A semiconductor defect localization memory holding increment generalization method, system and device

Abstract

The application relates to a semiconductor defect positioning memory holding incremental generalization method, system and equipment, which comprises the following steps: S1, acquiring an incremental data field of semiconductor defect positioning detection, a defect positioning detection model and a historical characterization agent of a current stage; S2, updating the historical characterization agent by using a current stage characterization prototype; S3, constructing mutual information constraint loss and defect positioning loss, combining the defect positioning loss and the mutual information constraint loss to obtain a combined loss; and S4, training the defect positioning detection model by using the incremental data field of the current stage and the combined loss to obtain a defect positioning detection model of a next stage. In the method, historical domain statistical information retained in past stage models or characterizations is fully utilized without replaying historical original images or accessing future data in advance, cross-domain knowledge reference is provided for current stage model training, and the model is guided to learn more stable and migratable defect characterizations.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, and in particular relates to a semiconductor defect location memory retention incremental generalization method, system and device. Background Technology

[0002] In microscopic optical inspection scenarios, common defects include scratches, cracks, chipped edges, dirt, foreign objects, pinholes, indentations, bubbles, welding anomalies, and missing local structures. These defects are typically characterized by small size, significant morphological variations, unclear boundaries, and complex background textures. Relying on manual inspection is not only inefficient but also easily affected by the experience and fatigue of the inspectors. Therefore, deep learning-based microscopic image defect localization and detection methods are gradually becoming an important technical route for industrial intelligent inspection.

[0003] In the microscopic inspection of semiconductor workpieces, microscopic images are typically acquired using different microscopic equipment, industrial cameras, light source configurations, magnification parameters, exposure conditions, inspection stations, and product batches. Even if the defect category is the same, images from different sources may still exhibit significant differences in brightness, contrast, texture detail, noise level, local reflections, and background structure, leading to a distribution shift between training data and actual detection data. If the defect localization and detection model is trained only on a fixed imaging condition or a limited batch of data, problems such as missed defects, false detections, localization shifts, or decreased confidence may occur when applied to new workpiece batches, new inspection stations, or new microscopic imaging conditions.

[0004] Existing deep learning-based defect localization and detection methods typically rely on relatively sufficient and stable training data. For general object detection models, defect segmentation models, or anomaly localization models, their detection performance is often affected by the distribution of the training data. As new batches of workpieces, new defect morphologies, or new imaging styles continuously appear during subsequent inspections, the model usually needs continuous updates. However, in real-world industrial scenarios, data is often not acquired all at once, but rather arrives in stages as production processes, inspection tasks, or customer needs change. Each stage's model can typically only access the data for that stage; original microscopic images from historical stages may not be retained or reused long-term due to storage costs, data access restrictions, or quality traceability requirements, and data from future stages cannot be obtained in advance. In this training scenario with continuously arriving data, the model faces two interrelated problems. One is the problem of historical knowledge forgetting, where the model's parameters gradually shift towards the current data distribution as it learns new batches or workstation data, leading to a decrease in its ability to detect historical defect categories, historical imaging styles, or historical workpiece batches. Secondly, there is the problem of insufficient generalization in the future domain. That is, the current data usually only covers limited imaging conditions and defect morphology. If the model is over-adapted to the current data distribution, it will be difficult to maintain stable detection performance under unknown batches, unknown inspection stations, or unknown microscopic imaging conditions in the future.

[0005] Traditional incremental learning methods primarily focus on mitigating historical knowledge forgetting, with common techniques including historical sample replay, knowledge distillation, and parameter regularization. Among these, methods based on historical sample replay require preserving or re-accessing data from past stages. However, in the context of microscopic semiconductor workpiece inspection, historical original images may not be long-term viable or suitable for repeated use in subsequent stages. While methods based on knowledge distillation or parameter regularization can mitigate forgetting to some extent, they typically prioritize the preservation of existing historical data, neglecting generalization capabilities for future, unknown imaging domains. On the other hand, domain generalization methods attempt to train models robust to unknown target domains using data from multiple source domains. However, traditional domain generalization methods often assume that multiple source domains can be accessed simultaneously or participate in parallel training, which is not entirely consistent with the reality of data arriving sequentially in stages during semiconductor workpiece microscopic inspection. When the model can only access a single current data domain at each training stage, traditional multi-source parallel training-based domain generalization methods are difficult to apply directly. Therefore, in the task of defect localization and detection of microscopic semiconductor workpieces, it is necessary to propose an incremental generalization method that can simultaneously consider both historical knowledge preservation and future domain generalization capabilities. Summary of the Invention

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] This invention provides a semiconductor defect location memory retention incremental generalization method, comprising the following steps:

[0008] Step S1: Obtain the incremental data domain, defect location detection model, and historical representation agent for the current stage of semiconductor defect location detection. The defect location detection model includes a feature extraction module and a defect location module. The historical representation agent is used to save and aggregate the defect representation statistics learned in the past incremental stages.

[0009] Step S2: Train the defect localization and detection model through the incremental data domain of the current stage, then input the incremental data domain of the current stage into the trained feature extraction module to obtain the current stage representation, obtain the current stage representation prototype through the current stage representation, and update the historical representation agent using the current stage representation prototype.

[0010] Step S3: Construct the mutual information constraint loss between the current stage representation and the historical representation agent, and construct the defect localization loss of the defect localization module. Combine the defect localization loss and the mutual information constraint loss to obtain the joint loss.

[0011] Step S4: Train the defect localization and detection model using the incremental data domain and joint loss of the current stage to obtain the defect localization and detection model for the next stage. Repeat the above steps until all incremental stages are trained, and finally obtain the memory-preserving incremental generalization model.

[0012] Preferably, in step S3, the mutual information constraint loss between the current stage representation and the historical representation proxy includes:

[0013] By using a multilayer perceptron to predict the representation of the current stage, a historical representation proxy is obtained.

[0014]

[0015] in, Representing historical representation agency, This indicates the current stage of characteristics. Indicates the representation based on the current stage. Inferred historical representation proxy The conditional probability distribution, and These represent the characteristics of the current stage. The predicted mean and covariance;

[0016] Constructing mutual information constraint loss:

[0017]

[0018] in, Indicates the current incremental data field. This indicates that for all samples in the current incremental data domain Find the average value. Represents image data, Indicates a label, Representing historical characteristics Mean of MLP prediction The square of the Euclidean distance between them Represents the covariance matrix of MLP predictions The logarithm of the determinant.

[0019] Preferably, in step S3, the formulas for the joint defect localization loss and the mutual information constraint loss are as follows:

[0020]

[0021] in, Indicates joint loss, Indicates the loss in defect location. These are the weighting coefficients.

[0022] Preferably, the defect localization loss adopts classification loss, bounding box regression loss, mask segmentation loss, anomaly localization loss, or a combination of the above losses.

[0023] Preferably, in step S4, training the defect location and detection model using the incremental data domain of the current stage to obtain the defect location and detection model for the next stage includes:

[0024] Multiple checkpoints are saved within the convergence interval of model training;

[0025] A memory-retaining ensemble model is constructed by combining the model parameters corresponding to multiple checkpoints, which serves as the defect localization and detection model for the next stage.

[0026] Preferably, the formula for constructing the memory retention integrated model is:

[0027]

[0028] in, This indicates that the memory retains the parameter states of the ensemble model, and n represents the number of checkpoints. This indicates the parameter status corresponding to the i-th checkpoint. This represents the weight of the i-th checkpoint.

[0029] Preferably, the weight of the i-th checkpoint is:

[0030]

[0031] in, This represents the mutual information constraint loss corresponding to the i-th checkpoint. This represents the temperature coefficient.

[0032] Preferably, the prototype of the current stage representation is obtained by averaging the current stage representation.

[0033] The present invention also provides a semiconductor defect location memory retention incremental generalization system, comprising:

[0034] The incremental data acquisition module is used to acquire the incremental data domain, defect location detection model, and historical representation agent for the current stage of semiconductor defect location detection. The defect location detection model includes a feature extraction module and a defect location module. The historical representation agent is used to save and aggregate the defect representation statistics learned in the past incremental stages.

[0035] The historical representation update module is used to input the incremental data domain of the current stage into the feature extraction module of the defect location detection model to obtain the current stage representation, and use the current stage representation to update the historical representation agent.

[0036] The mutual information constraint module is used to construct the mutual information constraint loss between the current stage representation and the historical representation agent, as well as to construct the defect localization loss of the defect localization module, and to combine the defect localization loss with the mutual information constraint loss.

[0037] The dynamic model integration module trains the defect location and detection model using the incremental data domain of the current stage to obtain the defect location and detection model for the next stage.

[0038] The present invention also provides an electronic device, including a memory and a processor, characterized in that the memory stores a computer program, and the processor is configured to run the computer program to perform the above-described semiconductor defect location memory retention incremental generalization method.

[0039] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the above-described semiconductor defect location memory retention incremental generalization method.

[0040] The present invention has at least the following beneficial effects:

[0041] This invention can use any defect location and detection model as the base network. During the training process, which arrives in stages from different incremental data domains, it maintains historical representation agents to aggregate defect representation information from past stages and promotes consistency between the current stage representation and the historical representation agents through mutual information constraints.

[0042] This invention integrates dynamic models based on the degree of matching between different training checkpoints and historical knowledge constraints, thereby improving the stable detection capability of the defect localization and detection model for microscopic images of historical batches, current batches, and unknown future batches without limiting the specific detection network structure. Attached Figure Description

[0043] Figure 1 A flowchart of the semiconductor defect location memory retention incremental generalization method provided in an embodiment of the present invention;

[0044] Figure 2 A flowchart illustrating the characterization process of the incremental generalization method for semiconductor defect location memory retention provided in an embodiment of the present invention; Detailed Implementation

[0045] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0046] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0047] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0048] like Figure 1-2 As shown, an embodiment of the present invention provides a semiconductor defect location memory retention incremental generalization method, including the following steps:

[0049] Step S1: Obtain the incremental data domain, defect location detection model, and historical representation agent for the current stage of semiconductor defect location detection. The defect location detection model includes a feature extraction module and a defect location module. The historical representation agent is used to save and aggregate the defect representation statistics learned in the past incremental stages.

[0050] First, image data of microscopic semiconductor workpieces arriving sequentially according to time, batch, workstation, process stage, or imaging condition are acquired, forming multiple incremental data domains:

[0051]

[0052] in, Indicates the first Data domains accessible in each incremental phase This indicates the total number of incremental stages. Different data fields can correspond to different workpiece batches, different microscopic equipment, different light source conditions, different magnification parameters, different inspection stations, different process stages, or different defect acquisition environments.

[0053] In the In each incremental phase, the model can only access the current data domain. Historical data domain The original image is not used in the playback; future data domain Pre-access is not allowed. Samples in each data domain are represented as follows:

[0054]

[0055] in, Represents an image of a microscopic semiconductor workpiece. This indicates the corresponding defect location and detection label. The label may include at least one of the following: defect category, defect bounding box, defect mask, defect center point, abnormal area, or product conformity.

[0056] Construct a defect location and detection model:

[0057]

[0058] in, This represents the model parameters. The model includes a feature extraction module and a defect localization module. The feature extraction module is used to extract defect characteristics from the microscopic image, and the defect localization module is used to output the defect location and defect category.

[0059] Obtaining historical representation agent The historical representation agent is used to store and aggregate defect representation statistics learned in past incremental phases. This agent does not contain the original historical micrographs, but rather stores historical knowledge in the form of category prototypes, defect region feature prototypes, normal region feature prototypes, or other low-dimensional representation statistics.

[0060] Step S2: Train the defect localization and detection model using the incremental data domain of the current stage, then input the incremental data domain of the current stage into the trained feature extraction module to obtain the current stage representation, obtain the current stage representation prototype through the current stage representation, and update the historical representation agent using the current stage representation prototype.

[0061] In the In each incremental stage, the defect location and detection model is trained using the incremental data domain of the current stage to obtain the model parameters for the current stage. Defect location loss Depending on the specific task, classification loss, bounding box regression loss, mask segmentation loss, anomaly localization loss, or a combination of the above losses can be used.

[0062] Then, the incremental data domain of the current stage is input into the model, and the feature extraction module obtains the representation of the current stage:

[0063]

[0064] in, This represents the feature extraction function of the model at the current stage. Specifically, it is based on the incremental data domain at the current stage. The sample labels in the model are used to aggregate the features extracted by the model at the current stage into categories or regions, resulting in the current stage representation prototype for each type of defect:

[0065]

[0066] in, Indicates the number of prototype categories. Indicates the first stage in the current phase Class defect or first The prototype representation of the class region. Current stage. The class representation prototype can be obtained through mean calculation:

[0067]

[0068] in, Indicates the first stage in the current phase Number of class samples This represents the indicator function. For defect localization and detection tasks, the current stage representation prototype can be calculated based on defect category, defect candidate box region features, defect mask region features, normal background region features, or multi-scale features.

[0069] Using the current stage to represent the prototype Update historical representation agent The update method is as follows:

[0070]

[0071] in, The momentum coefficient is used to control the fusion ratio between historical knowledge and current stage knowledge. Through the above update method, the historical representation agent can continuously integrate defect representation information from different workpiece batches, different inspection stations, and different imaging conditions as the incremental stage progresses, thereby forming a low-dimensional representation memory of historical domain defect knowledge.

[0072] Step S3: Construct the mutual information constraint loss between the current stage representation and the historical representation agent, and construct the defect localization loss of the defect localization module. Combine the defect localization loss and the mutual information constraint loss to obtain the joint loss.

[0073] To enable the current stage model to learn stable feature representations consistent with historical representation proxies, this invention constructs the current stage representation. With historical representation agent Mutual information constraints between the distributions are used to introduce conditional distributions and predict their distribution parameters using parameterized functions. A lightweight multilayer perceptron can be used to predict the mean and covariance of a Gaussian distribution.

[0074]

[0075] in, Indicates the representation based on the current stage. Inferred historical representation proxy The conditional probability distribution, Indicates a normal distribution. Representing historical representation agency, This indicates the current stage of characteristics. and These represent the characteristics of the current stage. The predicted mean and covariance.

[0076] Based on the variational modeling approach described above, a mutual information constraint loss is constructed:

[0077]

[0078] in, Indicates the current incremental data field. This indicates the current incremental data field All samples Find the average value. Represents image data, Indicates a label, Representing historical characteristics Mean of MLP prediction The square of the Euclidean distance between them Represents the covariance matrix of MLP predictions The logarithm of the determinant.

[0079] This loss is used to constrain the microscopic defect representations extracted by the model in the current stage to remain consistent with the historical representation proxy, so that the model can preserve historical knowledge without accessing the original historical images.

[0080] The joint loss can be obtained by acquiring the joint defect localization loss and the mutual information constraint loss:

[0081]

[0082] in, , where is the weighting coefficient, used to balance the defect localization loss and the mutual information constraint loss. By jointly optimizing the above losses, the current model can maintain consistency with historical representation proxies while learning the defect localization capability of the current workpiece batch or the current imaging conditions, thereby mitigating the forgetting of historical defect knowledge and improving the generalization ability to future unknown data domains.

[0083] Step S4: Train the defect localization and detection model using the incremental data domain of the current stage to obtain the defect localization and detection model for the next stage. Repeat the above steps until all incremental stages are trained, and finally obtain the memory-preserving incremental generalization model.

[0084] During the training of the current incremental stage model using joint loss, multiple model checkpoints are saved during the model convergence interval:

[0085]

[0086] in, This indicates the number of checkpoints saved. Different checkpoints correspond to different parameter states in the current training phase, and their ability to preserve historical knowledge and generalize to the future domain may vary.

[0087] Specifically, the mutual information constraint loss corresponding to each model checkpoint is first calculated. , indicating the first The mutual information constraint loss corresponds to each model checkpoint. The dynamic ensemble weights are calculated based on the mutual information constraint loss.

[0088]

[0089] in This represents the temperature coefficient. The smaller the mutual information loss, the more consistent the model's representation at checkpoints is with historical representations, and the larger the corresponding dynamic ensemble weights.

[0090] Based on the dynamic ensemble weights, a weighted ensemble is performed on multiple model checkpoints saved in the current stage to obtain the memory-preserving ensemble model for the current stage:

[0091]

[0092] in, This indicates that the memory retains the parameter states of the ensemble model, and n represents the number of checkpoints. This indicates the parameter status corresponding to the i-th checkpoint. This represents the weight of the i-th checkpoint.

[0093] Compared to a single checkpoint model, this ensemble model more stably retains historical domain defect knowledge and guides model parameters to a flatter, more transferable optimization region. When the next stage of the data domain... Upon arrival, the memory-retaining ensemble model obtained in the previous stage is used as the initialization model, and the above process is repeated until all incremental stages are trained, ultimately obtaining a memory-retaining incremental generalization model for the task of locating and detecting defects in microscopic semiconductor workpieces.

[0094] This method, without replaying historical original images or accessing future data in advance, makes full use of historical domain statistical information retained in past stage models or representations to provide cross-domain knowledge references for current stage model training and guides the model to learn more stable and transferable defect representations.

[0095] An embodiment of the present invention also provides a semiconductor defect location memory retention incremental generalization system, comprising:

[0096] The incremental data acquisition module is used to acquire the incremental data domain, defect location detection model, and historical representation agent for the current stage of semiconductor defect location detection. The defect location detection model includes a feature extraction module and a defect location module. The historical representation agent is used to save and aggregate the defect representation statistics learned in the past incremental stages.

[0097] The historical representation update module is used to input the incremental data domain of the current stage into the feature extraction module of the defect location detection model to obtain the current stage representation, and use the current stage representation to update the historical representation agent.

[0098] The mutual information constraint module is used to construct the mutual information constraint loss between the current stage representation and the historical representation agent, as well as to construct the defect localization loss of the defect localization module, and to combine the defect localization loss with the mutual information constraint loss.

[0099] The dynamic model integration module trains the defect location and detection model using the incremental data domain of the current stage to obtain the defect location and detection model for the next stage.

[0100] It should be noted that the system provided in the above embodiments is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the corresponding method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the corresponding method embodiments, which will not be repeated here.

[0101] Furthermore, one embodiment of the present invention provides an electronic device including a processor and a memory, the memory storing at least one instruction or at least one program, the at least one instruction or the at least one program being loaded and executed by the processor to implement the method provided in the above embodiments.

[0102] Memory can be used to store software programs and modules. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory. Memory can primarily include a program storage area and a data storage area. The program storage area can store the operating system, application programs required for the functions, etc.; the data storage area can store data created based on the use of the device, etc. Furthermore, memory can include high-speed random access memory, and can also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, memory can also include a memory controller to provide the processor with access to the memory.

[0103] The method embodiments provided in this invention can be executed in a terminal, server, or similar computing device; that is, the aforementioned electronic device may include a terminal, server, or similar computing device. Servers can vary significantly due to differences in configuration or performance, and may include one or more central processing units (CPUs) (e.g., one or more processors) and memory, and one or more storage media (e.g., one or more mass storage devices) for storing applications or data. The memory and storage media may be temporary or persistent storage. The program stored in the storage media may include one or more modules, each module including a series of instruction operations on the server. Furthermore, the CPU may be configured to communicate with the storage media and execute a series of instruction operations stored in the storage media on the server. The server may also include one or more power supplies, one or more wired or wireless network interfaces, one or more input / output interfaces, and / or one or more operating systems, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0104] Input / output interfaces can be used to receive or send data via a network. Specific examples of such networks may include a wireless network provided by the server's communication provider. In one example, the input / output interface includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the input / output interface may be a Radio Frequency (RF) module used for wireless communication with the Internet. This wireless communication can use any communication standard or protocol, including but not limited to Global System for Mobile Communications (GSM), General Packet Radio Service (GPRS), Code Division Multiple Access (CDMA), Wideband Code Division Multiple Access (WCDMA), Long Term Evolution (LTE), email, and Short Messaging Service (SMS).

[0105] One embodiment of the present invention also provides a computer-readable storage medium that can be disposed in an electronic device to store at least one instruction or at least one program for implementing the above-described method, wherein the at least one instruction or the at least one program is loaded and executed by the processor to implement the method in the above-described embodiment.

[0106] Optionally, in embodiments of the present invention, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0107] One embodiment of the present invention also provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method provided in the above embodiments.

[0108] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0109] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0110] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A semiconductor defect location memory retention incremental generalization method, characterized in that, Including the following steps: Step S1: Obtain the incremental data domain, defect location detection model, and historical representation agent for the current stage of semiconductor defect location detection. The defect location detection model includes a feature extraction module and a defect location module. The historical representation agent is used to save and aggregate the defect representation statistics learned in the past incremental stages. Step S2: Train the defect localization and detection model through the incremental data domain of the current stage, input the incremental data domain of the current stage into the trained feature extraction module to obtain the current stage representation, obtain the current stage representation prototype through the current stage representation, and update the historical representation agent using the current stage representation prototype. Step S3: Construct the mutual information constraint loss between the current stage representation and the historical representation agent, and construct the defect localization loss of the defect localization module. Combine the defect localization loss and the mutual information constraint loss to obtain the joint loss. Step S4: Train the defect localization and detection model using the incremental data domain and joint loss of the current stage to obtain the defect localization and detection model for the next stage. Repeat the above steps until all incremental stages are trained to obtain the memory-preserving incremental generalization model. In step S4, training the defect location and detection model using the incremental data domain of the current stage to obtain the defect location and detection model for the next stage includes: Multiple checkpoints are saved within the convergence interval of model training; A memory-retaining ensemble model is constructed by combining the model parameters corresponding to multiple checkpoints, which serves as the defect localization and detection model for the next stage. The formula for constructing a memory retention integration model is: in, This indicates that the memory retains the parameter states of the ensemble model, and n represents the number of checkpoints. This indicates the parameter status corresponding to the i-th checkpoint. This represents the weight of the i-th checkpoint; The weight of the i-th checkpoint is: in, This represents the mutual information constraint loss corresponding to the i-th checkpoint. This represents the temperature coefficient.

2. The semiconductor defect location memory retention incremental generalization method according to claim 1, characterized in that, In step S3, constructing the mutual information constraint loss between the current stage representation and the historical representation proxy includes: By using a multilayer perceptron to predict the representation of the current stage, a historical representation proxy is obtained. in, Indicates a normal distribution. Representing historical representation agency, This indicates the current stage of characteristics. Indicates the representation based on the current stage. Inferred historical representation proxy The conditional probability distribution, and These represent the characteristics of the current stage. The predicted mean and covariance; Constructing mutual information constraint loss: in, Indicates the current incremental data field. This indicates that for all samples in the current incremental data domain Find the average value. Represents image data, Indicates a label, Representing historical representation agents Mean of MLP prediction The square of the Euclidean distance between them Represents the covariance matrix of MLP predictions The logarithm of the determinant.

3. The semiconductor defect location memory retention incremental generalization method according to claim 2, characterized in that, In step S3, the formulas for the joint defect localization loss and the mutual information constraint loss are as follows: in, Indicates joint loss, Indicates the loss in defect location. These are the weighting coefficients.

4. The semiconductor defect location memory retention incremental generalization method according to claim 3, characterized in that, The defect localization loss can be classified loss, bounding box regression loss, mask segmentation loss, anomaly localization loss, or a combination of the above losses.

5. The semiconductor defect location memory retention incremental generalization method according to claim 1, characterized in that, The prototype of the current stage representation is obtained by averaging the current stage representation.

6. A semiconductor defect location memory retention incremental generalization system, used to execute the semiconductor defect location memory retention incremental generalization method according to any one of claims 1-5, characterized in that, include: The incremental data acquisition module is used to acquire the incremental data domain, defect location detection model, and historical representation agent for the current stage of semiconductor defect location detection. The defect location detection model includes a feature extraction module and a defect location module. The historical representation agent is used to save and aggregate the defect representation statistics learned in the past incremental stages. The historical representation update module trains the defect localization and detection model through the current stage incremental data domain, then inputs the current stage incremental data domain into the trained feature extraction module to obtain the current stage representation, and uses the current stage representation to update the historical representation agent. The mutual information constraint module is used to construct the mutual information constraint loss between the current stage representation and the historical representation agent, and to construct the defect localization loss of the defect localization module, and to combine the defect localization loss with the mutual information constraint loss. The dynamic model integration module trains the defect location and detection model using the incremental data domain of the current stage to obtain the defect location and detection model for the next stage.

7. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the semiconductor defect location memory retention incremental generalization method according to any one of claims 1-5.

Images