Detection model training method, device and system and detection method

Abstract

The detection model training method provided by one embodiment of the invention comprises the following steps: acquiring a plurality of training data; the plurality of training data correspond to a plurality of workpiece categories; performing collection processing on the plurality of training data, and determining a plurality of training data sets in one-to-one correspondence with the plurality of workpiece categories; and performing model training based on the plurality of training data sets to obtain a detection model. The detection model trained by the detection model training method disclosed by the embodiment of the invention is higher in detection accuracy and less in missing detection.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor manufacturing technology, and in particular to a detection model training method, a detection model training device, a detection model training system and a detection method. Background Art

[0002] Due to the complex background of semiconductor wafers, traditional defect detection algorithms based on template matching will generate a large number of over-inspections, and these large numbers of over-inspections still require quality inspectors to check them one by one, which is very time-consuming. Due to the subjectivity and fatigue of humans, the detection effect is not good, so deep learning-based methods are currently widely used. However, the effect of deep learning algorithms is strongly dependent on training data. The amount of raw data that can be obtained on the actual production line is very large, but these data often have very serious imbalances. The model will learn much better for categories with large data volumes than for categories with small data volumes. When training deep learning models based on unbalanced data, the model will be more inclined to judge the data to be tested as a detection category with large data volumes, which will cause a great risk of missed detection and affect the detection accuracy.

[0003] Therefore, there is an urgent need to provide a new detection model training solution to solve the problem caused by the imbalance of training data. Summary of the invention

[0004] Therefore, in order to overcome at least some of the defects and problems in the prior art, the embodiments of the present invention provide a detection model training method, a detection model training device, a detection model training system and a detection model trained by a detection method, which have the characteristics of higher detection accuracy and fewer missed detections.

[0005] Specifically, on the one hand, an embodiment of the present invention provides a detection model training method, including: acquiring multiple training data; the multiple training data correspond to multiple artifact categories; aggregating the multiple training data to determine multiple training data sets corresponding one-to-one to the multiple artifact categories; and performing model training based on the multiple training data sets to obtain a detection model.

[0006] In one embodiment, the model training based on the multiple training data sets to obtain the detection model includes: a basic training step, performing basic training based on the multiple training data sets to obtain a basic model; a self-check step, using the basic model to perform self-check on the multiple training data sets to obtain self-check information; and an iterative training step, iteratively training the basic model based on the self-check information to obtain the detection model.

[0007] In one embodiment, the iterative training step specifically includes: when the self-check information meets a preset judgment condition, completing the iterative training; otherwise, updating the multiple training data sets according to the self-check information and returning to the basic training step.

[0008] In one embodiment, the self-test information includes a plurality of self-test noise values ​​corresponding one-to-one to the plurality of training data sets; when the plurality of self-test noise values ​​are all less than or equal to a preset threshold, the self-test information meets the preset judgment condition.

[0009] In one embodiment, the self-check information includes abnormal data among the multiple training data; updating the multiple training data sets according to the self-check information includes: updating the abnormal data to a target updated training data set among the multiple training data sets based on a re-judgment result of the abnormal data.

[0010] In one embodiment, the basic training based on the multiple training data sets to obtain the basic model includes: sampling the multiple training data sets one by one according to a preset sampling rule to obtain multiple groups of sampling data corresponding one to one to the multiple training data sets; and performing the basic training based on the multiple groups of sampling data to obtain the basic model.

[0011] In one embodiment, the multiple training data correspond to multiple feature categories, and the multiple training data sets include a target sampling data set; the target sampling data set includes multiple target sampling data subsets corresponding one-to-one to the multiple feature categories; the preset sampling rule includes: the number of data samplings of the multiple target sampling data subsets in the target sampling data set is equal.

[0012] In one embodiment, the multiple training data correspond to multiple feature categories; the multiple training data sets include a target training data set; the detection model training method also includes: classifying the target training data set according to the feature categories to obtain multiple target training data subsets corresponding one-to-one to the multiple feature categories.

[0013] An embodiment of the present invention provides a detection model training device, including: a training data acquisition module, used to acquire multiple training data; the multiple training data correspond to multiple artifact categories; a data collection module, used to collect and process the multiple training data, and determine multiple training data sets corresponding one-to-one to the multiple artifact categories; a training module, used to perform model training based on the multiple training data sets to obtain a detection model.

[0014] An embodiment of the present invention provides a detection method, comprising: acquiring an image of a workpiece to be detected; detecting the image using a detection model, wherein the detection model is trained by the model training method described in any one of the aforementioned embodiments.

[0015] An embodiment of the present invention provides a detection model training system, including a processor and a memory connected to the processor, the memory storing instructions executed by the processor, and the instructions causing the processor to perform operations to perform the detection model training method described in any one of the aforementioned embodiments.

[0016] An embodiment of the present invention provides a readable medium, wherein the readable medium stores computer-readable instructions, wherein the computer-readable instructions include instructions for executing the detection model training method described in any of the aforementioned embodiments.

[0017] It can be seen from the above that the above embodiments of the present invention can achieve one or more of the following beneficial effects: by aggregating multiple training data sets to obtain multiple workpiece categories corresponding one to one, compared to directly using multiple training data mixed with different workpiece categories as training samples, this embodiment can achieve data balance at the workpiece category level, can implement model training based on different workpiece categories, improve the accuracy of the model, and reduce missed detections. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 A flowchart of a detection model training method provided by one embodiment of the present invention.

[0020] Figure 2 A schematic diagram of multiple training data sets in one embodiment of the present invention.

[0021] Figure 3 It is a schematic diagram of multiple target training data subsets in a target training data set in one embodiment of the present invention.

[0022] Figure 4 Schematic diagram of self-check information in one embodiment of the present invention.

[0023] Figure 5 A schematic diagram of the structure of a detection model training device provided by an embodiment of the present invention.

[0024] Figure 6A schematic diagram of the structure of a detection model training system provided for one embodiment of the present invention.

[0025] Figure 7 A schematic diagram of the structure of a readable medium provided for one embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0027] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work should fall within the scope of protection of the present invention.

[0028] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0029] It should also be noted that the division of multiple embodiments in the present invention is only for the convenience of description and should not constitute a special limitation. The features in various embodiments can be combined and referenced to each other without contradiction.

[0030] [First embodiment]

[0031] Reference Figure 1 The first embodiment of the present invention provides a detection model training method, comprising:

[0032] Step S1: Acquire a plurality of training data; the plurality of training data correspond to a plurality of workpiece categories;

[0033] Step S3: performing a collection process on the plurality of training data to determine a plurality of training data sets corresponding one-to-one to the plurality of workpiece categories;

[0034] Step S5: Performing model training based on the multiple training data sets to obtain a detection model.

[0035] Among them, the detection model trained by the detection model training method can be used for feature detection of workpieces such as wafers, for example, for detecting the defect type of the workpiece.

[0036] The multiple training data obtained in step S1 are, for example, images of multiple workpieces (which may be called devices). The multiple training data correspond to multiple categories of workpieces. For example, some of the multiple training data are images of category A workpieces (which may be called device A), some are images of category B workpieces (which may be called device B), and some are images of other categories of workpieces. Images of different categories of workpieces have, for example, different texture backgrounds. In step S3, the multiple training data are grouped together. For example, the training data of the aforementioned device A is grouped into one training data set, and the training data of the aforementioned device B is grouped into another training data set. The training data in each training data set corresponds to the same category of device. Figure 2 As shown, each folder (ie, each training data set) corresponds to a device category. In step S5, model training can be performed based on multiple training data sets according to the device category.

[0037] The model training in the related technology is a hybrid training, for example, only the training data is classified according to the defect type, for example, all defect-free images (such as Good data) are composed of a Good data set, and all defective images (such as Bad data) are composed of a Bad data set. This hybrid training has serious imbalance. For example, the Good data of device A occupies 1,000 in the total Good data set, while the Bad data of device A only occupies 20 in the total Bad data set. Direct training will make the model easy to miss device A. Therefore, through the detection model training method provided in this embodiment, by grouping multiple training data according to the workpiece category in step S2, targeted data balancing, that is, device-by-device balancing, can be achieved, thereby reducing missed detections and improving detection accuracy.

[0038] In some embodiments, the plurality of training data corresponds to a plurality of feature categories. The plurality of training data sets include a target training data set, and the detection model training method further includes step S4: classifying the target training data set according to feature categories to obtain a plurality of target training data subsets corresponding to the plurality of feature categories.

[0039] Among them, the multiple feature categories refer to defect categories, for example, according to the binary classification, they can be divided into non-defective data (i.e., Good data) and defective data (i.e., Bad data). Of course, there can be more classifications, such as further classifying defective data according to defect type, location, size and other features. In step S4, each of the multiple training data sets can be used as a target training data set and classified according to the feature category. Taking the binary classification as an example, in step S4, the training data in the target training data set is divided into a Good training data subset and a Bad training data subset according to Good data and Bad data. For example, refer to Figure 3 , the training data set corresponding to the number 2AI0CW6P2 is taken as the target training data set, and the training data set includes a Good training data subset and a Bad training data subset. The training data in the Good training data subset are all training data with the artifact category of 2AI0CW6P2 and the feature category of Good; the training data in the Bad training data subset are all training data with the artifact category of 2AI0CW6P2 and the feature category of Bad.

[0040] In this embodiment, by further performing classification within the device, the overall data set and the distribution of each training data set can be better understood, so as to facilitate data balancing at the feature category level.

[0041] In one embodiment, step S3 includes, for example, steps S31 to S33.

[0042] Step S31: a basic training step, performing basic training based on the multiple training data sets to obtain a basic model;

[0043] Step S33: a self-checking step, using the basic model to perform self-checking on the plurality of training data sets to obtain self-checking information;

[0044] Step S35: an iterative training step, in which the basic model is iteratively trained based on the self-test information to obtain the detection model.

[0045] In one embodiment, step S31 specifically includes step S311: sampling the multiple training data sets one by one according to a preset sampling rule to obtain multiple groups of sampling data corresponding one to one with the multiple training data sets; step S313: performing the basic training based on the multiple groups of sampling data to obtain the basic model.

[0046] Before step S311, for example, the preset sampling rules are also defined, and before step S313, for example, the appropriate deep learning model and parameters are also defined. In one embodiment, the multiple training data sets include a target sampling data set; the target sampling data set includes multiple target sampling data subsets corresponding one-to-one to the multiple feature categories; the preset sampling rules include, for example: the number of data samplings of the multiple target sampling data subsets in the target sampling data set is equal. In step S311, each training data set in the multiple training data sets can be used as a target sampling data set to sample the multiple training data sets one by one. For the target sampling data set, taking the binary method as an example, the sampling number of the Good training data subset should be equivalent to the sampling number of the Bad training data subset. For example, the number of data in the Good training data subset is usually greater than the number of data in the Bad training data subset, and the data in the Bad training data subset can be sampled after incremental processing.

[0047] In step S33, a basic model is used to perform self-inspection on multiple training data sets. Specifically, the basic model is used to infer the training data of different feature categories under each training data set one by one, and the result obtained by using the basic model to infer is compared with the label of the feature category of the training data itself (that is, which training data subset the training data belongs to). When the inference result is inconsistent with the label category, the training data is judged to be abnormal data. For example, a target comparison training data in the Good training data subset in the target training data set is inferred using the basic model. If the inference result determines that the type of the target comparison training data is Bad, then the target comparison training data is abnormal data. The ratio of the number of all abnormal data detected in each training data set to the total amount of all training data participating in the self-inspection in the training data set is recorded as the self-inspection noise value of the training data set, and the range of the self-inspection noise value is 0 to 1. Multiple self-inspection noise values ​​corresponding to multiple training data sets can be obtained. The self-inspection information includes the multiple self-inspection noise values, and the self-inspection information also includes, for example, abnormal data detected in the self-inspection, and label information of the abnormal data itself. Refer to Figure 4 The first field of each line is the identifier of the training data set, that is, the device ID, and the following numbers are the self-test noise values ​​of the device (the artifact category corresponding to the training data set).

[0048] Step S35 specifically includes: (S35.1) when the self-test information meets the preset judgment condition, completing the iterative training; otherwise, (S35.2) updating the multiple training data sets according to the self-test information, and returning to the basic training step. Specifically, when the self-test information does not meet the preset judgment condition, the multiple training data sets are updated according to the self-test information and then the process returns to step S31, and steps S31 to S35 are repeatedly executed until the self-test information meets the preset judgment condition, and the iterative training is completed to obtain the final detection model.

[0049] Specifically, the self-test information includes the aforementioned multiple self-test noise values ​​corresponding one-to-one to the multiple training data sets. In step S35, when it is determined that the multiple self-test noise values ​​are all less than or equal to the preset threshold, the self-test information meets the preset judgment condition. The preset threshold can be 0.001, for example, and can of course also be set according to actual needs.

[0050] In step S35.2 of some embodiments, for multiple training data sets, for example, the first N (N is a custom value) training data sets are updated according to the arrangement of the self-check noise value from large to small; step S35.2, for example, specifically includes: updating the abnormal data to the target updated training data set in the multiple training data sets according to the re-judgment result of the re-judgment of the abnormal data. For example, before the update, the label of a certain abnormal data is Good data, and the self-check determines that it is Bad data. According to the re-judgment of the abnormal data, for example, it is confirmed through manual re-judgment that it is actually Bad data, then the abnormal data is updated from the original Good training data subset to the Bad training data subset. Of course, in the manual re-judgment process, it can also be determined whether the training data set corresponding to the abnormal data before the update is correct. If it is not correct, the abnormal data set should also be updated to the correct training data set. This embodiment introduces a cyclic process of self-checking, re-judgment, and iteration, which can improve the training accuracy of model training and save labor costs.

[0051] This embodiment also provides a detection method, which can be used for feature detection of workpieces such as wafers, specifically, for example, wafer defect detection. The detection method includes: obtaining a test image of the workpiece to be tested; using a detection model to detect the test image, and the detection model is trained by any of the aforementioned model training methods.

[0052] Using the detection model trained by the model training method provided in the above embodiment of the present invention for detection can improve the detection accuracy and reduce the occurrence of missed detection.

[0053] [Second embodiment]

[0054] Reference Figure 5The second embodiment of the present invention provides a detection model training device 100, including a training data acquisition module 11, a data collection module 12 and a training module 13. The training data module 11 is used to execute step S1 in the aforementioned detection model training method, the data collection module 12 is used to execute step S3 in the aforementioned detection model training method, and the training module 13 is used to execute step S5 in the aforementioned detection model training method. In some embodiments, the data collection module 12 is also used to execute step S4 in the aforementioned detection model training method. For the specific principles of the detection model training device 100 provided in this embodiment, reference can be made to the description in the aforementioned first embodiment. The detection model training device 100 provided in this embodiment is used to execute the aforementioned detection model training method, and has the same beneficial effects as the aforementioned first embodiment, which will not be repeated here.

[0055] [Third embodiment]

[0056] Reference Figure 6 The third embodiment of the present invention provides a detection model training system 200, which includes, for example, a processor 201 and a memory 202 electrically connected to the processor 201, wherein the memory 202 stores instructions executed by the processor 201, and the instructions enable the processor 201 to perform operations to perform any one of the detection model training methods in the aforementioned first embodiment.

[0057] The functions and technical effects of the detection model training system 200 provided in the third embodiment of the present invention can be referred to the relevant description of the detection model training method in the aforementioned first embodiment, and will not be repeated here.

[0058] [Fourth embodiment]

[0059] Reference Figure 7 The fourth embodiment of the present invention provides a readable medium 300, which stores computer-readable instructions 301. The computer-readable instructions 301 include instructions for executing any one of the detection model training methods in the first embodiment.

[0060] The functions and technical effects of the readable medium 300 provided in the fourth embodiment of the present invention can refer to the technical effects of the detection model training method in the aforementioned first embodiment, and will not be repeated here.

[0061] It should be noted that in the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and / or methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units / modules is only a logical function division. There may be other division methods in actual implementation, such as multiple units or modules can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0062] The units / modules described as separate components may or may not be physically separated, and the components shown as units / modules may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units / modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0063] In addition, each functional unit / module in each embodiment of the present invention may be integrated into one processing unit / module, or each unit / module may exist physically separately, or two or more units / modules may be integrated into one unit / module. The above-mentioned integrated unit / module may be implemented in the form of hardware or in the form of hardware plus software functional units / modules.

[0064] The above-mentioned integrated unit / module implemented in the form of a software functional unit / module can be stored in a computer-readable storage medium. The above-mentioned software functional unit is stored in a storage medium, including a number of instructions for causing one or more processors of a computer device (which can be a personal computer, a server, or a network device, etc.) to perform some steps of the method described in each embodiment of the present invention.

[0065] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A detection model training method, characterized in that: include: Get multiple training data; The plurality of training data correspond to a plurality of artifact categories; Performing grouping processing on the plurality of training data to determine a plurality of training data sets corresponding one-to-one to the plurality of workpiece categories; Model training is performed based on the multiple training data sets to obtain a detection model.

2. The detection model training method according to claim 1, characterized in that: The performing model training based on the multiple training data sets to obtain the detection model includes: A basic training step, performing basic training based on the multiple training data sets to obtain a basic model; A self-checking step, using the basic model to perform self-checking on the multiple training data sets to obtain self-checking information; An iterative training step is to iteratively train the basic model based on the self-test information to obtain the detection model.

3. The detection model training method according to claim 2, characterized in that: The iterative training step specifically includes: completing the iterative training when the self-check information meets the preset judgment condition; otherwise, updating the multiple training data sets according to the self-check information and returning to the basic training step.

4. The detection model training method according to claim 3, characterized in that: The self-test information includes a plurality of self-test noise values ​​corresponding one-to-one to the plurality of training data sets; when the plurality of self-test noise values ​​are all less than or equal to a preset threshold, the self-test information meets the preset judgment condition.

5. The detection model training method according to claim 3, characterized in that: The self-check information includes abnormal data in the plurality of training data; The updating of the plurality of training data sets according to the self-check information includes: updating the abnormal data to a target updated training data set in the plurality of training data sets according to a re-judgment result of re-judgment of the abnormal data.

6. The detection model training method according to claim 2, characterized in that: The step of performing basic training based on the multiple training data sets to obtain a basic model includes: Sampling the multiple training data sets one by one according to a preset sampling rule to obtain multiple groups of sampling data corresponding to the multiple training data sets one by one; The basic model is obtained by performing the basic training based on the multiple groups of sampling data.

7. The detection model training method according to claim 6, characterized in that: The plurality of training data corresponds to a plurality of feature categories, the plurality of training data sets include a target sampling data set; the target sampling data set includes a plurality of target sampling data subsets corresponding one-to-one to the plurality of feature categories; The preset sampling rule includes: the data sampling numbers of the multiple target sampling data subsets in the target sampling data set are equal.

8. The detection model training method according to claim 1, characterized in that: The multiple training data correspond to multiple feature categories; The multiple training data sets include a target training data set; The detection model training method also includes: classifying the target training data set according to feature categories to obtain multiple target training data subsets corresponding to the multiple feature categories.

9. A detection model training device, characterized in that: include: A training data acquisition module, used to acquire multiple training data; The plurality of training data correspond to a plurality of artifact categories; A data collection module, used for collecting the plurality of training data and determining a plurality of training data sets corresponding to the plurality of workpiece categories; A training module is used to perform model training based on the multiple training data sets to obtain a detection model.

10. A detection method, characterized in that: include: Acquire the image of the workpiece to be tested; The image to be tested is detected using a detection model, wherein the detection model is trained by any one of the model training methods described in claims 1 to 8.

11. A detection model training system, characterized in that: It includes a processor and a memory connected to the processor, the memory stores instructions executed by the processor, and the instructions enable the processor to perform operations to perform the detection model training method as described in any one of claims 1-8.

Images