula is:

[0062]

[0063] It is not difficult to see from the above formula that The larger the value is, the greater the probability that the sample is a negative transfer sample, and vice versa. Therefore, calculating the negative transfer probability of the sample through the network can effectively remove the negative transfer samples from the data to avoid affecting the overall performance of the model.

[0064] Due to the existence of D0, the original single model is divided into two due to different situations, namely the general detection model when the source domain label space and the target domain label space are the same and the negative transfer detection model in the case of negative transfer.

[0065] 4) Defect detection of known and unknown categories based on feature reconstruction error

[0066] This step is to verify the results of the previous test again to determine whether the previous test is reasonable.

[0067] For unknown fault detection, the present invention introduces a feature reconstruction module, such as Figure 2 As shown. According to the feature z obtained by the generator network, the decoding network G' is trained. Given a category, the ideal encoder-decoder network can perform a reversible encoding and decoding operation on the data, that is, the decoding result after the data is encoded can restore the original data. If it cannot be restored, it indicates that the given category is wrong and the wrong encoder-decoder combination is used. Borrowing this point, the decoder network is used to restore the feature z output by the generator network. By reconstructing the restored The error is compared with the real x, and a judgment method based on the error threshold is used to identify whether the reconstruction of x is accurate. When the predicted label of x given by the classifier is correct, the decoding model corresponding to its category can restore z to x within a small error range. If the reconstruction error is greater than the threshold, it is considered that the classification label of x is Error, that is, x belongs to an unknown defect type. With the help of this method, the known and unknown faults of the wafer can be judged.

[0068] Collecting data in the semiconductor processing production line. Since this method requires two types of data, source domain and target domain, it is necessary to continuously collect a small amount of wafer image data first, and this part of the data needs to be labeled manually or in other ways. After that, the target domain data can be collected directly on the production line, and no fault labels are required. After obtaining the two data sets, use Figure 3The network is built in the manner shown, and the data set is input for training. The finally trained network is deployed on the production line for online detection. The network model that already has cross-domain detection can effectively identify the surface defect types of the wafer.

[0069] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0070] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0071] In another embodiment provided in the present application, a computer program product comprising instructions is also provided, which, when executed on a computer, enables the computer to execute any cross-domain detection method for wafer surface defects of known and unknown categories in the above-mentioned embodiments.

[0072] It is understandable that the system provided by the embodiment of the present invention corresponds to the method provided by the embodiment of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the above method.

[0073] The embodiment of the present application also provides an electronic device, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus.

[0074] Memory, used to store computer programs;

[0075] The processor is used to implement the above-mentioned cross-domain detection method for surface defects of wafers of known and unknown categories when executing the program stored in the memory.

[0076] The communication bus mentioned in the above electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc.

[0077] The communication interface is used for communication between the above electronic device and other devices.

[0078] The memory may include a random access memory (RAM) or a non-volatile memory (NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0079] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, and discrete hardware components.

[0080] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk SolidStateDisk (SSD)), etc.

[0081] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0082] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0083] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cross-domain detection method for wafer surface defects of known and unknown categories, characterized in that: The following processes are included: S1, using the generative network to extract source domain wafer image features for subsequent detection; S2, using the known characteristics of the source domain wafer image labels, the source domain features and labels extracted in S1 are used to train the classification network model, so that it can accurately identify the fault category of the wafer image according to the source domain features; S3, transplanting the generation network and classification network used for source domain data processing to the target domain to extract the features of the target domain wafer image and generate pseudo labels; S4. The obtained source domain features, source domain labels, target domain features, and target domain pseudo labels are sent as input to the discriminant module. First, the discriminant network will filter the data to eliminate the negative transfer effect. Then, the discriminant network will judge the features of the source domain and the target domain to identify whether they come from the source domain or the target domain, so as to extract domain-invariant features. S5, finally discriminate the network output result, perform back propagation based on the obtained recognition error, and update the network parameters of the entire model; S6. For the recognition of unknown categories, the decoding network is first trained based on the data features obtained by the generating network. The decoding network is used to restore the features to data under the guidance of the data label category; S7. During the detection, the data features are restored according to the results of the classification network and the output of the generation network, and the restored error is judged based on the error threshold to identify whether it is a known category fault or an unknown category fault; The step S7 includes introducing a feature reconstruction module for unknown fault detection, and generating a feature reconstruction module based on the feature , train the decoding network , given a category, the ideal codec network can perform a reversible encoding and decoding operation on the data, that is, decode the result after data encoding to restore the original data. If it cannot be restored, it means that the previously given category is wrong and the wrong codec combination is used; Borrowing this point, the decoding network is used to restore the features of the generated network output , obtained by reconstructing With the real Compare the errors and use the error threshold-based discrimination method to identify Is the reconstruction accurate? Given the classifier When the predicted label of is correct, the decoding model corresponding to its category can convert Restore to ; If the reconstruction error is greater than the threshold, it is considered Category tags Error, that is The defect type is unknown, so the known and unknown faults of the wafer can be judged.

2. The cross-domain detection method for wafer surface defects of known and unknown categories according to claim 1 is characterized in that: S1, using the generative network to extract source domain wafer image features for subsequent detection, specifically includes: Generative Network A CNN feature extraction network is used. A typical CNN model consists of an input layer, a convolutional layer, a pooling layer, and a fully connected layer. CNN extracts features through the convolutional layer and reduces the number of network parameters by sharing convolution weights and pooling operations. The update of its parameters will use the subsequent discriminant network to extract the invariant features of the source domain and target domain data.

3. The cross-domain detection method for wafer surface defects of known and unknown categories according to claim 1 is characterized in that: In S2, the source domain features and labels extracted from S1 are used to train the classification network model, including: First, a classification network for the source domain is trained in a supervised manner using the source domain data and labels. The source domain data features extracted in the previous step are As input, the classification network outputs the corresponding predicted label: Its loss function is as follows: in, and Represent the source domain data and labels respectively, represents the number of samples, Represents the cross entropy loss function as follows: in, represents the indicator function; In the subsequent processing of target domain data, the classification network is used to extract pseudo labels of the target domain data to facilitate the domain invariant feature determination of the subsequent discriminant network.

4. The cross-domain detection method for wafer surface defects of known and unknown categories according to claim 1 is characterized in that: The step S4 includes, in the processing of the target domain data, the first step is to generate a network To extract features, in the source domain, since the data is labeled Therefore, the generated network processes the input Get Features After that, Input into the discriminant network; In order to ensure that the generated features are domain invariant and class discriminative, a pseudo-label learning block is embedded, and the pseudo-labels of the target domain data are generated based on the domain classifier; Using source domain classification network To generate pseudo labels for target domain data, after obtaining the pseudo labels for the target domain, the target domain will Input discriminant network; The discriminant network is based on the input features , determine its source and identify whether it comes from the source domain or the target domain; for generating the network , if it does not have good domain-invariant feature extraction, the discriminant network can successfully identify Is it from the source domain or the target domain? Conversely, if it has a good domain-invariant feature extraction function, it will be difficult for the discriminant network to discriminate the source of the input. At this time, it is believed that the generative network already has the domain-invariant feature extraction capability.

5. The cross-domain detection method for wafer surface defects of known and unknown categories according to claim 4 is characterized in that: The discriminant network is divided into the discriminant network and Two parts, discriminant network The role is to discriminate the network Before making a judgment, the input features are screened to remove those features that are too different from the target domain features. ; Discriminative network It is a two-class network used to distinguish whether the input belongs to negative transfer or not. The output layer will weight Mapped to 0 or 1, indicating whether it is a negative transfer sample. The specific method of weight calculation is: in, is the optimal value of the network, and its calculation formula is: but The larger the value is, the greater the probability that the sample is a negative transfer sample, and vice versa. Therefore, calculating the negative transfer probability of the sample through the network can effectively remove the negative transfer samples from the data to avoid affecting the overall performance of the model.

6. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 5.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 5.

Images