TW202343613A - Detection formula configuration and optimization method and apparatus, electronic device and storage medium - Google Patents

Abstract

Provided in the present invention is a detection formula configuration and optimization method and apparatus, an electronic device and a storage medium. The method comprises: labeling a first data sample to obtain a second data sample; wherein the first data sample comprises a plurality of pieces of detection result data, and the second data sample comprises the detection result data and a label corresponding to each piece of data; according to the second data sample, obtaining data feature distribution information of a detection object; using a preset outlier statistical analysis strategy, performing outlier statistical analysis on the data feature distribution information, so as to obtain defect distribution boundary information and determine a detection formula; finally, according to the defect distribution boundary information and the preset outlier statistical analysis strategy, determining or optimizing the values of detection parameters of the detection formula by means of reverse derivation. In the invention, the coupling relationship between the parameters is considered, such that repeated adjustment of parameters can be avoided, and meanwhile, a whole set of detection parameters are inferred, so that rapid modeling of the detection formula is achieved; and manpower and time costs can be saved.

Description

Testing recipe setup and optimization methods, devices, electronic equipment and storage media

The invention relates to the field of semiconductor technology, and in particular to a detection recipe setting and optimization method, device, electronic equipment and storage medium.

In the manufacturing process of semiconductor wafers, wafer warpage (Bow) and wafer surface morphology are key parameters that affect process stability and product yield, and are critical to wafer yield (Yield). influence. For example, after the wafer undergoes different processes such as etching or thin film deposition, the wafer will warp to varying degrees or the wafer surface will be uneven; another example is that a robot may scratch the wafer during the manufacturing process of semiconductor integrated circuits. . Therefore, wafer defects are what all chip manufacturers pay most attention to during yield inspection. Once a wafer is defective, it is difficult to remedy it through subsequent processes. Therefore, it is crucial to quickly and accurately detect defects on the wafer surface to avoid wasting production resources due to defective products flowing into the next process.

In the existing technology, the wafer defect detection process usually uses forward process parameter adjustment. However, due to the diversity of on-site processes, a large amount of information needs to be generated each time, and coupled with the lack of prior knowledge, the detection parameters of the detection process are usually adjusted one by one. , since the coupling relationship between parameters cannot be taken into account, repeated adjustments of a single parameter may lead to deviations in the parameter adjustment results. In order to achieve better detection results, the detection formula needs to repeatedly adjust parameters, which brings manpower and time costs. increase. Moreover, due to the diversity of processes, existing detection formulas are difficult to apply to defect detection in new processes. Adjusting the parameters of the detection formula requires a certain algorithm background, so the requirements for users are high.

It should be noted that the information disclosed in the background technology section of this invention is only intended to deepen the understanding of the general background technology of the invention, and should not be regarded as an admission or any form of implication that the information constitutes a component that is already known to those skilled in the art. existing technology.

The purpose of the present invention is to provide a detection formula setting and optimization method, system, electronic equipment and storage medium in view of the defects existing in the prior art. The detection formula setting and optimization method provided by the invention is based on the prior knowledge of the detection result data. And fully consider the coupling relationship between parameters to determine the strategy and parameter setting values of the detection formula at one time, which not only ensures high efficiency of the detection process, but also improves the detection accuracy of the detection formula.

In order to achieve the above objectives, the present invention provides a detection formula setting and optimization method, a detection formula setting and optimization method, including:

Annotate the first data sample to obtain a second data sample; wherein, the first data sample includes several pieces of detection result data; the second data sample includes the detection result data and the corresponding data of each piece of detection result data. label;

According to the second data sample, obtain the data characteristic distribution information of the detection object;

Using a preset outlier statistical analysis strategy, perform outlier statistical analysis on the data feature distribution information, obtain defect distribution boundary information, and determine the detection formula according to the preset outlier statistical analysis strategy;

According to the defect distribution boundary information and the preset outlier statistical analysis strategy, the values of the detection parameters of the detection formula are set or optimized through reverse derivation.

Optionally, the detection result data includes basic information and characteristic data information of the detection object; wherein the characteristic data information includes position information of the detection result on the detection object, and the process flow of the detection object. Information, one or more of the grayscale information, shape information and texture information of the data information of the detection results;

Annotating the first data sample to obtain the second data sample includes:

Obtain the basic information of the detection object corresponding to each piece of detection result data in the first data sample;

For each piece of detection result data, obtain the original information corresponding to the detection result data on the detection object based on the basic information of the detection object and the location information of the detection result on the detection object;

According to the original information, it is judged whether the defect marked by the data information of the detection result is a real defect. If so, the detection result data is marked as true defect data; if not, the detection result data is marked as Noisy data;

The second data sample is obtained based on all the detection result data and the label corresponding to each piece of detection result data.

Optionally, the detection object includes a Wafer; the basic information of the Wafer includes the number of the Wafer, the number of Dies it contains, and the basic information of each Die; the basic information of the Die includes the Die number and the Die number of the Die. image information;

Obtaining the original information corresponding to the detection result data on the detection object based on the basic information of the detection object and the location information of the detection result on the detection object includes:

According to the basic information of the Wafer, obtain the Die number of each Die of the Wafer and the basic information of each Die;

According to the position information of the detection result on the Die and the image information of the Die, the image information of the detection result corresponding to the detection result data on the Die is obtained.

Optionally, obtaining the data feature distribution information of the detection object according to the second data sample includes:

Determine the characteristic data axis and the segmentation data axis, and establish a feature space based on the characteristic data axis and the segmentation data axis; wherein the characteristic data axis represents the characteristic data information of the detection result data, and the segmentation data axis represents the segmentation feature Information; wherein the segmentation feature information includes other feature data information in addition to the feature data axis;

Arrange the second data samples according to the feature space to obtain data feature distribution information of the detection object.

Optionally, the feature space includes one or more feature data axes and one or more segmentation data axes.

Optionally, the second data samples are arranged according to the feature space to obtain the data feature distribution information of the detection object, including:

Use the feature data axis as the horizontal axis and the segmented data axis as the vertical axis to establish a rectangular coordinate system;

In the rectangular coordinate system, in the horizontal axis direction, the feature value size of the feature data information represented by the feature data axis, and in the vertical axis direction, according to the feature data represented by the segmented data axis. The second data samples are arranged according to the characteristic value size of the information to obtain a defect characteristic distribution map.

Optionally, using a preset outlier statistical analysis strategy to perform outlier statistical analysis on the data feature distribution information to obtain defect distribution boundary information includes:

Determine whether to automatically find the defect distribution boundary information. If so, train the outlier statistical analysis model according to the selected outlier statistical analysis model to obtain the defect distribution boundary information; if not, use the data segmentation method to analyze the data. Conduct outlier statistical analysis on feature distribution information to obtain defect distribution boundary information;

Wherein, training the outlier statistical analysis model includes: training the selected outlier statistical analysis model according to the detection result data and the data feature distribution information until the detection result is obtained. The defect distribution boundary information of the object satisfies the first preset condition;

The use of data segmentation method to perform outlier statistical analysis on the data feature distribution information includes: based on the detection result data and the data feature distribution information, on the feature data axis and/or the segmented data axis Obtain at least one first segmentation threshold; and acquire the defect boundary information according to the first segmentation threshold until the obtained defect distribution boundary information of the detection object satisfies the second preset condition.

Optionally, the segmented data axis represents process flow information; and based on the detection result data and the data feature distribution information, threshold segmentation is performed on the characteristic data axis and/or the segmented data axis until the The defect distribution boundary information of the detection object satisfies the second preset condition, including:

Determine the first segmentation threshold of the segmented data axis based on the data feature distribution information and the consistency of the data distribution of the detection results labeled as true defect data and labeled as noise data;

Determine the second segmentation threshold of the feature data axis based on the data feature distribution information and the consistency of the data distribution of the detection results labeled as true defect data and labeled as noise data;

According to the first segmentation threshold of the segmentation data axis and the second segmentation threshold of the feature data axis, the defect distribution boundary information of the detection object is obtained.

Optionally, the use of preset outlier statistical analysis strategies also includes: an outlier statistical analysis strategy that combines data segmentation and model learning;

The outlier statistical analysis strategy that combines data segmentation and model learning includes: obtaining at least one first segmentation threshold on the segmentation data axis of the detection result data labeled as a true defect based on the data feature distribution information. ; And according to the first segmentation threshold and the data feature distribution information, train the selected outlier statistical analysis model until the obtained defect distribution boundary information of the detection object meets the third preset condition.

Optionally, setting or optimizing the values of detection parameters of the detection formula through reverse derivation based on the defect distribution boundary information and the preset outlier statistical analysis strategy includes:

Determine a reverse derivation strategy according to the preset outlier statistical analysis strategy;

According to the reverse derivation strategy, determine the input data information of the reverse derivation strategy;

Determine the data distribution model of the detection result data according to the input data information;

Determine the detection parameters of the detection formula according to the data distribution model and the defect distribution boundary information;

According to the strategy of the detection recipe and the input data information of the reverse derivation, the values of the detection parameters of the detection recipe are set or optimized.

Optionally, the preset outlier statistical analysis strategy is a data segmentation method;

According to the data segmentation method, the data distribution density of the detection result data of the detection object is counted as the reverse derivation strategy;

According to the reverse derivation strategy of the statistical data distribution density, all detection result data of the detection object are used as the input data information;

According to all detection result data, it is assumed that the data distribution density of the characteristic data information of all the detection result data in the feature space is divided into normal areas, noise areas and true defect areas; the normal area is a data distribution density greater than The area of the first density threshold, the noise area is the area where the data density is less than or equal to the first density threshold and greater than the second density threshold, and the true defect area is the area where the data density is less than or equal to the second density threshold;

Calculate the first density threshold and the second density threshold according to all detection result data and labels of all detection result data; wherein the first density threshold is greater than the second density threshold;

Calculate the displacement parameter of the detection formula according to the first density threshold, the second density threshold and the defect distribution boundary information.

Optionally, the preset outlier statistical analysis strategy is an outlier statistical analysis strategy based on Gaussian model;

According to the outlier statistical analysis strategy based on the Gaussian model, the Gaussian distribution of the detection result data of the detection object is obtained as the reverse derivation strategy, and Gaussian model detection is used as the detection formula strategy;

According to the reverse derivation strategy of statistical Gaussian distribution, all detection result data of the detection object are used as the input data information and the defect distribution boundary information is used as the input data information;

According to all detection result data, it is assumed that the data distribution density of the feature values of all the feature data information of the detection result data in the feature space obeys Gaussian distribution;

According to the input data information and the defect distribution boundary information, the parameters of the Gaussian model detection are determined.

Optionally, the preset outlier statistical analysis strategy is a machine learning outlier statistical analysis strategy;

According to the outlier statistical analysis strategy of machine learning, the density threshold and distance threshold for obtaining the detection result data of the detection object are used as the reverse derivation strategy, and the machine learning model is used as the strategy of detection formula;

According to the reverse derivation strategy of obtaining the density threshold and distance threshold of the detection result data of the detection object, the obtained density and distance of the detection result data of the detection object are used as the input data information;

Based on all inspection result data and the defect boundary distribution information, the density parameters and distance parameters of the inspection strategy of the machine learning model are reversely derived.

Optionally, the detection recipe setting and optimization method also includes:

According to the detection formula and the values of the detection parameters of the detection formula, defect analysis of the object to be detected is performed to obtain defect data information of the object to be detected.

In order to achieve the above object, the present invention also provides a detection formula setting and optimization device. The detection parameter and adjustment device includes:

The true defect and noise marking unit is configured to mark the first data sample to obtain a second data sample; wherein the first data sample includes several pieces of detection result data; the second data sample includes the detection result data Result data and labels corresponding to each test result data;

The feature distribution information acquisition unit is configured to obtain the data feature distribution information of the detection object based on the second data sample;

The defect distribution boundary acquisition unit is configured to use a preset outlier statistical analysis strategy to perform outlier statistical analysis on the data feature distribution information, obtain defect distribution boundary information, and is used to perform outlier statistical analysis according to the preset outlier statistical analysis strategy. Determine the test formula;

The detection parameter setting and optimization unit is configured to determine or optimize the value of the detection parameter of the detection formula through reverse derivation based on the defect distribution boundary information and the preset outlier statistical analysis strategy.

Optionally, the detection recipe setting and optimization device also includes:

The detection recipe application unit is configured to perform defect analysis on the object to be detected according to the detection formula and the values of detection parameters of the detection formula, and obtain defect data information of the object to be detected.

In order to achieve the above object, the present invention also provides an electronic device, including a processor and a storage. A computer program is stored on the storage. When the computer program is executed by the processor, the detection formula described above is realized. Setup and optimization methods.

In order to achieve the above object, the present invention also provides a readable storage medium. A computer program is stored in the readable storage medium. When the computer program is executed by the processor, the above-mentioned detection recipe setting and optimization method is realized.

Compared with the existing technology, the detection recipe setting and optimization method, device, electronic equipment and storage medium provided by the present invention have the following advantages:

The detection recipe setting and optimization method provided by the present invention first obtains a second data sample by annotating the first data sample; wherein the first data sample includes several pieces of detection result data; the second data sample includes all The detection result data and the label corresponding to each of the detection result data; then obtain the data feature distribution information of the detection object according to the second data sample, and determine the detection formula according to the preset outlier statistical analysis strategy; Then, a preset outlier statistical analysis strategy is used to perform outlier statistical analysis on the data feature distribution information to obtain defect distribution boundary information; finally, based on the defect distribution boundary information and the preset outlier statistical analysis strategy, through reverse Through direct derivation, the detection formula determines or optimizes the detection parameters of the detection formula. Therefore, in the detection recipe setting and optimization method provided by the present invention, the first data sample includes several pieces of detection result data, and the detection result data includes auxiliary parameter adjustment information (such as the basic information and characteristic data of the detection object). Information, the characteristic data information includes but is not limited to the grayscale, shape, texture and other information of defects indicated by the detection results). Through data annotation, real defect data and noise data can be distinguished, which can effectively use historical information for subsequent data analysis and Inference provides an important basis for obtaining accurate prior knowledge, which can improve the detection accuracy of detection formulas. Furthermore, in the detection recipe setting and optimization method provided by the present invention, the detection recipe strategy and detection parameter values are obtained through reverse derivation based on the defect distribution boundary information and the preset outlier statistical analysis strategy. Therefore, the present invention can deduce a set of detection parameters at the same time (that is, adjust all parameters at the same time) through reverse derivation. The coupling relationship between parameters is also taken into account, realizing rapid modeling of detection formulas; avoiding repeated adjustments. Parameters can significantly save labor and time costs; moreover, for new process defect detection, users can set or optimize the strategy of the detection formula and the values of the detection parameters of the detection formula without having any algorithm foundation.

Since the detection recipe setting and optimization device, electronic equipment and storage medium provided by the present invention and the detection parameters and adjustment method provided by the present invention belong to the same inventive concept, the detection recipe setting and optimization device, electronic equipment and storage medium provided by the present invention It has all the advantages of the detection recipe setting and optimization method, which will not be described in detail here.

The detection recipe setting and optimization method, device, electronic equipment and storage medium proposed by the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation modes. The advantages and features of the present invention will become clearer from the following description. It should be noted that the drawings are in a very simplified form and use imprecise proportions, and are only used to conveniently and clearly assist in explaining the embodiments of the present invention. In order to make the objects, features and advantages of the present invention more apparent, please refer to the accompanying drawings. It should be noted that the structures, proportions, sizes, etc. shown in the drawings attached to this specification are only used to coordinate with the content disclosed in the specification for the understanding and reading of those familiar with this technology, and are not used to limit the implementation of the present invention. Conditions, any structural modifications, changes in proportions or adjustments in size should still fall within the technical content disclosed in the present invention, provided that they are the same as or similar to the effects that the present invention can produce and the purposes that can be achieved. within the scope that can be covered.

It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that these entities or operations are mutually exclusive. any such actual relationship or sequence exists between them. Furthermore, the terms "comprises," "comprises," or any other variations thereof are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that includes a list of elements includes not only those elements, but also those not expressly listed other elements, or elements inherent to the process, method, article or equipment. Without further limitation, an element defined by the statement "comprises a..." does not exclude the presence of additional identical elements in a process, method, article, or apparatus that includes the stated element.

One embodiment of the present invention provides a detection recipe setting and optimization method. Specifically, please refer to FIG. 1 , which schematically provides a flow chart of the detection recipe setting and optimization method provided by an embodiment of the present invention. As shown in Figure 1, the detection recipe setting and optimization method includes the following steps: S100: Annotate the first data sample to obtain a second data sample; wherein the first data sample includes several pieces of detection result data; the second data sample includes the detection result data and each of the detection results The label corresponding to the data; S200: Obtain the data feature distribution information of the detection object according to the second data sample; S300: Use a preset outlier statistical analysis strategy to perform outlier statistical analysis on the data feature distribution information, obtain defect distribution boundary information, and determine the detection formula according to the preset outlier statistical analysis strategy; S400: Based on the defect distribution boundary information and the preset outlier statistical analysis strategy, set or optimize the values of the detection parameters of the detection formula through reverse derivation.

Therefore, in the detection recipe setting and optimization method provided by the present invention, the first data sample includes several pieces of detection result data, and the detection result data includes a large amount of auxiliary parameter adjustment information (such as the basic information and characteristics of the detection object). Data information, the characteristic data information includes but is not limited to the grayscale, shape, texture and other information of the defects indicated by the detection results). Through data annotation, real defect data and noise data can be distinguished, and historical information can be effectively used for subsequent data analysis. And reasoning can provide an important basis for obtaining accurate prior knowledge, which can improve the detection accuracy of detection formulas. Furthermore, in the detection recipe setting and optimization method provided by the present invention, the detection recipe strategy and parameter setting values are obtained through reverse derivation based on the defect distribution boundary information and the preset outlier statistical analysis strategy. Therefore, the present invention can deduce a set of detection parameters at the same time (that is, adjust all parameters at the same time) through reverse derivation. The coupling relationship between parameters is also taken into account, realizing rapid modeling of the detection process and avoiding repeated adjustments. parameters, which can significantly save labor and time costs. Moreover, for new process defect detection, users can set or optimize the detection recipe strategy and detection parameter values without having any algorithm foundation.

It should be noted that the detection result data is the historical detection result data of the detection object. For example, when setting up the detection recipe strategy and detection parameters used in the defect detection process for the first time, you can first randomly or artificially select the detection recipe strategy and detection parameters to obtain a certain amount of defect detection data. The detection data is the detection result data (that is, the first sample data). When optimizing the strategy of the detection formula and the detection parameters, the detection result data (that is, the first sample data) includes all or part of the historical detection data of the detection formula to be optimized. For ease of understanding and explanation, the detection result data described below are historical detection data of wafer defects. Obviously, this is not a limitation of the present invention. In other embodiments, the detection recipe setting and optimization provided by the present invention The method can also be adapted to other detection formulas for initial detection of wafer defects, so no examples will be given one by one.

Preferably, in one of the preferred embodiments, the detection result data includes basic information and characteristic data information of the detection object; wherein the characteristic data information includes location information of the detection result on the detection object, and One or more of the process flow information of the detection object, the grayscale information, the shape information and the texture information of the data information of the detection result. As those skilled in the art can understand, it is obvious that the data information of the detection results must also include conclusion information (defective data or non-defective data) used to indicate the detection results. In order to facilitate understanding, specific examples of the data information of the detection results will be described below in comparison with the image information of the detection results. Here, the data information of the detection results will not be exemplified. It can be seen that the detection result data includes the basic information and characteristic data information of the detection object (such as the grayscale, shape, texture and other information of nuisance) and other auxiliary parameter adjustment information, and will be used in the subsequent drawing of the defect distribution map. The parameter reverse reasoning process is based on the detection result data. Therefore, the detection formula setting and optimization method provided by the present invention can improve the detection accuracy of the detection formula.

Preferably, in one of the implementations, please refer to Figure 2, which schematically shows a schematic flow chart of the data sample annotation method. As can be seen from Figure 2, in step S100, the first data sample is annotated to obtain the second data sample, including: S110: Obtain the basic information of the detection object corresponding to each piece of detection result data in the first data sample; S120: For each piece of detection result data, obtain the original information corresponding to the detection result data on the detection object based on the basic information of the detection object and the location information of the detection result on the detection object; S130: Based on the original information, determine whether the defect marked by the data information of the detection result is a true defect. If so, mark the detection result data as true defect data; if not, mark the detection result data as true defect data. Marked as noisy data; S140: Obtain the second data sample based on all the detection result data and the tag corresponding to each detection result data.

With such configuration, the detection recipe setting and optimization method provided by the present invention can accurately classify the real defect data and noise data (nusiance, noise interference) in the detection result data (historical data) by labeling the first data sample. Data) are accurately distinguished, thereby providing accurate prior knowledge for subsequent acquisition of data feature distribution information, and then obtaining defect distribution boundary information based on the data feature distribution information for further reverse derivation, thereby improving the detection accuracy of the detection formula.

It should be noted that those skilled in the art should be able to understand that the characteristic data information is all detection results of defect detection on the detection object, including defect data and non-defect data.

As one of the preferred examples of applying the detection recipe setting and optimization method provided by the present invention, the following description takes the detection object as a wafer (wafer) as an example. Obviously, the first data sample is the wafer. Historical test result data. More specifically, the basic information of the Wafer includes the number of the Wafer, the number of Dies (die) contained, and the basic information of each Die; the basic information of the Die includes the Die number and image information of the Die. . Correspondingly, in step S120, the original information corresponding to the detection result data on the detection object is obtained based on the basic information of the detection object and the location information of the defect on the detection object, including: S121: According to the basic information of the Wafer, obtain the Die number of each Die of the Wafer and the basic information of each Die; S122: According to the position information of the detection result on the Die and the image information of the Die, obtain the image information of the detection result corresponding to the detection result data on the Die.

In order to facilitate a more accurate understanding of the present invention, the data information of the detection results and the image information of the detection results are explained below. The data information of the detection results includes the detection results in the detection result data. Description of the image information of the result, and the image information of the detection result is the original image corresponding to the data information of the detection result on the detection object. In other words, the data information of the detection result includes the detection result. Data representation of the resulting image information. Still taking the wafer as the detection object, for example: if the defect is a texture defect, the data information of the detection result records the texture characteristics of the texture defect, such as the roughness of the texture, etc., and the data information of the detection result The image information is the original image corresponding to the texture defect. Therefore, according to the image information of the detection result, the detection result data corresponding to the image information of the detection result can be re-judged as true defect data. Or noisy data.

Specifically, please refer to FIG. 3 , which schematically illustrates one of the interface diagrams for defect marking of data samples provided by an embodiment of the present invention. As can be seen from Figure 3, there are three main functional areas on the defect annotation interface, including the Wafer display window area, the detection data list window area and the defect display area. Specifically, the Wafer display window area is used to graphically display the basic information of the Wafer, including but not limited to the position of each Die on the Wafer and the number of the Die. Below the Wafer display window area, the user can select the Die number to be marked for defects. According to the Die number selected by the user, the historical detection data results of the Die corresponding to the selected Die number will be refreshed in the detection data list window area. . Therefore, according to the list of test result data in the test data list window area, the user can select the test result data one by one, and the original information corresponding to the test result data (that is, all the test result data) will be displayed in the defect display area. The image information of the detection result is the image information indicated by the position information of the detection result on the Die). Therefore, according to various characteristics of the original information (texture, size, curvature, shape, etc.), it is possible to Through manual re-judgment or machine re-judgment, etc., it is further confirmed whether the defect indicated by the data information of the test result is a true defect. If so, the test result data is marked as true defect data (for example, it will be included in the test result). The label of the detection result data in the data list window area is marked as a true defect, and the value corresponding to the manual judgment column of whether it is a real defect is set to yes); if not, the detection result data is marked as noise data (such as Mark the label of the detection result data in the detection data list window area as a false defect, and set the value corresponding to the column of manual judgment whether it is a real defect to No). Repeat the above process, select each Die number in turn, and manually annotate each detection data result under the current Die, then you can complete the annotation of the entire Wafer detection data results, and so on, you can add the above-mentioned first Die number. Label one data sample to obtain a second data sample.

It should be noted that although manual labeling is used as an example to illustrate the labeling method of the first data sample, it is obvious that this is not a limitation of the present invention. In other implementations, machines can also be used to label the first data sample. Annotation may be performed by learning or other methods, which is not limited by the present invention. Furthermore, as mentioned above, although the detection recipe setting and optimization method provided by the present invention is explained by taking a wafer as an example as a detection object, as those skilled in the art can understand, this is only a preferred embodiment. To illustrate, but not to limit the present invention, in other embodiments, the detection object may also be other products besides wafers, including but not limited to lenses, display screens, 3D printing products, etc. Explain with examples one by one.

Preferably, in one of the exemplary implementations, in step S200, obtaining the data feature distribution information of the detection object based on the second data sample includes: S210: Determine the characteristic data axis and the segmented data axis, and establish a feature space based on the characteristic data axis and the segmented data axis; wherein the characteristic data axis represents the characteristic data information of the detection result data, and the segmented data axis represents Segmentation feature information; wherein the segmentation feature information includes other feature data information except for the feature data axis; S220: Arrange the second data samples according to the feature space to obtain data feature distribution information of the detection object.

So configured, the detection recipe setting and optimization method provided by the present invention arranges the second data samples through the feature space, and the purpose is to make the distribution of the detection result data in the feature space show a certain trend. , making the distinction between true defect data and noise data more obvious, so as to facilitate the acquisition of defect distribution boundary information.

Preferably, the feature space includes one or more feature data axes and one or more segmentation data axes.

In the detection formula and optimization method provided by the present invention, the feature space may include multiple feature data axes and multiple segmentation data axes, and the feature space may be a multi-dimensional feature space. For example, there are two feature data axes, one of which is used to represent the grayscale information of the defect, and the other is used to represent the texture information of the defect; one of the split data axes is used to represent the shape information of the defect, and the other is used to represent all the defects. Describe the size of the defect. Therefore, the detection formula and optimization method provided by the present invention refer to more characteristic information of the defects, thus laying a good foundation for further improving the detection accuracy of the detection formula. It should be noted that the above is only an illustrative description and not a limitation of the present invention. In practical applications, the characteristic data axis, the segmented data axis and their respective numbers should be reasonably selected according to actual needs.

Preferably, in one of the exemplary implementations, in step S220, the second data samples are arranged according to the feature space to obtain the data feature distribution information of the detection object, including: S221: Use the feature data axis as the horizontal axis and the segmented data axis as the vertical axis to establish a rectangular coordinate system; S222: In the rectangular coordinate system, in the horizontal axis direction, according to the characteristic value size of the characteristic data information represented by the characteristic data axis, and in the vertical axis direction, according to the characteristic value represented by the segmented data axis. The second data samples are arranged according to the characteristic value size of the characteristic data information to obtain a defect characteristic distribution map.

Specifically, please refer to Figure 4, which schematically shows an example diagram of the distribution of detection result data in a two-dimensional feature space of one specific example. As can be seen from Figure 4, in this example, the horizontal axis represents the feature data axis, and the vertical axis represents the two-dimensional data feature distribution map formed by dividing the data axis. That is, the abscissa of each point in the coordinate system represents the size of the feature value, and the ordinate represents the size of the corresponding segmentation feature value. In this way, the feature values of all detection result data constitute the entire feature distribution map.

It should be noted that, as mentioned above, although the above example takes two-dimensional feature space distribution as an example, in actual applications, the feature data axis and the segmentation data axis may be multi-dimensional. That is, multiple segmentation values can be selected for the segmented data axis to divide the detection result data (ie, the second sample data) into several different feature distributions.

Furthermore, the present invention does not limit the specific selection method of the feature space. In one embodiment, a feature selection algorithm can be used to select the feature data axis and the segmentation data axis to automatically select the feature space; in other embodiments, The feature data axis and segmentation data axis can also be selected manually, and the present invention does not impose any limitation on this. More specifically, the feature data axis can represent information such as color, texture, shape, size, etc., and the segmentation axis can be information such as a trained mean map.

Further, as one of the preferred embodiments, the criteria for selecting the feature space are: the segmented data axis can better distinguish different process areas, and the feature data axis can make true defect data and noise data (noise points), the ultimate goal is to make the distribution of detection result data in the feature space show a certain trend, making the distinction between real defects and noise points more obvious. For example, for the detection result data of wafer defects, if the shape in the feature data information is used as the feature data axis rather than the texture in the feature data information as the feature data axis, the detection result data can be better placed in the feature space. To make the distinction between real defects and noise data more obvious, the shape in the feature data information is used as the feature data axis instead of the texture in the feature data information as the feature data axis. It can be understood that the shape in the feature data information is no longer used as the segmentation data axis.

Preferably, in one of the preferred embodiments, please refer to FIG. 5 , which schematically provides a flow chart of a detection recipe setting and optimization method provided by an embodiment of the present invention. It can be seen from Figure 5 that in step S300, the preset outlier statistical analysis strategy is used to perform outlier statistical analysis on the data feature distribution information to obtain defect distribution boundary information, including:

Determine whether to automatically find the defect distribution boundary information. If so, train the outlier statistical analysis model according to the selected outlier statistical analysis model to obtain the defect distribution boundary information; if not, use the data segmentation method to analyze the data. Conduct outlier statistical analysis on feature distribution information to obtain defect distribution boundary information.

Specifically, please refer to FIG. 6 , which is a schematic diagram of defect distribution boundary information obtained by applying the outlier statistical analysis model provided by the present invention. In Figure 6, feature1 is the segmentation data axis, and feartrue2 is the feature data axis. It can be seen from Figure 6 that in this example, the defect distribution boundary information 3 is a curve. It can be seen that the detection formula setting and optimization method provided by the present invention determines the preset outlier statistical analysis strategy based on the detection result data and the data feature distribution information, and determines the preset outlier statistical analysis strategy based on the determined preset outlier. The group statistical analysis strategy performs outlier statistical analysis on the data feature distribution information to obtain defect distribution boundary information, which can enable the defect distribution boundary information to better separate the true defect data 2 and the noise data 1, that is, The defect distribution boundary information can reduce over-inspection problems as much as possible without causing missed defects, so as to filter out more noise data. This can ensure that the subsequent inspection formula determined by reverse derivation based on the defect distribution boundary information will not cause missed inspections or over-inspections, thereby improving the defect detection accuracy of the inspection process.

It should be noted that for the same second sample data using the same feature space, if different outlier statistical analysis strategies are adopted, the defect distribution boundary information obtained may be different. Therefore, the subsequent reverse The strategies for deriving and detecting recipes are closely related to the outlier statistical analysis strategy. For the same second sample data used in Figure 6, if the data segmentation method is used, the shape of the defect distribution boundary information is completely the same as that in Figure 6. Different, please refer to the description below for details. To avoid redundancy, we will not go into details here.

In order to facilitate understanding and explanation, the following description takes two-dimensional data distribution as an example. First, the outlier statistical analysis model is explained in detail, and then the data segmentation method is explained.

Specifically, training the outlier statistical analysis model includes: training the selected outlier statistical analysis model according to the detection result data and the data feature distribution information until the detection result is obtained. The defect distribution boundary information of the object satisfies the first preset condition.

More specifically, as those skilled in the art can understand, comprehensive analysis can be performed based on the detection result data and the data feature distribution information to select an outlier statistical analysis model. The outlier analysis statistical model includes but is not limited to Statistics-based outlier algorithms (such as the 3σ principle), distance and proximity-based clustering algorithms (such as K-means, etc.), density-based outlier algorithms (such as DBSCAN, etc.), tree-based outlier analysis algorithms (such as isolated forest, etc.). It should be noted that the choice of algorithm model is very critical. Different algorithm models mean different shapes of outlier boundaries. An optimal algorithm model can make the training of the data set neither underfitting nor outliers occur. Overfitting. For example, if the distribution of the second sample data in the feature space is closer to a normal distribution, the outlier analysis statistical model is preferably based on a statistical outlier algorithm (such as the 3σ principle). For another example, if The distribution of the second sample data in the feature space is close to the true defect data and the noise data, but the distance between the defect data and the noise data is far, then the outlier analysis statistical model is preferably based on the distance sum Proximity clustering algorithm. Those skilled in the art should be able to draw inferences based on this and will not go into details here.

Further, those skilled in the art should be able to understand that the purpose of the outlier analysis statistical model is to find the optimal boundary result. After determining the outlier analysis statistical model, the second sample data should be used to analyze all selected The outlier analysis statistical model is trained. Through continuous learning and target optimization processes, the model training results can find the best inflection point of the segmented data axis and classify the true defects and noise data (interference data) according to the characteristic data axis. noise points) to distinguish. Therefore, after the training of the outlier analysis statistical model is completed, a boundary result (i.e., defect distribution boundary information) is obtained. Please refer to Figure 6. As shown in Figure 6, defect distribution boundary curve 3 (i.e., defect distribution boundary information) It can better separate defect data and noise data to ensure that the detection results will not miss detection or cause over-inspection problems. That is, the first preset condition is that the defect distribution boundary information can distinguish the detection result data labeled as true defect data and the detection result data labeled as noise data in the second sample.

Further, using the data segmentation method to perform outlier statistical analysis on the data feature distribution information includes: based on the detection result data and the data feature distribution information, on the feature data axis and/or the segmented data At least one first segmentation threshold is obtained on the axis; and the defect boundary information is obtained according to the first segmentation threshold until the obtained defect distribution boundary information of the detection object satisfies the second preset condition.

As one of the preferred embodiments, the data segmentation method includes manually segmenting the feature space to obtain the first segmentation threshold. As those skilled in the art can understand, the present invention is not limited to the specific implementation of the data segmentation method. In other implementations, the first segmentation threshold can also be obtained through a data segmentation algorithm.

In order to facilitate understanding and explanation, the following takes two-dimensional data distribution and manual segmentation as an example to explain the data segmentation method as follows: S321: Determine the first segmentation threshold of the segmented data axis based on the data feature distribution information and the consistency of the data distribution of the detection results labeled as true defect data and labeled as noise data. S322: Determine the second segmentation threshold of the feature data axis based on the data feature distribution information and the consistency of the data distribution of the detection results labeled as true defect data and labeled as noise data; S323: Obtain the defect distribution boundary information of the detection object based on the first segmentation threshold of the segmentation data axis and the second segmentation threshold of the feature data axis.

Specifically, in step S321, the data feature distribution information is used as input, and the segmented data axis is segmented in this feature distribution map. The segmentation standard is the consistency of the detection result data distribution, and the data with consistent distribution is regarded as a Cluster, find the segmentation value between clusters, so that the data of different processes can be distinguished. The consistent distribution includes the distribution law of the characteristic data information of the detection result data, including but not limited to the distribution density in the characteristic space, the relative position relationship of the spatial points, etc., and is based on this to determine the segmentation axis and the characteristic axis threshold, such as , in one of the examples, two first segmentation thresholds segment_value1 and segment_value2 are set.

Correspondingly, in step S322, a second segmentation threshold is determined for the feature data axis in the feature distribution. Since the defect data points have been marked in the feature distribution, the principle of determining the second segmentation threshold is to separate the noise data and the real defect data as far as possible, so as to ensure that the detection result data will not be missed at the same time. Also minimize the occurrence of over-inspections. That is, the second preset condition is preferably that the defect boundary information can separate the true defect data and the noise data.

Therefore, after the first segmentation threshold of the segmentation data axis and the second segmentation threshold of the feature data axis are determined respectively, the defect separation boundary information of the outlier statistical analysis can be obtained. The figure below still takes the two-dimensional feature data distribution as an example to display the manually segmented defect distribution boundary information. The detection result data is segmented using two first segmentation thresholds segment_value1 and segment_value2 on the segmentation axis, and all the detection result data is divided into three different distributions. In each segmentation threshold interval, three different second segmentation thresholds are used on the feature data axis to distinguish true defects from noise data, and the final defect distribution boundary information is obtained. That is, the defect distribution boundary information includes two straight lines parallel to the feature data axis featureu1 formed by the two first segmentation thresholds segment_value1 and segment_value2, and are respectively located on the feature data axis featreu1 and the first segmentation thresholds segment_value1 and segment_value2. between the first line segment that intersects the feature data axis featreu1 and the first segmentation threshold segment_value1, the second line segment that intersects the first segmentation thresholds segment_value1 and segment_value2, and the first segmentation threshold segment_value2 A third straight line that intersects and extends upward along the segmented data axis feature2.

Preferably, in one of the exemplary implementations, the use of a preset outlier statistical analysis strategy further includes: an outlier statistical analysis strategy that combines data segmentation and model learning. The outlier statistical analysis strategy that combines data segmentation and model learning includes: obtaining at least one first segmentation threshold on the segmentation data axis of the detection result data labeled as a true defect based on the data feature distribution information. ; And according to the first segmentation threshold and the data feature distribution information, train the selected outlier statistical analysis model until the obtained defect distribution boundary information of the detection object meets the third preset condition.

With such configuration, the detection recipe setting and optimization method provided by the present invention can further reduce the uncertainty of machine learning model training through an outlier statistical analysis strategy that combines data segmentation and model learning when obtaining outlier distribution boundary information. The input of the machine learning model has certain constraints, and the results of manual segmentation are used as constraints, which can further improve the efficiency of obtaining defect boundary distribution information.

The third preset condition is preferably to ensure that the detection result data does not miss detection while also minimizing the occurrence of over-inspection. That is, the second preset condition is preferably that the defect boundary information can reduce the true defect to The data and the noise data are separated or the number of training times of the outlier statistical analysis model reaches a preset value.

As those skilled in the art can understand, unlike the data segmentation method, the defect distribution boundary information obtained by using the outlier statistical analysis strategy that combines data segmentation and model learning is different from the defect distribution boundary information obtained by the above-mentioned data segmentation method. The defect distribution boundary information obtained by the outlier statistical analysis strategy that combines data segmentation and model learning includes two straight lines parallel to the feature data axis featureure1 formed by the two first segmentation thresholds segment_value1 and segment_value2, and two straight lines located on the feature data respectively. The three intervals formed by axis featreu1, the first segmentation thresholds segment_value1 and segment_value2 are closed curves surrounding the true defect data. Due to the different outlier statistical analysis strategies used, the defect boundary distribution information obtained is completely different. However, it is obvious that no matter what outlier statistical analysis strategy is used, the defect boundary distribution information obtained can be compared with the detection result data. Accurately distinguish between true defect data and noise data. As mentioned above, based on this, the present invention does not limit the specific implementation of the outlier statistical analysis strategy.

In addition, for details on the data segmentation method and model learning in the outlier statistical analysis strategy that combines data segmentation and model learning, please refer to the detailed description of the data segmentation method and outlier statistical analysis model above. In order to avoid redundancy , will not be described in detail here.

Preferably, in one of the exemplary implementations, please refer to FIG. 7 , which schematically provides a detailed flow chart of step S400 in FIG. 1 . It can be seen from Figure 7 that in step S400, based on the defect distribution boundary information and the preset outlier statistical analysis strategy, through reverse derivation, the values of the detection parameters for setting or optimizing the detection formula are determined, including : S410: Determine the reverse derivation strategy according to the preset outlier statistical analysis strategy; S420: Determine the input data information of the reverse derivation strategy according to the reverse derivation strategy; S430: Determine the data distribution model of the detection result data according to the input data information; S440: Determine the detection parameters of the detection formula according to the data distribution model and the defect distribution boundary information; S450: Set or optimize the values of detection parameters of the detection recipe according to the strategy of the detection recipe and the input data information of the reverse derivation.

Therefore, compared with the forward parameter setting in the prior art, which adopts the adjustment of corresponding parameters based on the results of forward parameter adjustment feedback (maybe only one or two detection parameters are adjusted), the detection formula setting and optimization provided by the present invention Method, use reverse derivation to determine the detection recipe strategy, and use the defect boundary distribution information to reversely infer all parameter settings of the detection recipe (key parameters, such as data density, data sparse distance and/or tolerance range, etc.) , the coupling relationship between the parameters of the detection process is also taken into account, thereby avoiding the repeated parameter adjustment process; and the parameter adjustment process is based on the user's annotation results, the user does not need to have prior knowledge, and can automatically deduce a relatively accurate set of parameters. For the parameters of the detection process, all detection parameters are adjusted to the optimal level at one time, which not only improves the efficiency of parameter adjustment in the detection process, but also improves the detection accuracy of the detection formula.

More specifically, please refer to FIG. 8 , which schematically shows a specific example of reverse derivation using the detection recipe setting and optimization method provided by the present invention. It can be seen from Figure 8 that in the detection recipe setting and optimization method provided by the present invention, the outlier statistical analysis strategy, the reverse derivation strategy and the parameter setting values of the detection process are closely related: that is, the reverse derivation strategy The core of the strategy of directional derivation and detection of recipes is consistent with the outlier statistical analysis strategy of obtaining the defect boundary distribution information. For example, if the outlier segmentation method is used as the strategy for outlier statistical analysis, the basic principles of the strategy for reverse derivation and detection of recipes should also be consistent with the basic principles of the outlier segmentation method.

In order to facilitate the understanding of the present invention, the following uses the data segmentation method as the outlier statistical analysis strategy, the outlier statistical analysis strategy based on Gaussian model and the outlier statistical analysis strategy of machine learning as examples to perform reverse derivation to obtain the parameters of the detection formula. The process of setting values is explained in detail.

1. Data segmentation method, reversely deriving new data processes and parameter setting values

Before describing the specific steps of reverse derivation of the basic principles of the data segmentation method to obtain the detection formula and parameter setting values, the core idea of the method is explained as follows:

In order to facilitate understanding of the present invention, please refer to FIG. 9 , which schematically provides a schematic diagram of the data density distribution of one of the detection result data provided in an implementation manner of this embodiment. The basic idea of this method is to define the area where the density of the detection result data points (the characteristic value of the detection result data) in the feature distribution map is greater than the first threshold as a normal (normal) area, that is, the normal area is expressed as the sum of the data density related functions. Therefore, all data points (feature values of detection result data) whose data density data_density is greater than the first threshold are normal, then data density data_density is one of the detection parameters that requires reverse inference. Further, an area where the data density is less than or equal to the first threshold and greater than the second threshold is defined as a nuisance area. The detection parameters of the nuisance area indicate that the detection result data in this area contains noise, and these noises are Allowable errors (that is, noisy areas due to process errors and noise effects) are not defective data. That is to say, the noise area is considered to be a tolerance value (displacement parameter) added to the normal area to describe the noise area, expressed by the following formula: nuisance_threshold = f1(data_density)                 (1)

Since the true defect data has been annotated in the outlier statistical analysis (that is, the boundaries between the noise area and the true defect area have been obtained), the displacement parameters (tolerance values) can be reversed based on the defect boundary distribution information. inferred. Locate the area where the data density is less than or equal to the second threshold (that is, outside the noise area) as the true defect area. Specifically, it can be expressed by the following formula: boundary_threshold = f2(inspection_data)                 (2) defect_threshold = f3(boundary_threshold)                 (3) offset_parameter = abs (defect_threshold - nuisance_threshold) (4)

In the formula, boundary threshold is the defect distribution boundary result obtained by the outlier statistical analysis algorithm, defect_threshold is a function related to the defect distribution boundary boundary_threshold, and finally the displacement parameter offset_parameter can be calculated using defect_threshold and nuisance_threshold.

According to the above analysis, as one of the preferred implementation methods, if the preset outlier statistical analysis strategy is the data segmentation method, the displacement parameters of the detection formula are obtained through the following steps: Step A1: According to the data segmentation method, count the data distribution density of the detection result data of the detection object as the reverse derivation strategy. Step A2: According to the reverse derivation strategy of the statistical data distribution density, use all detection result data of the detection object as the input data information. Step A3: Based on all the detection result data, it is assumed that the data distribution density of the characteristic data information of all the detection result data in the feature space is divided into normal areas, noise areas and true defect areas; the normal area is the data The area where the distribution density is greater than the first density threshold. The noise area is the area where the data density is less than or equal to the first density threshold and greater than the second density threshold. The true defect area is the area where the data density is less than or equal to the second density threshold. area. Step A4: Calculate the first density threshold and the second density threshold according to all detection result data and the labels of all detection result data; wherein the first density threshold is greater than the second density threshold; Step A5: Calculate the displacement parameter of the detection formula according to the first density threshold, the second density threshold and the defect distribution boundary information.

More specifically, in order to understand the present invention more clearly, next, taking wafer macro-defect detection as an example, the data segmentation method is used to obtain defect boundary distribution information, and reverse derivation is performed to obtain parameter setting values of the inspection process.

Referring to FIG. 10 , a schematic diagram of true defect distribution within the average gray level range of the standard segmentation axis provided by an implementation of this embodiment is provided. As shown in Figure 10, it is assumed that there are defects in each gray level range of the standard segmentation axis (that is, the segmentation data axis, corresponding to the vertical axis Feature2 in the figure), and the true defect data and noise data are marked. Wherein, the standard dividing axis is an average value graph generated by the statistics of N (N can be set according to actual needs, such as N=10 in Figure 10, the present invention is not limited to this) standard defect-free process data graphs, that is, using The average value of the corresponding pixel grayscales of N standard images is used as the final result. Specifically, please refer to Figure 11(a)-Figure 11(c) and Figure 12. Figure 11(a) is an example of multiple test charts provided by an embodiment of the present invention, and Figure 11(b) is a diagram. Figure 11(a) is an example of the average value chart generated by multiple test charts. Figure 11(c) is an example of the standard deviation chart generated by multiple test charts in Figure 11(a). Figure 12 is provided for applying the present invention. Schematic diagram of grayscale dynamic threshold. In the figure, pixel A is a pixel in the test image, and pixel A1 and A2 are the corresponding pixels of pixel A in the mean map and standard deviation map respectively.

a. Selected samples: Select samples (as shown in Figure 11(a)), and obtain the average value map and the standard deviation map based on the statistics (training) of N test charts.

b. Determine the feature data axis and segmentation data axis. See Figure 10. Feature1 is the feature data axis in Figure 10, and fearture2 is the segmentation data axis in Figure 10. According to the gray value of the test image and the gray value of the segmented data axis, the value of each test image on the feature data axis feature1 is calculated, as shown in the following formula: feature1=test-mean (5)

In the formula, feature1 is the feature data axis in Figure 10, test is the gray value of the test image, and mean is the gray value of the average image obtained by statistics of N test images.

As mentioned before, the segmented data axis feature2 is obtained by the following formula: feature2=mean                                                                                                                                        

In the formula, mean is the gray value of the average image obtained by statistics of N test images.

c. It is assumed that the defect distribution boundary information (threshold) is expressed by the following formula: defect_threshold=mean+/-（sigma*std+gray） （7）

In the formula, mean is the gray value of the average image obtained by statistics of N test images, std is the standard deviation corresponding to one of the pixels in the test image, sigma is the coefficient of the standard deviation, and is the parameter to be solved , gray is the dynamic threshold. The dynamic threshold gray is equivalent to the displacement parameter offset_parameter mentioned above, which can be defined as any curve. There is the following relationship between it and the average value of gray value: gray=b+a1*mean + a2*mean^2 + a3*mean^3+……+am*mean^m (8)

Among them, when only the first two terms of the polynomial are taken, the dynamic threshold gray =b+a1*mean is in the form of a straight line, and when the subsequent polynomials are continued to be taken, it becomes a curve form. Put the corresponding values of multiple points into equation (7), and get:

From this, the problem of reverse inference detection parameters is transformed into the problem of using the least squares method to solve the optimal solution of the above equations, where the value of sigma is the coefficient of the variance std in the boundary threshold hypothesis formula, and [b a1 a2 ... an] are all coefficients in the above piecewise curve to be fitted.

d. Solve each coefficient in the polynomial

By solving the system of equations and obtaining the values of sigma and [b a1 a2 ... an] of each coefficient, all parameters involved in the algorithm can be analyzed. As follows:

Convert the above system of equations into matrix form as follows: A x = b A’ A x= A’ b x =(A’ A )^(-1) * (A’ b)

Therefore, x is the final solution, and the vector can be obtained through the above matrix operation:

According to the above vector, the coefficient sigma of the standard deviation in the detection process can be obtained, as well as the multiple coefficients required for the dynamic threshold curve. From this, the curve of the dynamic threshold gray can also be obtained. Therefore, the true defect threshold of each pixel in the test image can be obtained by using the following formula in the inspection process: defect_threshold =std*sigma+gray                   (9)

That is, pixels greater than the above threshold defect_threshold are normal points, and pixels less than or equal to the threshold defect_threshold are defective points.

Specifically, please refer to Figure 11(d) and Figure 11(e). Figure 11(d) is an enlarged example of one of the test images, and Figure 11(e) is the defect location detected using a machine learning algorithm. Schematic diagram. By comparing Figure 11(d) and Figure 11(e), it is easy to find that the detection formula obtained by using the detection formula setting and optimization method provided by the present invention can accurately detect the true defects of the object to be detected.

2. Outlier statistical analysis strategy based on Gaussian model, reversely deriving new data processes and parameter setting values

In order to facilitate the understanding of the present invention, before specifically describing the reverse derivation based on outlier learning provided by the present invention to obtain new data processes and parameter setting values, the outlier statistical analysis strategy based on the Gaussian model is first used to reversely deduce new data. The core ideas of the process and parameter setting values are explained. The basic principle of this method is to assume that the distribution of all data points (detection result data) in the feature distribution map obeys Gaussian distribution. Then based on the defect boundary distribution information in the outlier statistical analysis, the parameters such as mean, variance and variance coefficient that need to be used in the detection model (strategy of the detection process) are reversely inferred to obtain the correlation required for Gaussian model detection. parameters. Similar to the data segmentation method and the process of reversely deriving new data processes and parameter settings, the Gaussian model-based outlier statistical analysis strategy to reversely derive new data processes and parameter settings includes the following steps: Step B1: The preset outlier statistical analysis strategy is an outlier statistical analysis strategy based on Gaussian model; Step B2: According to the outlier statistical analysis strategy based on the Gaussian model, use the Gaussian distribution of the detection result data of the detection object as the reverse derivation strategy, and use Gaussian model detection as the detection formula strategy; Step B3: According to the reverse derivation strategy of statistical Gaussian distribution, use all detection result data of the detection object as the input data information and the defect distribution boundary information as the input data information; Step B4: Based on all detection result data, assume that the data distribution density of the characteristic values of all the characteristic data information of the detection result data in the characteristic space obeys Gaussian distribution; Step B5: Determine the parameters of the Gaussian model detection based on the input data information and the defect distribution boundary information.

More specifically, it is expressed as follows through the following functional relationship expressions: boundary_threshold = f2(inspection_data)             (2) 𝜇=𝑓4(inspection_data )                                                                                             ∑ = f5(inspection_data,,𝜇)                                (11) ∏ = f6(boundary_threshold, inspection_data, 𝜇, ,∑) (12)

In the formula, boundary_threshold is the defect boundary distribution result obtained by the outlier algorithm. This boundary matrix boundary_threshold can already be obtained. The mean value 𝜇 can be obtained from the detection result data, which is obtained by averaging the grayscale of the current detection data image. The variance Σ is calculated by subtracting the sum of squares from the gray value of the pixels of the image to be detected and the mean 𝜇, and then averaging. The weight ∏ can be expressed as a function associated with boundary_threshold, inspection_data, 𝜇 and ∑, which is expressed as the coefficient of the variance ∑, according to the following formula: 𝜇 +∏ * ∑= boundary_threshold               (13)

In the above formula, since 𝜇, variance ∑ and boundary_threshold have been calculated, the equation can be solved to obtain the weight ∏.

3. Outlier statistical analysis strategy based on machine learning, reversely deriving new data processes and parameter setting values

Preferably, in one of the exemplary implementations, the outlier statistical analysis strategy based on machine learning to reversely derive new data processes and parameter setting values includes the following steps: Step C1: The preset outlier statistical analysis strategy is a machine learning outlier statistical analysis strategy; Step C2: According to the outlier statistical analysis strategy of machine learning, the density threshold and distance threshold for obtaining the detection result data of the detection object are used as the reverse derivation strategy, and the machine learning model is used as the detection formula strategy; Step C3: According to the reverse derivation strategy for obtaining the density threshold and distance threshold of the detection result data of the detection object, use the obtained density and distance of the detection result data of the detection object as the input data information; Step C4: Based on all detection result data and the defect boundary distribution information, reversely derive the density parameters and distance parameters of the detection strategy of the machine learning model.

As those skilled in the art can understand, since the machine learning model needs to formulate multiple parameters, the determination of the parameters of the outlier statistical analysis algorithm based on machine learning directly affects the detection accuracy. For example: the initial clustering center in the k-means algorithm, the neighborhood and number threshold in the DBSCAN algorithm, etc. Therefore, by conducting reverse reasoning on these machine learning parameters through the defect boundary distribution information (results) in outlier statistical analysis, a machine learning model with prior knowledge can be obtained, thereby improving the accuracy of model detection. Specifically, the following formulas can be used: boundary_threshold = f7(inspection_data) density_parameters = f8(boundary_threshold, inspection_data) distance_parameters = f9(boundary_threshold, inspection_data)

In the formula, boundary_threshold is the defect boundary distribution information obtained by the outlier algorithm, which is related to the detection result data and has been obtained during the defect boundary analysis process. The two important parameters of the clustering algorithm based on distance and density are density density_parameters and distance distance_parameters. Density density_parameters and distance distance_parameters are derived from the detection result data and the boundary matrix. By inverting the distance and density parameters, the defects are exactly located at the preset threshold. can be detected; while normal pixels are located within a threshold range with a larger density and are filtered out, thereby improving detection accuracy.

Preferably, in one of the exemplary implementations, please continue to refer to Figure 1. As can be seen from Figure 1, the detection recipe setting and optimization method also includes:

S500: Perform defect analysis on the object to be detected according to the detection formula and the values of the detection parameters of the detection formula to obtain defect data information of the object to be detected.

Please refer to Figure 13, which schematically shows a comparison diagram of the detection result data obtained by the detection process using the detection recipe setting and optimization method proposed by the present invention and the detection result data obtained by the original detection process. It can be seen from Figure 13 that by applying the strategy and parameter setting values of the detection process obtained by reverse derivation of the present invention for the detection process, the nuisance noise data is filtered out, the true defect data (defect defect data) is retained, and the detection result data is passed The distribution in the feature space can visually test the correctness of the results.

To sum up, in the detection recipe setting and optimization method provided by the present invention, the first data sample includes several pieces of detection result data, and the detection result data includes a large amount of auxiliary parameter adjustment information, which can be used effectively for subsequent use through data annotation. Historical information can be used for data analysis and reasoning to obtain accurate prior knowledge, which provides an important basis and can improve the detection accuracy of detection formulas. Furthermore, in the detection recipe setting and optimization method provided by the present invention, the detection recipe strategy and parameter setting values are obtained through reverse derivation based on the defect distribution boundary information and the preset outlier statistical analysis strategy. Therefore, the present invention can deduce a set of detection parameters at the same time (adjusting all parameters at the same time) through reverse derivation. The coupling relationship between parameters is also taken into account, realizing rapid modeling of the detection process; avoiding repeated adjustment of parameters. , which can significantly save labor and time costs; moreover, for new process defect detection, users can determine the strategy and parameter setting values of the detection process without the need for algorithm foundation.

Yet another embodiment of the present invention provides a detection recipe setting and optimization device. Specifically, please refer to FIG. 14 , which schematically provides a structural block diagram of the detection recipe setting and optimization device provided by this embodiment. As can be seen from Figure 14, the detection recipe setting and optimization device provided by this embodiment includes: a true defect and noise marking unit 100, a feature distribution information acquisition unit 200, a defect distribution boundary acquisition unit 300, and a detection parameter setting and optimization unit. 400.

Specifically, the true defect and noise marking unit 100 is configured to mark a first data sample to obtain a second data sample; wherein the first data sample includes several pieces of detection result data; and the second The data sample includes the detection result data and the label corresponding to each piece of the detection result data. The feature distribution information acquisition unit 200 is configured to obtain data feature distribution information of the detection object based on the second data sample. The defect distribution boundary acquisition unit 300 is configured to use a preset outlier statistical analysis strategy to perform outlier statistical analysis on the data feature distribution information, obtain defect distribution boundary information, and use it to perform outlier statistical analysis according to the preset outlier statistics. Analyze strategies and determine detection recipes. The detection parameter setting and optimization unit 400 is configured to set or optimize the values of detection parameters of the detection recipe through reverse derivation based on the defect distribution boundary information and the preset outlier statistical analysis strategy.

Preferably, as one of the exemplary implementations, the detection recipe setting and optimization device further includes a detection recipe application unit 500 . Specifically, the detection recipe application unit 500 is configured to perform defect analysis on the object to be detected according to the detection recipe and the values of detection parameters of the detection recipe, and obtain defect data information of the object to be detected.

Since the basic principles of the detection recipe setting and optimization device provided by the present invention are similar to the detection recipe setting and optimization methods provided by the above embodiments, in order to avoid redundancy, the specific content of the above detection recipe setting and optimization device implementation is introduced. It is relatively rough. For detailed information, please refer to the detailed description of the detection recipe settings and optimization methods above. Furthermore, since the detection recipe setting and optimization device provided by the present invention and the detection recipe setting and optimization method provided by the above embodiments belong to the same inventive concept, the detection recipe setting and optimization device provided by the present invention at least has the same features as the detection recipe setting and optimization method. The recipe setting and optimization method have the same beneficial effects. You can refer to the relevant content in the detection recipe setting and optimization method above, so this will not be described again. In addition, since the detection formula setting and optimization device in the present invention and the detection formula setting and optimization method described above belong to the same inventive concept, the introduction to the detection formula setting and optimization device in this article is relatively simple. Regarding how, you can Please refer to the relevant content in the detection recipe settings and optimization methods above, so we will not go into details here.

Based on the same inventive concept, the present invention also provides an electronic device. Please refer to FIG. 15 , which schematically shows a block structure diagram of the electronic device provided by an embodiment of the present invention. As shown in Figure 15, the electronic device includes a processor 601 and a storage 603. A computer program is stored on the storage 603. When the computer program is executed by the processor 601, the detection described above is realized. Recipe setting and optimization methods. Since the electronic device provided by the present invention belongs to the same inventive concept as the above-mentioned detection recipe setting and optimization method, it has all the advantages of the above-mentioned detection recipe setting and optimization method, and thus will not be described again.

As shown in FIG. 15 , the electronic device also includes a communication interface 602 and a communication bus 604 , wherein the processor 601 , the communication interface 602 , and the storage 603 complete communication with each other through the communication bus 604 . The communication bus 604 may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus 604 can be divided into an address bus, a data bus, a control bus, etc. For ease of presentation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus. The communication interface 602 is used for communication between the above electronic device and other devices.

The processor 601 in the present invention can be a central processing unit (CPU), or other general-purpose processor, a digital signal processor (Digital Signal Processor, DSP), or an application specific integrated circuit (Application Specific Integrated Circuit). ASIC), off-the-shelf programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general processor may be a microprocessor or the processor may be any conventional processor, etc. The processor 601 is the control center of the electronic device and uses various interfaces and lines to connect various parts of the entire electronic device.

The storage 603 can be used to store the computer program. The processor 601 implements the electronic device by running or executing the computer program stored in the storage 603 and calling the data stored in the storage 603. various functions.

The storage 603 may include non-volatile and/or volatile storage. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

The present invention also provides a readable storage medium. A computer program is stored in the readable storage medium. When the computer program is executed by a processor, the above-mentioned detection recipe setting and optimization method can be realized. Since the readable storage medium provided by the present invention and the detection recipe setting and optimization method described above belong to the same inventive concept, it has all the advantages of the detection recipe setting and optimization method described above, so this will not be discussed further. Repeat.

The readable storage medium in the embodiment of the present invention may be any combination of one or more computer-readable media. The readable medium may be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared or semiconductor system, device or device, or any combination thereof. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more conductors, a portable computer hard drive, a hard drive, random access memory (RAM), read only memory (ROM) ), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above. As used herein, a computer-readable storage medium may be any tangible medium that contains or stores a program for use by or in combination with an instruction execution system, device, or device.

A computer-readable signal medium may include a data signal that carries computer-readable program code in baseband or as part of a carrier wave. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, device, or device .

Computer code for performing operations of the present invention may be written in one or more programming languages, including object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming, or a combination thereof. Programming language - such as "C" or a similar programming language. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer, or entirely on the remote computer or server. In situations involving remote computers, the remote computer can be connected to the user's computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., using an Internet service provider to connect via the Internet).

To sum up, compared with the existing technology, the detection recipe setting and optimization method, device, electronic equipment and storage medium provided by the present invention have the following advantages: the first data sample includes several detection result data, and the detection result The data includes auxiliary parameter adjustment information (such as the basic information and characteristic data information of the detection object. The characteristic data information includes but is not limited to the grayscale, shape, texture and other information of the defects indicated by the detection results). Through data annotation, the data can be Distinguishing between true defect data and noise data provides an important basis for subsequent effective use of historical information for data analysis and reasoning to obtain accurate prior knowledge, which can improve the detection accuracy of detection formulas. Furthermore, in the detection recipe setting and optimization method provided by the present invention, the detection recipe strategy and detection parameter values are obtained through reverse derivation based on the defect distribution boundary information and the preset outlier statistical analysis strategy. Therefore, the present invention can deduce a set of detection parameters at the same time (adjusting all parameters at the same time) through reverse derivation. The coupling relationship between parameters is also taken into account, realizing rapid modeling of detection formulas; avoiding repeated adjustment of parameters. , which can significantly save labor and time costs; moreover, for new process defect detection, the user can determine the strategy of the detection formula and the values of the detection parameters of the detection formula without having any algorithm foundation.

It should be noted that the devices and methods disclosed in the embodiments of this article can also be implemented in other ways. The above-described device implementations are only illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate possible implementation architectures, functions, and implementations of devices, methods, and computer program products according to various embodiments of this document. operate. In this regard, each block in the flowchart or block diagram may represent a module, program, or portion of code that contains one or more logic for implementing the specified Functional executable instructions. The module, code segment or part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two consecutive blocks may actually execute substantially in parallel, or they may sometimes execute in the reverse order, depending on the functionality involved. It is also noted that each block in the block diagram and/or flowchart illustration, and combinations of blocks in the block diagram and/or flowchart illustration, can be configured to perform special purpose hardware-based hardware to perform the specified functions or actions. system, or can be implemented using a combination of dedicated hardware and computer instructions.

In addition, each functional module in each embodiment of this article can be integrated together to form an independent part, each module can exist alone, or two or more modules can be integrated to form an independent part.

The above description is only a description of the preferred embodiments of the present invention, and does not limit the scope of the present invention in any way. Any changes or modifications made by those of ordinary skill in the field of the present invention based on the above disclosure fall within the scope of the present invention. Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations are within the scope of the present invention and equivalent technologies, the present invention is also intended to include these modifications and variations.

S100~S500: steps 1: Noisy data 2: True defect data 3: Defect distribution boundary curve segment_value1, segment_value2: first segmentation threshold A, A1, A2: pixels 100: True defect and noise marking unit 200: Feature distribution information acquisition unit 300: Defect distribution boundary acquisition unit 400: Detection parameter setting and optimization unit 500: Detection formula application unit 601: Processor 602: Communication interface 603:Storage 604: Communication bus

Figure 1 is a schematic flow chart of a detection recipe setting and optimization method provided by an embodiment of the present invention; Figure 2 is a schematic flow chart of a data sample labeling method provided by an embodiment of the present invention; Figure 3 is a schematic diagram of an interface for defect marking of data samples provided by an embodiment of the present invention; Figure 4 is an example diagram showing the distribution of detection result data in a two-dimensional feature space in one specific example of applying the present invention; Figure 5 is a schematic diagram of the principle of outlier statistical analysis provided by an embodiment of the present invention; Figure 6 is a schematic diagram of defect distribution boundary information obtained by applying the outlier statistical analysis model provided by the present invention; Figure 7 is a detailed flow diagram of step S400 in Figure 1; Figure 8 is a specific example diagram of reverse derivation using the detection formula setting and optimization method provided by the present invention; Figure 9 is a schematic diagram of the data density distribution of one of the detection result data provided by an embodiment of the present invention; Figure 10 is a schematic diagram of true defect data distribution within the average gray level range of the standard segmentation axis provided by an embodiment of the present invention; Figure 11(a) is an example of multiple test charts provided by an embodiment of the present invention; Figure 11(b) is an example of the mean graph generated from multiple test images in Figure 11(a); Figure 11(c) is an example of the standard deviation chart generated from multiple test charts in Figure 11(a); Figure 11(d) is an enlarged example of one of the test images; Figure 11(e) is a schematic diagram of defect locations detected using machine learning recipes; Figure 12 is a schematic diagram of the grayscale dynamic threshold provided by the application of the present invention; Figure 13 is a schematic diagram comparing the detection result data obtained by applying the detection formula setting and optimization method provided by the present invention and the detection result data obtained by the original detection formula; Figure 14 is a structural block diagram of a detection recipe setting and optimization device in an embodiment of the present invention; FIG. 15 is a schematic block structure diagram of an electronic device in an embodiment of the present invention.

S100~S500: steps

Claims (18)

A detection recipe setting and optimization method, which is characterized by including: Annotate the first data sample to obtain a second data sample; wherein, the first data sample includes several pieces of detection result data; the second data sample includes the detection result data and the corresponding data of each piece of detection result data. label; According to the second data sample, obtain the data characteristic distribution information of the detection object; Using a preset outlier statistical analysis strategy, perform outlier statistical analysis on the data feature distribution information to obtain defect distribution boundary information; and determine the detection formula according to the preset outlier statistical analysis strategy; According to the defect distribution boundary information and the preset outlier statistical analysis strategy, the values of the detection parameters of the detection formula are set or optimized through reverse derivation. The detection recipe setting and optimization method as described in claim 1, wherein the detection result data includes basic information and characteristic data information of the detection object; wherein the characteristic data information includes the detection result on the detection object. Position information, as well as one or more of the process flow information of the detection object, the grayscale information, shape information and texture information of the detection result; Annotating the first data sample to obtain the second data sample includes: Obtain the basic information of the detection object corresponding to each piece of detection result data in the first data sample; For each piece of detection result data, obtain the original information corresponding to the detection result data on the detection object based on the basic information of the detection object and the location information of the detection result on the detection object; According to the original information, it is judged whether the defect marked by the data information of the detection result is a real defect. If so, the detection result data is marked as true defect data; if not, the detection result data is marked as Noisy data; The second data sample is obtained based on all the detection result data and the label corresponding to each piece of detection result data. The detection recipe setting and optimization method as described in claim 2, wherein the detection object includes a wafer; the basic information of the wafer includes the number of the wafer, the number of Dies included, and the basic information of each Die; The basic information of Die includes the Die number and image information of the Die; Obtaining the original information corresponding to the detection result data on the detection object based on the basic information of the detection object and the location information of the detection result on the detection object includes: According to the basic information of the Wafer, obtain the Die number of each Die of the Wafer and the basic information of each Die; According to the position information of the detection result on the Die and the image information of the Die, the image information of the detection result corresponding to the detection result data on the Die is obtained. The detection recipe setting and optimization method as described in claim 1, wherein the data feature distribution information of the detection object is obtained according to the second data sample, including: Determine the characteristic data axis and the segmentation data axis, and establish a feature space based on the characteristic data axis and the segmentation data axis; wherein the characteristic data axis represents the characteristic data information of the detection result data, and the segmentation data axis represents the segmentation feature Information; wherein the segmentation feature information includes other feature data information in addition to the feature data axis; Arrange the second data samples according to the feature space to obtain data feature distribution information of the detection object. The detection recipe setting and optimization method as described in claim 4, wherein the feature space includes one or more feature data axes and one or more segmentation data axes. The detection recipe setting and optimization method as described in claim 4, wherein the second data samples are arranged according to the feature space to obtain the data feature distribution information of the detection object, including: Use the feature data axis as the horizontal axis and the segmented data axis as the vertical axis to establish a rectangular coordinate system; In the rectangular coordinate system, in the horizontal axis direction, the feature value size of the feature data information represented by the feature data axis, and in the vertical axis direction, according to the feature data represented by the segmented data axis. The second data samples are arranged according to the characteristic value size of the information to obtain a defect characteristic distribution map. The detection recipe setting and optimization method as described in claim 6, wherein the preset outlier statistical analysis strategy is used to perform outlier statistical analysis on the data feature distribution information to obtain defect distribution boundary information, including: Determine whether to automatically find the defect distribution boundary information. If so, train the outlier statistical analysis model according to the selected outlier statistical analysis model to obtain the defect distribution boundary information; if not, use the data segmentation method to analyze the data. Conduct outlier statistical analysis on feature distribution information to obtain defect distribution boundary information; Wherein, training the outlier statistical analysis model includes: training the selected outlier statistical analysis model according to the detection result data and the data feature distribution information until the detection result is obtained. The defect distribution boundary information of the object satisfies the first preset condition; The use of data segmentation method to perform outlier statistical analysis on the data feature distribution information includes: based on the detection result data and the data feature distribution information, on the feature data axis and/or the segmented data axis Obtain at least one first segmentation threshold; and acquire the defect boundary information according to the first segmentation threshold until the obtained defect distribution boundary information of the detection object satisfies the second preset condition. The detection recipe setting and optimization method as described in claim 7, wherein the segmented data axis represents process flow information; and based on the detection result data and the data characteristic distribution information, the characteristic data axis and/or The segmented data axis is threshold segmented until the obtained defect distribution boundary information of the detection object meets the second preset condition, including: Determine the first segmentation threshold of the segmented data axis based on the data feature distribution information and the consistency of the data distribution of the detection results labeled as true defect data and labeled as noise data; Determine the second segmentation threshold of the feature data axis based on the data feature distribution information and the consistency of the data distribution of the detection results labeled as true defect data and labeled as noise data; According to the first segmentation threshold of the segmentation data axis and the second segmentation threshold of the feature data axis, the defect distribution boundary information of the detection object is obtained. The detection recipe setting and optimization method as described in claim 7, wherein the use of a preset outlier statistical analysis strategy also includes: an outlier statistical analysis strategy that combines data segmentation and model learning; The outlier statistical analysis strategy that combines data segmentation and model learning includes: obtaining at least one first segmentation threshold on the segmentation data axis of the detection result data labeled as a true defect based on the data feature distribution information. ; And according to the first segmentation threshold and the data feature distribution information, train the selected outlier statistical analysis model until the obtained defect distribution boundary information of the detection object meets the third preset condition. The detection recipe setting and optimization method as described in claim 1, wherein the detection parameters of the detection recipe are set or optimized through reverse derivation based on the defect distribution boundary information and the preset outlier statistical analysis strategy. The values include: Determine a reverse derivation strategy according to the preset outlier statistical analysis strategy; According to the reverse derivation strategy, determine the input data information of the reverse derivation strategy; Determine the data distribution model of the detection result data according to the input data information; Determine the detection parameters of the detection formula according to the data distribution model and the defect distribution boundary information; According to the strategy of the detection recipe and the input data information of the reverse derivation, the values of the detection parameters of the detection recipe are set or optimized. The detection recipe setting and optimization method as described in claim 10, wherein the preset outlier statistical analysis strategy is a data segmentation method; According to the data segmentation method, the data distribution density of the detection result data of the detection object is counted as the reverse derivation strategy; According to the reverse derivation strategy of the statistical data distribution density, all detection result data of the detection object are used as the input data information; According to all detection result data, it is assumed that the data distribution density of the characteristic data information of all the detection result data in the feature space is divided into normal areas, noise areas and true defect areas; the normal area is a data distribution density greater than The area of the first density threshold, the noise area is the area where the data density is less than or equal to the first density threshold and greater than the second density threshold, and the true defect area is the area where the data density is less than or equal to the second density threshold; Calculate the first density threshold and the second density threshold according to all detection result data and labels of all detection result data; wherein the first density threshold is greater than the second density threshold; Calculate the displacement parameter of the detection formula according to the first density threshold, the second density threshold and the defect distribution boundary information. The detection recipe setting and optimization method as described in claim 10, wherein the preset outlier statistical analysis strategy is an outlier statistical analysis strategy based on a Gaussian model; According to the outlier statistical analysis strategy based on the Gaussian model, the Gaussian distribution of the detection result data of the detection object is obtained as the reverse derivation strategy, and Gaussian model detection is used as the detection formula strategy; According to the reverse derivation strategy of statistical Gaussian distribution, all detection result data of the detection object are used as the input data information and the defect distribution boundary information is used as the input data information; According to all detection result data, it is assumed that the data distribution density of the feature values of all the feature data information of the detection result data in the feature space obeys Gaussian distribution; According to the input data information and the defect distribution boundary information, the parameters of the Gaussian model detection are determined. The detection recipe setting and optimization method as described in claim 10, wherein the preset outlier statistical analysis strategy is a machine learning outlier statistical analysis strategy; According to the outlier statistical analysis strategy of machine learning, the density threshold and distance threshold for obtaining the detection result data of the detection object are used as the reverse derivation strategy, and the machine learning model is used as the strategy of detection formula; According to the reverse derivation strategy of obtaining the density threshold and distance threshold of the detection result data of the detection object, the obtained density and distance of the detection result data of the detection object are used as the input data information; Based on all inspection result data and the defect boundary distribution information, the density parameters and distance parameters of the inspection strategy of the machine learning model are reversely derived. The detection recipe setting and optimization method as described in any one of requests 1 to 13, which also includes: According to the detection formula and the values of the detection parameters of the detection formula, defect analysis of the object to be detected is performed to obtain defect data information of the object to be detected. A detection formula setting and optimization device, which is characterized by including: The true defect and noise marking unit is configured to mark the first data sample to obtain a second data sample; wherein the first data sample includes several pieces of detection result data; the second data sample includes the detection result data Result data and labels corresponding to each test result data; The feature distribution information acquisition unit is configured to obtain the data feature distribution information of the detection object based on the second data sample; The defect distribution boundary acquisition unit is configured to use a preset outlier statistical analysis strategy to perform outlier statistical analysis on the data feature distribution information, obtain defect distribution boundary information, and is used to perform outlier statistical analysis according to the preset outlier statistical analysis strategy. Determine the test formula; The detection parameter setting and optimization unit is configured to set or optimize the values of detection parameters of the detection formula through reverse derivation based on the defect distribution boundary information and the preset outlier statistical analysis strategy. The detection recipe setting and optimization device as described in claim 15, which also includes: The detection recipe application unit is configured to perform defect analysis on the object to be detected according to the detection formula and the values of detection parameters of the detection formula, and obtain defect data information of the object to be detected. An electronic device, characterized in that it includes a processor and a storage, and a computer program is stored on the storage. When the computer program is executed by the processor, the method described in any one of claims 1 to 14 is realized. Test recipe settings and optimization methods. A readable storage medium, characterized in that a computer program is stored in the readable storage medium. When the computer program is executed by a processor, the detection recipe setting and optimization described in any one of claims 1 to 14 are realized. method.