Automatic defect classification system and its training and classification method

Abstract

The present invention provides an automatic defect classification system and its training and classification methods. Among them, the over-detection and missed detection judgment unit in the automatic defect classification system is used to judge whether the data belongs to over-detection data or missed detection data, so as to separate the over-detection and missed detection data. The over-detection and missed detection cause classifier is trained after extracting the first feature from the over-detection and missed detection data set, and is used to classify the over-detection and missed detection causes of the over-detection and missed detection data, which is beneficial to quickly locate the causes of over-detection and missed detection during the online detection process, and reduce the probability of over-detection and missed detection. The supplementary classifier is independent of the automatic defect classifier, and is at least trained using data belonging to the causes of the defect classification system as samples, which reduces the problem of incorrect data classification due to the reasons of the automatic defect classifier itself. It does not interfere with the normal operation of the automatic defect classifier, and improves the accuracy of defect classification of the detected object, the detection accuracy of the detection equipment, and the stability of the machine.

Description

Technical Field

[0001] The present invention relates to the technical field of integrated circuit manufacturing, and in particular to an automatic defect classification system and a training and classification method thereof. Background Art

[0002] During the semiconductor product manufacturing process, defects may appear on the wafer surface during various processes such as single crystal pulling, slicing, grinding, polishing, deposition, photolithography, doping, heat treatment, probe testing, and scribing. These defects directly affect the product's service life and reliability. Therefore, wafer defect classification is particularly important.

[0003] Existing wafer defect classification typically utilizes an automatic defect classifier (ADC) for defect identification and classification. However, the classification rules or machine learning algorithms based on the ADC often result in over-inspections and missed inspections of defects. Over-inspection refers to the rejection of images without defects, while missed inspection refers to the acceptance of images with defects. Therefore, as the number of over-inspections and missed inspections increases, the wafer defect recognition accuracy decreases and the detection recall rate increases. To address this, the existing solution is to modify the inspection recipe when the number of over-inspection and missed inspection data exceeds a set threshold. However, over-inspection and missed inspection data caused by the ADC itself still exist, and modifying the inspection recipe will also cause the original ADC's classification results for non-over-inspection and missed inspection data to deteriorate.

[0004] Therefore, a new automatic defect classification system is needed to be able to perform fault analysis and defect classification on the data that have been inspected or missed, as well as to classify the data that have been inspected or missed due to reasons of the automatic defect classifier itself, without affecting the classification of the data that have not been inspected or missed, thereby improving the accuracy of wafer defect classification. Summary of the Invention

[0005] The object of the present invention is to provide an automatic defect classification system and a training and classification method thereof to solve at least one of the problems of how to improve the accuracy of wafer defect classification and how to improve the detection accuracy of detection equipment.

[0006] To solve the above technical problems, the present invention provides a training method for an automatic defect classification system, comprising:

[0007] According to the pre-acquired defect classification database, the passed inspection data and missed inspection data are collected to form a passed and missed inspection data set;

[0008] Marking the data in the over-detection and missed detection data set with the reasons for over-detection and missed detection; wherein the reasons for over-detection and missed detection include at least detection system reasons and defect classification system reasons;

[0009] Performing a first feature extraction on the data to train a missed detection reason classifier, wherein the missed detection reason classifier is used to classify missed detection reasons of the data during the detection process;

[0010] At least taking the data belonging to the cause of the defect classification system as a sample, and labeling the sample with a defect category;

[0011] A second feature is extracted from the sample of the marked defect category to train a supplementary classifier, which is used to reclassify the data classified as defect classification system reasons by the over-detection and missed detection reason classifier during the detection process.

[0012] Optionally, in the training method of the automatic defect classification system, performing first feature extraction on the data to train a classifier for causes of over-detection and missed-detection includes:

[0013] The first feature is extracted from the data to form a first feature vector; wherein the first feature includes multiple feature values, and the types of the feature values ​​in the first feature extracted from each data are the same; multiple first feature vectors formed by multiple data constitute a first feature space, and the first feature vector is matched with the over / miss detection reasons annotated by the corresponding data to train and form the over / miss detection reason classifier.

[0014] Optionally, in the training method of the automatic defect classification system, training a classifier for missed detection reasons includes:

[0015] A plurality of classification algorithms are selected for training respectively, and the classification algorithm that meets the set requirements is selected from the plurality of evaluation results obtained as the classification algorithm of the over / under detection reason classifier.

[0016] Optionally, in the training method of the automatic defect classification system, extracting a second feature from the sample with the labeled defect category to train a supplementary classifier includes:

[0017] The second feature is extracted from the sample to form a second feature vector; wherein the second feature includes multiple feature values, and the feature values ​​in the second feature extracted from each sample are of the same type; the multiple second feature vectors formed by the multiple samples constitute a second feature space, and the second feature vectors are matched with the defect categories marked by the corresponding samples to train and form the supplementary classifier.

[0018] Optionally, in the training method of the automatic defect classification system, the feature values ​​include: inspection machine, inspection parameters, software attributes, inspection object attributes, inspection object process information, inspection object defect information or inspection object defect classification results.

[0019] Optionally, in the training method of the automatic defect classification system, the feature value in the second feature also includes: defect grayscale information and / or detection optical mode.

[0020] Optionally, in the training method of the automatic defect classification system, training to form the supplementary classifier further includes: selecting a plurality of classification algorithms for separate training, and selecting the classification algorithm that meets the set requirements from the plurality of evaluation results obtained as the classification algorithm of the supplementary classifier.

[0021] Optionally, in the training method of the automatic defect classification system, the classification algorithm includes: K-nearest neighbor method, decision tree, random forest ensemble learning, neural network or support vector machine.

[0022] Optionally, in the training method of the automatic defect classification system, the classification algorithm adopted by the pre-obtained defect classification database is different from the classification algorithm of the supplementary classifier; and\or, the type of the feature value in at least part of the second feature is different from the type of feature value selected by the pre-obtained defect classification database.

[0023] Optionally, in the training method of the automatic defect classification system, the samples used to train the supplementary classifier include the over-detection and missed detection dataset.

[0024] Optionally, in the training method of the automatic defect classification system, the samples used to train the supplementary classifier include data of normal inspection objects.

[0025] Based on the same inventive concept, the present invention also provides an automatic defect classification system, comprising: a data acquisition unit, a data storage unit, an over / miss detection determination unit, an automatic defect classifier, an over / miss detection cause classifier, and a supplementary classifier;

[0026] The data acquisition unit is used to acquire data of the detected object;

[0027] The data storage unit is used to store data acquired and generated by the data acquisition unit, the automatic defect classifier, the over / under detection determination unit, the over / under detection cause classifier, and the supplementary classifier;

[0028] The over-detection / missed detection determination unit is used to perform similarity comparison on the data of the detected object to determine whether the data is over-detected data or missed detection data;

[0029] The automatic defect classifier is used to perform defect recognition and defect classification on data that does not belong to the passed-inspection data or the missed-inspection data;

[0030] The over-detection and missed detection reason classifier is used to classify the over-detection data or the missed detection data according to the over-detection and missed detection reasons; wherein the over-detection and missed detection reasons at least include detection system reasons or defect classification system reasons;

[0031] The supplementary classifier is at least used to classify the data belonging to the defect classification system into defect categories.

[0032] Optionally, in the automatic defect classification system, the similarity comparison method includes Euclidean distance, Manhattan distance, Chebyshev distance, Minkowski distance or Pearson correlation coefficient.

[0033] Based on the same inventive concept, the present invention also provides an automatic defect classification method, comprising:

[0034] The data acquisition unit acquires the data of the detected object and stores it in the data storage unit;

[0035] The over / under detection determination unit obtains the data of the inspection object from the data storage unit and determines whether the data is within the over / under detection determination threshold range. If so, the data is input into the over / under detection cause classifier; if not, the data is input into the automatic defect classifier to perform defect identification and defect classification on the data.

[0036] The over / miss detection reason classifier classifies the input data into over / miss detection reasons; wherein the data belonging to the defect classification system reasons is input into the supplementary classifier;

[0037] The supplementary classifier classifies the input data into defect categories.

[0038] Optionally, in the automatic defect classification method, after the over-detection and missed detection reason classifier classifies the input data according to the over-detection and missed detection reasons, the data belonging to the detection system reasons is input into the data detection unit after the detection conditions are adjusted.

[0039] In summary, the present invention provides an automatic defect classification system and its training and classification methods. The automatic defect classification system includes: a data acquisition unit, a data storage unit, an over / miss detection determination unit, an automatic defect classifier, an over / miss detection cause classifier, and a supplementary classifier. The data acquisition unit is used to acquire data from the inspection object. The data storage unit is used to store the data. The over / miss detection determination unit is used to determine whether the data is over / miss detection data or missed detection data, thereby separating the over / miss detection data for subsequent separate defect classification of the over / miss detection data. The automatic defect classifier is used to identify and classify defects in data that is not over / miss detection data or missed detection data. The over / miss detection cause classifier is trained using a over / miss detection dataset through first feature extraction and classification, and is used to classify the over / miss detection causes of over / miss detection data. This facilitates rapid identification of the causes of over / miss detection during online testing, reduces the probability of over / miss detection, and improves detection accuracy. The supplementary classifier, independent of the automatic defect classifier, is used to classify the defect category of the data caused by the defect classification system, thus avoiding the problem of data being unable to be classified due to the automatic defect classifier itself. Furthermore, there is no need to modify the inspection formula of the automatic defect classifier, and there is no impact on the defect recognition and classification of the inspection objects that are not missed inspections.

[0040] Therefore, the present invention realizes fault analysis and defect classification of over-detection and missed detection data separately through offline trained over-detection and missed detection cause classifiers and supplementary classifiers, and realizes defect classification of over-detection and missed detection data caused by the automatic defect classifier itself, without interfering with the normal operation of the automatic defect classifier, thereby improving the accuracy of defect classification of the detected object and the detection accuracy of the detection equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of the composition of the automatic defect classification system in an embodiment of the present invention;

[0042] Figure 2 This is the process intention of the training method of the automatic defect classification system in the embodiment of the present invention;

[0043] Figure 3 is a schematic diagram of a first eigenvector in an embodiment of the present invention;

[0044] Figure 4 4 is a flow chart of an automatic defect classification method in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to make the objects, advantages and features of the present invention clearer, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the drawings are all in a very simplified form and are not drawn to scale, and are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention. In addition, the structure shown in the drawings is often a part of the actual structure. In particular, the emphasis required to be shown in each drawing is different, and sometimes different scales are used. It should also be understood that, unless otherwise specified or indicated, the terms "first", "second", "third" and the like in the specification are only used to distinguish between the various components, elements, steps, etc. in the specification, and are not used to represent the logical relationship or sequential relationship between the various components, elements, steps, etc.

[0046] To address the above technical issues, this embodiment provides an automatic defect classification system for classifying defects in an inspection object. The inspection object in this embodiment includes, but is not limited to, wafers, and may also refer to specific products such as CMOS devices. In the following description, wafers are used as an example of the inspection object.

[0047] See also Figure 1 The automatic defect classification system includes: a data acquisition unit 100, a data storage unit 101, an over / miss detection judgment unit 102, an automatic defect classifier 103, an over / miss detection cause classifier 104, and a supplementary classifier 105. The data acquisition unit 100 is used to acquire data of the inspection object. That is, by taking a photo of the wafer, the information in the photo is extracted as the wafer data. The data storage unit 101 is used to store data in the entire system, such as the data acquired and generated by the data acquisition unit 100, the over / miss detection judgment unit 102, the automatic defect classifier 103, the over / miss detection cause classifier 104, and the supplementary classifier 105. The over / miss detection judgment unit 102 is used to determine whether the data is over / miss detection data or missed detection data based on the data of the inspection object. The automatic defect classifier 103 is used to identify and classify defects in data that does not belong to the over-inspected data or the missed inspection data. It can be understood as an automatic defect classifier that comes with the existing defect identification system. According to the set detection formula, the wafer is detected and classified, and the acquired data is stored in the defect classification database. The defect classification database is used to store information on all the wafers and the corresponding defect types. The over-inspection and missed inspection cause classifier 104 is used to classify the over-inspection data and the missed inspection data into over-inspection and missed inspection causes; wherein, the over-inspection and missed inspection causes include at least detection system causes and defect classification system causes. The supplementary classifier 105 is used to at least classify the data that belongs to the defect classification system causes into defect categories.

[0048] It can be seen that the automatic classification system provided in this embodiment, by adding an over / miss detection judgment unit 102, an over / miss detection cause classifier 104 and a supplementary classifier 105, can realize fault analysis and defect classification of over / miss detection data separately, and can realize defect classification of over / miss detection data caused by the automatic defect classifier 103 itself, without interfering with the normal operation of the automatic defect classifier 103, thereby improving the accuracy of defect classification of the detected object and the detection accuracy of the detection equipment.

[0049] Furthermore, the data acquisition unit 100 usually includes an objective lens and a camera to take pictures of the wafer as data. The over-detection and missed detection judgment unit 102 is for screening over-detection and missed detection data. A similarity comparison is performed on the data of the detection object to obtain a distance value, and it is judged whether the distance value is within the over-detection and missed detection threshold range. If so, it can be identified as over-detection and missed detection data. If not, it does not belong to over-detection and missed detection data, so that all data that may be over-detected and missed detection can be screened out for subsequent separate defect classification. Optionally, a distance measurement method is used to perform a similarity comparison on the data. The distance measurement method includes but is not limited to: Euclidean distance, Manhattan distance, Chebyshev distance, Minkowski distance or Pearson correlation coefficient. Among them, the offline over-detection and missed detection data can be used for similarity comparison to obtain the over-detection and missed detection threshold range. Then, the distance measurement method is used to calculate the distance value corresponding to the data, and then the distance relationship between the data and the over-detection and missed detection threshold range is judged. If the distance between the detected data exceeds the over / miss detection threshold, it is considered not to be over / miss detection data. If the distance falls within the over / miss detection threshold, it is considered to be over / miss detection data, indicating that the data may have an over / miss detection fault. The greater the distance, the lower the over / miss detection similarity; the smaller the distance, the higher the over / miss detection similarity.

[0050] The over / miss detection reason classifier 104 classifies the causes of detection failures of over / miss detection data. That is, it refines the causes of the over / miss detection failures of the data, allowing engineers to quickly locate the causes of the over / miss detection failures and provide solutions to reduce the frequency of subsequent over / miss detection failures, thereby addressing the root cause of low detection accuracy.

[0051] The supplementary classifier 105 separately classifies the data that is caused by the defect classification system into defect categories. In the prior art, due to the defect classification system's own reasons, over-detection and missed detection may occur, and it is impossible to classify these missed detection data, resulting in low detection accuracy. In addition, solving this problem requires changing the detection formula of the inherent defect classification system, but this solution will affect the normal operation of the defect classification system and may even worsen the defect classification results of all data. Therefore, the automatic defect classification system provided in this embodiment separately sets up a classifier, namely the supplementary classifier 105, to solve the defect classification of missed detection data caused by the automatic defect classifier 103 itself. This design solution does not require modifying the detection formula of the automatic defect classifier 103, has no impact on the detection of the automatic defect classifier 103, and can realize defect classification of missed detection data caused by the automatic defect classifier 103 itself, thereby improving the system's defect classification accuracy and detection accuracy, and ensuring the stability of the machine.

[0052] To further introduce the automatic defect classification system, this embodiment also provides a training method for the automatic defect classification system, please refer to Figure 1-3 , the method comprising:

[0053] Step S10: Collect all passed inspection data and missed inspection data according to the pre-acquired defect classification database to form a passed and missed inspection data set;

[0054] The training method for the automatic defect classification system is performed offline. The pre-acquired defect classification database is a collection of data stored by the automatic defect classifier in the existing machine during all previous line inspections. The training method for the automatic defect classification system is established based on this database. The data that was missed or over-detected is then collected into a dataset to form a missed or over-detected dataset, which is then used for separate defect identification and classification of the missed or over-detected data.

[0055] Step S20: marking the reasons for over-detection and missed detection for the data in the over-detection and missed detection data set; wherein the reasons for over-detection and missed detection include at least reasons of the detection system and reasons of the defect classification system.

[0056] In actual operations, there are many reasons for over-detection and missed detection, such as: the machine not running to the correct position, incorrect detection parameter input, or classification rule reasons. To ensure that the reasons for the automatic defect classifier itself can be analyzed, this embodiment divides the reasons for over-detection and missed detection into at least two categories, one of which is the detection system reasons, and the other is the defect classification system reasons. The detection system reasons refer to over-detection and missed detection caused by the software and hardware in the entire detection machine. For example, the reasons for software parameter issuance errors, the error in the image template given in the detection algorithm, and the failure of the lens to move to the specified position. The defect classification system reasons refer to reasons belonging to the automatic defect classifier itself. For example, ADC algorithm defects, inconsistent distribution of the training data set and the actual detection data set, and non-optimal ADC features. Of course, according to needs, it can also be divided into multiple reasons for over-detection and missed detection, such as manual reasons. Furthermore, the reasons for over-detection and missed detection of each piece of data can be marked manually or automatically.

[0057] Step S30: performing first feature extraction on the data to train a missed detection reason classifier 104. The missed detection reason classifier 104 is used to classify missed detection reasons of the data during the detection process.

[0058] Furthermore, in the process of training the over / miss detection reason classifier 104, the first feature extracted from the data corresponds to the over / miss detection reason of the data to form a corresponding first feature vector. Wherein, the first feature includes multiple feature values, which correspond to multiple dimensions, and the types of the feature values ​​in the first feature extracted from each of the data are the same, that is, the first feature extraction method of each of the data is the same. The multiple first feature vectors formed by the multiple data constitute a first feature space, which is a multi-dimensional feature space. The first feature vector is matched with the over / miss detection reason marked by the corresponding data to train and form the over / miss detection reason classifier 104.

[0059] For example Figure 3As shown, the over-detection and missed detection data set includes a total of 6 data, each of which is marked with the reason for over-detection and missed detection. It is assumed that there are only two reasons for over-detection and missed detection, one is the detection system reason, and the other is the defect classification system reason, and 4 of the data are defect classification system reasons, and the other 2 are detection system reasons. Assuming that the first feature includes 3 eigenvalues, the same 3 eigenvalues ​​are extracted from each of the data as the first feature of each of the data. Therefore, each data corresponds to a reason for over-detection and missed detection, and corresponds to one of the first features, which is a first eigenvector. Among them, the type of eigenvalue in the first feature represents the feature dimension of the feature vector, and the 3 eigenvalues ​​correspond to 3 dimensions. Correspondingly, all the data in the over-detection and missed detection data set can form a 3*6-dimensional first feature space. Therefore, when the over-detection and missed detection data set includes N data and the first feature includes M eigenvalues, an M*N-dimensional first feature space can be formed.

[0060] Specifically, for example, if the reason for over- or under-detection of the data is due to a detection system problem and includes an offset X1 of the coordinate trajectory point of the lens, then the comparison result X1-N1 of the offset X1 and the offset N1 of the trajectory point under normal lens motion in the historical data is defined as feature_extract_1.

[0061] Feature_1=feature_extract_1(X1,N1);

[0062] Feature_1=X1–N1;

[0063] Feature_1 is a feature value. When there is only one feature value, Feature_1, the first feature is Feature_1. Accordingly, the data corresponds to the detection system cause and corresponds to Feature_1, which serves as the first feature vector of the data.

[0064] Therefore, in the process of detecting the causes of over-detection and missed detection, the causes of over-detection and missed detection of the data can be found in a targeted manner based on the first feature extracted from the data to be detected, so that engineers can promptly determine the type of fault and take corresponding solutions.

[0065] Furthermore, during the training process to form the over / miss detection reason classifier 104, multiple classification algorithms can be selected for separate training. From the multiple evaluation results obtained, the classification algorithm that meets the set requirements is selected as the classification algorithm for the over / miss detection reason classifier 104. The classification algorithms include, but are not limited to, the K-nearest neighbor method, decision tree, random forest ensemble learning, neural network, or support vector machine. Furthermore, the types of feature values ​​for the first feature include, but are not limited to, the following: inspection machine, inspection parameters, software attributes, wafer attributes, wafer process information, wafer defect information, or wafer defect classification results. Therefore, assuming training mode one, the first feature selected includes the following feature values: inspection machine, inspection parameters, software attributes, wafer attributes, wafer process information; and the K-nearest neighbor method is selected as the classification algorithm. Alternatively, training mode two includes the following feature values ​​for the first feature selected: inspection machine, inspection parameters, software attributes; and the neural network is selected as the classification algorithm. Then, based on the evaluation results of the two training modes, an optimal algorithm and feature value type are selected. Among them, the selection of the optimal evaluation result can comprehensively consider the following factors: classification accuracy, classification precision and classification recall rate.

[0066] Step S40: At least the data belonging to the defect classification system causes is used as a sample, and the sample is labeled with a defect category. In other words, each piece of data belonging to the defect classification system causes is considered a sample, and multiple pieces of data constitute multiple samples, forming a sample dataset. Each sample is then labeled with a defect category, either manually or automatically, to facilitate subsequent training of the supplementary classifier 105.

[0067] Step S50: extracting the second feature from the sample with the marked defect category to train a supplementary classifier 105, wherein the supplementary classifier 105 is used to reclassify the data classified as defect classification system reasons by the over-detection and missed detection reason classifier 105 during the detection process.

[0068] In this embodiment, the supplementary classifier 105 is used to identify and classify defects in data that was missed or over-detected due to reasons inherent in the automatic defect classifier 103. Of course, the supplementary classifier 105 can also be used to identify and classify defects in data caused by other reasons. However, data missed or over-detected due to issues with the inspection system does not require the use of the supplementary classifier 105 for defect identification and classification (see the introduction to the automatic defect classification method for details).

[0069] In addition, the sample data set for training the supplementary classifier 105 may also include the entire missed inspection data set, and may also include data of normal inspection objects. The normal inspection object data refers to data of wafers without defects or with defects within an allowable range.

[0070] Furthermore, during the training process for the supplementary classifier 105, the second features extracted from the samples correspond to the defect categories of the samples to form a corresponding second feature vector. The second features include multiple feature values, meaning the second feature vector has multiple dimensions. The feature values ​​in the second features extracted from each sample are of the same type, meaning the second features of each sample are extracted using the same method. The multiple second feature vectors generated by the multiple samples constitute a second feature space. The second feature vectors are matched to the defect categories labeled for the corresponding samples to train the supplementary classifier 105.

[0071] It will be appreciated that, similar to the method used to train the missed detection cause classifier 104, the supplementary classifier 105 also extracts a second feature for each sample. Specifically, each piece of data corresponds to a defect category and a second feature, recorded as a second feature vector. The types of eigenvalues ​​in the second feature also represent the dimensionality of the second feature vector. If the sample dataset includes A pieces of data and the second feature includes B types of eigenvalues, a B*A-dimensional second feature space can be formed.

[0072] Furthermore, the characteristic values ​​in the second feature include but are not limited to: detection machine, detection parameters, software properties, properties of the object to be detected, process information of the object to be detected, wafer defect information, wafer defect classification results, defect grayscale information and / or detection optical mode. In order to ensure the best effect, in the process of training to form the supplementary classifier 105, a variety of classification algorithms are selected for training respectively, and the classification algorithm that meets the set requirements is selected from the multiple evaluation results obtained as the classification algorithm of the supplementary classifier 105. In this regard, the classification algorithm can be screened with reference to the method of selecting the training mode during the training of the missed detection cause classifier 104. Further, the classification algorithm includes: K nearest neighbor method, decision tree, random forest ensemble learning, neural network or support vector machine. Further, the factors considered in selecting the optimal classification algorithm may include: classification accuracy, classification precision and classification recall.

[0073] Furthermore, to ensure accurate detection and classification of data that was missed or over-detected due to the automatic defect classifier 103 itself, the classification algorithm used by the pre-obtained defect classification database differs from the classification algorithm used by the supplementary classifier 105; and / or, the types of feature values ​​in at least some of the second features differ from the types of feature values ​​used in the pre-obtained defect classification database. In other words, the supplementary classifier 105 distinguishes itself from the defect classification criteria of the automatic defect classifier 103 by changing the classification algorithm and feature types, thereby accurately detecting and classifying data that was missed or over-detected due to the automatic defect classifier 103 itself. Furthermore, the supplementary classifier 105 is independent of the automatic defect classifier 103 and does not affect the detection of data that was not missed or over-detected.

[0074] Based on the above-mentioned training method of the automatic defect classification system, an automatic defect classification system can be established to realize fault analysis and defect classification of data that have been detected or not detected separately, as well as defect classification of data that have been detected or not detected due to reasons of the automatic defect classifier itself, without affecting the classification of data that have not been detected or not detected, thereby improving the accuracy of wafer defect identification.

[0075] Based on the same inventive concept, this embodiment also provides an automatic defect classification method, which uses the automatic defect system formed by the above training, such as Figure 1 and 4 As shown, the method includes:

[0076] Step 1: The data acquisition unit 100 acquires the data of the detected object and stores it in the data storage unit 101 .

[0077] For example, the data acquisition unit 100 takes a picture of a wafer, and stores the picture and wafer-related data as one piece of data in the data storage unit 101 .

[0078] Step 2: The over / under detection judgment unit 102 obtains the data of the inspection object from the data storage unit 101, and determines whether the data is within the over / under detection judgment threshold range. If so, the data is input into the over / under detection cause classifier 104; if not, the data is input into the automatic defect classifier 103 to perform defect identification and defect classification on the data.

[0079] The over / under detection determination unit 102 compares the data with a preset over / under detection threshold to screen out data that is or may be over / under detected, thereby further segmenting the data. The automatic defect classifier 103 performs defect detection and classification on the data that is not over / under detected. When all detected data is not over / under detected, the automatic defect classifier 103 completes the defect classification for all data. If there is a possibility of over / under detection, the corresponding data is input into the over / under detection cause classifier 104 to determine the cause of the over / under detection.

[0080] Step 3: The over-detection and missed detection reason classifier 104 classifies the input data into over-detection and missed detection reasons; wherein, the data belonging to the defect classification system reasons is input into the supplementary classifier 105 .

[0081] The over / under detection cause classifier 104 extracts the first feature of the data and determines which category of over / under detection cause the data belongs to based on the first feature space of the over / under detection cause classifier 104. When the data belongs to the defect classification system cause, the data needs to be input into the supplementary classifier 105. When the data belongs to the detection system cause, the engineer can overcome the fault by adjusting the detection conditions, such as changing the software and hardware parameters or initialization processing. After adjusting the detection conditions, the data enters the data acquisition unit 100 again. Among them, the data belonging to the defect classification system cause is further defect-classified by the supplementary classifier 105.

[0082] Step 4: The supplementary classifier 105 classifies the input data into defect categories.

[0083] If the data is due to a defect classification system error, indicating that the machine's automatic defect classifier has missed a detection error, the original automatic defect classifier is unable to identify and classify the data. Therefore, the data is input into the supplementary classifier 105, which extracts the data's second feature and, based on the supplementary classifier's second feature space, determines which defect category the data belongs to, thereby achieving defect identification and classification for the data.

[0084] Therefore, the automatic defect classification method can not only perform preventive screening for over-detection and missed detection, but also determine the causes of over-detection and missed detection, and accurately classify defects in over-detection and missed detection data caused by the automatic defect classifier itself, thereby improving detection accuracy and machine stability.

[0085] In summary, the automatic defect classification system and its training and classification methods provided in this embodiment utilize the offline trained over / under detection reason classifier 104 and supplementary classifier 105 to perform separate fault analysis and defect classification on over / under detection data, as well as defect classification on over / under detection data caused by the automatic defect classifier 103 itself, without interfering with the normal operation of the automatic defect classifier 103. Furthermore, the over / under detection determination unit 102 enables pre-screening of the inspection object, extracting data that may be over / under detection in advance to facilitate subsequent separate defect classification of the data, thereby further improving the accuracy of defect classification of the inspection object and the detection accuracy of the inspection equipment.

[0086] Furthermore, it should be recognized that although the present invention has been disclosed above with reference to preferred embodiments, the above embodiments are not intended to limit the present invention. Any person skilled in the art can utilize the above disclosed technical content to make many possible changes and modifications to the technical solution of the present invention, or modify it into equivalent embodiments with equivalent variations, without departing from the scope of the technical solution of the present invention. Therefore, any simple modifications, equivalent variations, and modifications 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, shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A training method for an automatic defect classification system, characterized in that: include: According to the pre-acquired defect classification database, the passed inspection data and missed inspection data are collected to form a passed and missed inspection data set; Marking the data in the over-detection and missed detection data set with the reasons for over-detection and missed detection; wherein the reasons for over-detection and missed detection include at least detection system reasons and defect classification system reasons; Performing a first feature extraction on the data to train a missed detection reason classifier, wherein the missed detection reason classifier is used to classify missed detection reasons of the data during the detection process; At least taking the data belonging to the cause of the defect classification system as a sample, and labeling the sample with a defect category; A second feature is extracted from the sample of the marked defect category to train a supplementary classifier, which is used to reclassify the data classified as defect classification system reasons by the over-detection and missed detection reason classifier during the detection process.

2. The training method of the automatic defect classification system according to claim 1, characterized in that: Performing a first feature extraction on the data to train a classifier for causes of over-detection and missed-detection, including: The first feature is extracted from the data to form a first feature vector; wherein the first feature includes multiple feature values, and the types of the feature values ​​in the first feature extracted from each data are the same; multiple first feature vectors formed by multiple data constitute a first feature space, and the first feature vector is matched with the over / miss detection reasons annotated by the corresponding data to train and form the over / miss detection reason classifier.

3. The training method of the automatic defect classification system according to claim 1, characterized in that: The training of the missed detection reason classifier includes: A plurality of classification algorithms are selected for training respectively, and the classification algorithm that meets the set requirements is selected from the plurality of evaluation results obtained as the classification algorithm of the over / under detection reason classifier.

4. The training method of the automatic defect classification system according to claim 1, characterized in that: Extracting a second feature from the sample of the labeled defect category to train a supplementary classifier includes: The second feature is extracted from the sample to form a second feature vector; wherein the second feature includes multiple feature values, and the feature values ​​in the second feature extracted from each sample are of the same type; the multiple second feature vectors formed by the multiple samples constitute a second feature space, and the second feature vectors are matched with the defect categories marked by the corresponding samples to train and form the supplementary classifier.

5. The training method of the automatic defect classification system according to claim 2 or 4, characterized in that: The characteristic values ​​include: testing machine, testing parameters, software attributes, test object attributes, test object process information, test object defect information or test object defect classification results.

6. The training method of the automatic defect classification system according to claim 5, characterized in that: The feature value in the second feature further includes: defect grayscale information and / or detection optical mode.

7. The training method of the automatic defect classification system according to claim 4, characterized in that: The training to form the supplementary classifier further includes: selecting a plurality of classification algorithms for training respectively, and selecting the classification algorithm that meets the set requirements from the plurality of evaluation results obtained as the classification algorithm of the supplementary classifier.

8. The training method of the automatic defect classification system according to claim 3 or 7, characterized in that: The classification algorithms include: K-nearest neighbor method, decision tree, random forest ensemble learning, neural network or support vector machine.

9. The training method of the automatic defect classification system according to claim 7, characterized in that: The classification algorithm used by the pre-acquired defect classification database is different from the classification algorithm of the supplementary classifier; and\or, the types of the feature values ​​in at least part of the second features are different from the types of feature values ​​selected by the pre-acquired defect classification database.

10. The training method of the automatic defect classification system according to claim 1, characterized in that: The samples used to train the supplementary classifier include the missed detection dataset.

11. The training method of the automatic defect classification system according to claim 1, characterized in that: The samples used to train the supplementary classifier include data of normal test objects.

12. An automatic defect classification system, characterized in that: include: Data acquisition unit, data storage unit, over / miss detection judgment unit, automatic defect classifier, over / miss detection cause classifier and supplementary classifier; The data acquisition unit is used to acquire data of the detected object; The data storage unit is used to store data acquired and generated by the data acquisition unit, the automatic defect classifier, the over / under detection determination unit, the over / under detection cause classifier, and the supplementary classifier; The over-detection / missed detection determination unit is used to perform similarity comparison on the data of the detected object to determine whether the data is over-detected data or missed detection data; The automatic defect classifier is used to perform defect recognition and defect classification on data that does not belong to the passed-inspection data or the missed-inspection data; The over-detection and missed detection reason classifier is used to classify the over-detection data or the missed detection data according to the over-detection and missed detection reasons; wherein the over-detection and missed detection reasons at least include detection system reasons or defect classification system reasons; The supplementary classifier is at least used to classify the data belonging to the defect classification system into defect categories.

13. The automatic defect classification system according to claim 12, characterized in that: The similarity comparison method includes Euclidean distance, Manhattan distance, Chebyshev distance, Minkowski distance or Pearson correlation coefficient.

14. An automatic defect classification method, characterized in that: Use of the automatic defect classification system according to any one of claims 12-13, comprising: The data acquisition unit acquires the data of the detected object and stores it in the data storage unit; The over / under detection determination unit obtains the data of the inspection object from the data storage unit, and determines whether the data is within the over / under detection determination threshold range. If yes, the data is input into the over / under detection cause classifier; if no, the data is input into the automatic defect classifier to perform defect identification and defect classification on the data. The over / miss detection reason classifier classifies the input data into over / miss detection reasons; wherein the data belonging to the defect classification system reasons is input into the supplementary classifier; The supplementary classifier classifies the input data into defect categories.

15. The automatic defect classification method according to claim 14, characterized in that: After the over / under detection reason classifier classifies the input data for over / under detection reasons, the data belonging to the detection system reasons is input to the data detection unit after the detection conditions are adjusted.

Images