Defective root cause analysis method and system based on relative feature importance

Abstract

The invention provides a bad root cause analysis method and system based on relative feature importance, and relates to the technical field of bad root cause analysis, and the method comprises the steps: obtaining the resume data and the reject ratio of a product, and carrying out the preprocessing of the resume data and the reject ratio of the product; performing format conversion and data splicing on the preprocessed resume data and the reject ratio to obtain training sample data; inputting the training sample data into the decision tree model to calculate real importance distribution and zero importance distribution of the sites; calculating the relative importance of the sites according to the change difference between the real importance distribution and the zero importance distribution of the sites; performing feature importance sorting on the sites according to the relative importance of the sites, and calculating the bad root cause of each site step by step according to a feature importance sorting result. The problems that an existing bad root cause analysis method is long in time consumption and high in cost, bad root cause analysis is inaccurate, and a root cause positioning result cannot be stably obtained are solved through relative feature importance analysis of sites.

Description

Technical Field

[0001] The present invention relates to the technical field of defect root cause analysis, and in particular to a defect root cause analysis method and system based on contribution. Background Art

[0002] In modern industrial production, the smooth operation of each production line site or equipment directly determines the overall production efficiency and the quality level of the final product. When product quality problems occur, it is crucial to trace their root causes. This tracing process is called root cause analysis, and its core lies in identifying and understanding the specific production line or equipment problems that cause quality problems.

[0003] Root cause analysis is not only a means to solve current problems, but also a key strategy to prevent future problems. By deeply analyzing production data, monitoring equipment operating status and collecting quality feedback, companies can quickly locate the source of problems and take effective measures to repair and improve them. This systematic approach not only helps to improve product quality and production efficiency, but also helps to optimize resource utilization and reduce production costs. By continuously optimizing the root cause analysis process, companies can establish a robust quality management system to ensure the production of products that meet high standards, thereby enhancing market competitiveness.

[0004] The process of the existing defect root cause analysis method is as follows: Step 1: Manually collect data on defective products, preliminarily locate the possible causes of the defect based on expert experience, and calculate the difference in yield between devices; Step 2: Analyze the differences between parallel devices and parallel chambers at each site through statistical methods, and then locate the root cause site; Step 3: Fit the equipment and parameters passed by the product through a machine learning algorithm, predict the product's defect rate, and obtain the root cause results based on feature importance. The existing defect root cause analysis method has the following defects when performing defect root cause analysis: ① It requires a lot of expert experience, is time-consuming and costly; ② The analysis results of statistical methods are inaccurate; ③ The feature importance ranking in the machine learning method is unstable, and the defect root cause results cannot be obtained stably. Summary of the invention

[0005] The present invention provides a method and system for analyzing root causes of defects based on relative feature importance, which solves the problems of the existing root cause analysis methods being time-consuming, costly, inaccurate, and unable to stably obtain root cause location results.

[0006] In a first aspect, an embodiment of the present invention provides a method for analyzing root causes of defects based on relative feature importance, the method comprising the following process:

[0007] Obtain product history data and defect rate, and pre-process the product history data and defect rate;

[0008] Perform format conversion and data splicing on the pre-processed resume data and defect rate to obtain training sample data;

[0009] A decision tree model is constructed, and the training sample data is input into the decision tree model to calculate the true importance distribution and zero importance distribution of the site;

[0010] The relative importance of the sites is calculated based on the difference between the true importance distribution of the sites and the change of the zero importance distribution;

[0011] The feature importance is ranked according to the relative importance of the sites, and the root cause of the badness of each site is calculated level by level based on the feature importance ranking results.

[0012] In the above embodiment, the present invention first calculates the real feature importance distribution and the zero feature importance distribution of each site, and then reorders the sites according to the relative difference between the two distributions, thereby improving the hit rate and stability of the sites in the top ranking results, thereby facilitating experts to quickly verify and take treatment measures to improve product yield; at the same time, the root cause device is split into first locating the root cause site and then calculating the device with a high defect rate, which reduces the difficulty of directly locating the root cause device and improves the hit rate of the root cause device; and the site is pre-screened in the data preprocessing stage, which reduces the difficulty of the model to directly find the root cause site from a large number of sites and improves the hit rate of the site; and the defect root cause analysis is highly automated, and the final defect root cause can be directly obtained from the input data through end-to-end modeling.

[0013] As some optional implementation modes of the present application, the preprocessing includes data type conversion, outlier removal, and site type screening.

[0014] In the above embodiment, in order to prevent the historical data and defective rate obtained in the fully automatic root cause analysis process from being affected by various problems such as missing data records and incorrect data import formats, it is necessary to uniformly preprocess the obtained historical data and defective rate.

[0015] As some optional implementations of the present application, the process of format conversion and data splicing of the pre-processed resume data and defect rate to obtain training sample data is as follows:

[0016] De-duplicate the pre-processed historical data according to product ID, site ID, and device ID.

[0017] The structure of the deduplicated historical data is converted according to the table structure of behavior samples, columns as sites, and values ​​as devices of each product under each site;

[0018] The structural conversion history data and defect rate are concatenated according to the product ID to obtain the training sample data.

[0019] In the above embodiment, the present invention performs format conversion and data splicing on the pre-processed historical data and defect rate, so that the acquired training sample data can be recognized by the decision tree model.

[0020] As some optional implementation methods of the present application, the process of calculating the real importance distribution of the site is as follows:

[0021] Set the number of random trainings of the decision tree model to n, and randomly obtain n training samples;

[0022] The randomly obtained n training sample data are input into the decision tree model to sort the feature importance to obtain the true importance distribution of the site.

[0023] In the above embodiment, since the true feature importance distribution reflects the approximate value range of the feature importance of a feature in multiple random trainings, the importance of a site can be judged more reliably by obtaining the true feature importance distribution.

[0024] As some optional implementations of the present application, the process of calculating the zero importance distribution of a site is as follows:

[0025] Set the number of random trainings of the decision tree model to m, and randomly obtain m training samples;

[0026] The defective rates corresponding to the randomly obtained m training sample data are shuffled, and the m training sample data with shuffled labels are input into the decision tree model for feature importance sorting to obtain the zero importance distribution of the site.

[0027] In the above embodiment, since the zero feature importance distribution is the result of scrambling the labels, the effect of an effective feature should be greatly reduced after the labels are scrambled, while the feature importance of the originally invalid feature does not change much or increases, so it can be used as an important reference target.

[0028] As some optional embodiments of the present application, the relative importance of the site is expressed as the difference between the minimum value of the true importance distribution and the maximum value of the zero importance distribution, the difference between the median of the true importance distribution and the median of the zero importance distribution, or the difference between the upper quartile of the true importance distribution and the upper quartile of the zero importance distribution.

[0029] In the above embodiments, the present invention can accurately and quickly locate truly important features according to the difference between the real importance distribution and the zero importance distribution.

[0030] As some optional implementation methods of the present application, the process of calculating the defect rate of the equipment at each site step by step according to the feature importance ranking result is as follows:

[0031] According to the feature importance ranking results, several top-ranked sites are obtained;

[0032] Calculate the defect rates of different devices at each site respectively, and output the device with the highest defect rate at each site. The device with the highest defect rate is taken as the root cause device.

[0033] In the above embodiment, the present invention divides the root cause device positioning into first positioning the root cause site and then calculating the device with high defect rate, which reduces the difficulty of directly positioning the root cause device and improves the hit rate of the root cause device.

[0034] In a second aspect, the present invention provides a root cause analysis system based on relative feature importance, the system comprising:

[0035] A raw data acquisition unit, the raw data acquisition unit is used to acquire the product's history data and defect rate, and pre-process the product's history data and defect rate;

[0036] A training sample data acquisition unit, the training sample data acquisition unit is used to perform format conversion and data splicing on the pre-processed history data and defect rate to obtain training sample data;

[0037] An importance distribution unit, wherein the importance distribution unit is used to construct a decision tree model, and input the training sample data into the decision tree model to calculate the true importance distribution and zero importance distribution of the site;

[0038] A relative importance unit, wherein the relative importance unit calculates the relative importance of the site according to the difference between the real importance distribution of the site and the change of the zero importance distribution;

[0039] A bad root cause location unit is used to perform feature importance ranking according to the relative importance of the sites, and to calculate the bad root cause of each site step by step according to the feature importance ranking result.

[0040] In a third aspect, the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method for analyzing root causes of defects based on relative feature importance when executing the computer program.

[0041] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method for analyzing root causes of defects based on relative feature importance.

[0042] The beneficial effects of the present invention are as follows:

[0043] 1. The present invention first calculates the real feature importance distribution and the zero feature importance distribution of each site, and then reorders the sites according to the relative difference between the two distributions, thereby improving the hit rate and stability of the sites in the top ranking results, thereby facilitating experts to quickly verify and take processing measures to improve product yield.

[0044] 2. The present invention divides the root cause device positioning into first positioning the root cause site and then calculating the device with high defect rate, which reduces the difficulty of directly positioning the root cause device and improves the hit rate of the root cause device.

[0045] 3. The present invention pre-screens sites in the data preprocessing stage, which reduces the difficulty of the model directly finding the root cause site from a large number of sites and improves the hit rate of the site.

[0046] 4. The present invention adopts a decision tree model to achieve highly automated root cause analysis of defects. Through end-to-end modeling, the final root cause site can be directly obtained from the input data. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0048] Figure 1 is a schematic diagram of the computer device structure of the hardware operating environment described in the embodiment of the present invention;

[0049] Figure 2 is a flow chart of the defect root cause analysis method according to an embodiment of the present invention;

[0050] Figure 3 is a comparison diagram of the importance distribution of sites simulated in an embodiment of the present invention;

[0051] Figure 4 is a comparison diagram of the real importance distribution and the zero feature importance distribution of the two sites step_11 and step_402 described in the embodiment of the present invention;

[0052] Figure 5 is a schematic diagram comparing the relative feature importance of the two sites step_11 and step_402 in the embodiment of the present invention;

[0053] Figure 6 It is a structural block diagram of the defect root cause analysis system described in an embodiment of the present invention. DETAILED DESCRIPTION

[0054] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0055] In order to solve the problems that the existing root cause analysis methods are time-consuming and costly, the root cause analysis is inaccurate, and the root cause location results cannot be stably obtained. The embodiment of the present invention provides a root cause analysis method and system based on relative feature importance. Before introducing the specific technical solution of the present application, the hardware operating environment involved in the embodiment of the present application is first introduced.

[0056] See also Figure 1 , Figure 1 The figure is a schematic diagram of the computer device structure of the hardware operating environment involved in the embodiment of the present invention.

[0057] like Figure 1 As shown, the computer device may include: a processor, such as a central processing unit (CPU), a communication bus, a user interface, a network interface, and a memory. Among them, the communication bus is used to realize the connection and communication between these components. The user interface may include a display screen (Display), an input unit such as a keyboard (Keyboard), and optionally, the user interface may also include a standard wired interface and a wireless interface. The network interface may be a wired interface or a wireless interface (such as a wireless fidelity interface). The memory may be a high-speed random access memory (Random Access Memory, RAM), or a stable non-volatile memory (Non-Volatile Memory, NVM), such as a disk storage. The memory may optionally be a storage device independent of the aforementioned processor.

[0058] Those skilled in the art will understand that Figure 1 The structure shown in the figure does not constitute a limitation on the computer device, and may include more or less components than shown in the figure, or combine some components, or arrange the components differently.

[0059] like Figure 1 As shown, the memory as a storage medium may include an operating system module, a network communication module, a user interface module and a data storage module.

[0060] exist Figure 1 In the computer device shown, the network interface is mainly used for data communication with the network server; the user interface is mainly used for data interaction with the user; the processor and memory in the computer device of the present application can be set in the computer device, and the computer device calls the bad root cause analysis system based on relative feature importance stored in the data storage module through the processor, and executes the bad root cause analysis method based on relative feature importance provided in the embodiment of the present application.

[0061] Based on the hardware environment of the aforementioned embodiment, the embodiment of the present application provides a method for analyzing the root cause of a defect based on relative feature importance. Figure 2 , Figure 2 The flowchart of the defect root cause analysis method is as follows:

[0062] (1) Obtain product history data and defect rate, and pre-process the product history data and defect rate.

[0063] In the embodiment of the present invention, the historical data of the product includes but is not limited to the product ID at different stages, the site ID, the site type, the equipment ID passed by the site, the processing unit ID passed by the equipment, and the sub-unit ID passed by the processing unit, etc. The defect rate is the ratio of the product defect quantity to the product input quantity.

[0064] In an embodiment of the present invention, since the historical data and defective rate obtained during the fully automatic root cause analysis process may have multiple problems such as missing data records and incorrect data import formats that affect data modeling, it is necessary to uniformly preprocess the obtained historical data and defective rate.

[0065] Specifically, the preprocessing includes but is not limited to data type conversion, outlier removal, and site type screening. Among them, the data type conversion process is to force the product defect quantity and product input quantity corresponding to the defect rate to be converted into Float type. The outlier removal process is to remove the product ID whose defect rate corresponds to zero product input quantity, the product ID whose historical data corresponds to the site ID or device ID is empty, etc. The site type screening process is to retain the site type of processing or rework, because this type of site has a higher product defect rate.

[0066] (2) Convert the format of the preprocessed historical data and defect rate and perform data splicing to obtain training sample data.

[0067] In the embodiment of the present invention, since each product in the product history data only passes through one device at each site, but passes through multiple processing units and processing sub-units at each device, and the parallel equipment or a certain parallel processing sub-unit at the site will cause a batch of products to have different defective rates, the sample features of the model training are the equipment, processing units and processing sub-units that a certain product passes through at each site, and the feature label is the product's defective rate, that is, the sample features and feature labels are combined into a training sample data. Therefore, it is necessary to convert the format of the pre-processed history data and defective rate and perform data splicing.

[0068] Specifically, the process of format conversion and data splicing for the pre-processed resume data and defect rate is as follows:

[0069] (2.1) De-duplicate the pre-processed historical data according to product ID, site ID, and device ID.

[0070] (2.2) The deduplication history data is converted into a structure according to the behavior sample, the column is the site, and the value is the device ID of each product under each site; at the same time, the device ID needs to be encoded. Preferably, the embodiment of the present invention uses numerical encoding (One-Hot encoding) to encode the device ID.

[0071] (2.3) The structural transformation history data and defect rate are concatenated according to the product ID to obtain training sample data that meets the algorithm input.

[0072] (3) Construct a decision tree model and input the training sample data into the decision tree model to calculate the true importance distribution and zero importance distribution of the site.

[0073] In an embodiment of the present invention, in a decision tree model, feature importance is determined by the split contribution of each feature during the model construction process, and the data is divided into purer subsets by recursively splitting the features. The decision tree model can gradually reduce impurity (such as variance or entropy, etc.) to find the features that can most effectively distinguish data categories. In the feature splitting process of the decision tree, each feature is measured by reducing the magnitude of impurity when it is selected as a splitting node, which is called feature splitting importance.

[0074] Specifically, the splitting process of the decision tree model depends on the correlation between each feature and the target variable. The decision tree model will automatically select the optimal feature for node splitting to minimize the variance or entropy within the node, thereby improving the decision tree model's ability to capture data patterns. For historical data with multiple sites, the decision tree model can measure the impact of each site on the final defect rate by calculating the contribution of each site in data splitting. Since the decision tree model is based on the information gain of the current node at each split, it can effectively identify the site that is most sensitive to defect rate fluctuations.

[0075] Specifically, the decision tree model takes all sites as features, and a certain device ID or subunit ID under the site as the value of the feature. By inputting the training sample data into the decision tree model to fit the defect rate, the feature importance of each site can be obtained. By sorting the feature importance of all sites, several sites with the highest split importance are selected. These sites are considered to be the root cause sites that are highly correlated with the defect.

[0076] In the embodiment of the present invention, in the decision tree model, since a certain proportion of samples and a certain proportion of features are randomly selected for splitting to construct the decision tree model, the feature importance obtained by the final decision tree model is also random. In order to fully utilize the generalization ability brought by the randomness of the decision tree model and increase the stability of the feature importance ranking as much as possible, it is necessary to input the training sample data into the decision tree model to calculate the real importance distribution and zero importance distribution of the site.

[0077] Specifically, the process of calculating the true importance distribution of sites through the decision tree model is as follows:

[0078] (3.1) The number of random trainings of the decision tree model is set to n, and n training samples are randomly obtained, where n is set according to actual conditions and is not limited in the embodiment of the present invention.

[0079] (3.2) The n randomly obtained training sample data are input into the decision tree model for feature importance sorting to obtain n kinds of feature importance sorting under n different random numbers. Based on these results, the true importance distribution of each site can be obtained.

[0080] Specifically, the process of calculating the zero importance distribution of a site through a decision tree model is as follows:

[0081] (3.3) The number of random trainings of the decision tree model is set to m, and m training samples are randomly obtained, where m is set according to actual conditions and is not limited in the embodiment of the present invention.

[0082] (3.4) The defective rates corresponding to the m randomly obtained training sample data are shuffled, and the m training sample data with shuffled labels are input into the decision tree model for feature importance sorting. At the same time, m kinds of feature importance sorting under m different random numbers are obtained. Based on these results, the zero importance distribution of each site can be obtained.

[0083] In the embodiment of the present invention, the real feature importance distribution reflects the approximate value range of the feature importance of a feature in multiple random trainings, and can more confidently judge the importance of the feature. A single random training may result in a feature importance that is too high or too low, thereby affecting the selection of K features before sorting. The zero feature importance is the result of scrambling the labels. The effect of an effective feature should be greatly reduced after the labels are scrambled, while the feature importance of the originally invalid feature will not change much or will increase. Therefore, the truly important features can be located based on the difference in changes before and after.

[0084] (4) Calculate the relative importance of the site based on the difference in the real importance distribution of the site and the change in the zero importance distribution.

[0085] In the embodiment of the present invention, after obtaining the two importance distributions, if the two importance distributions of each site are plotted on the same coordinate axis, if a site is very important, it is expected that its zero feature importance distribution is on the left side of the coordinate axis and the true importance distribution is on the right side of the coordinate axis, and the two are as far away as possible and do not intersect each other. If a site is not important, the two corresponding distributions have crossover values, and even the zero feature importance distribution is on the right side of the true importance distribution. Please refer to Figure 3 , Figure 3 This is a comparison chart of the simulated site importance distribution.

[0086] In the embodiment of the present invention, the relative importance of the site includes but is not limited to:

[0087] ①The difference between the minimum value of the true importance distribution and the maximum value of the zero importance distribution.

[0088] ②The difference between the median of the true importance distribution and the median of the zero importance distribution.

[0089] ③The difference between the upper quartile of the true importance distribution and the upper quartile of the zero importance distribution.

[0090] Specifically, considering the impact of abnormal data in the distribution, the embodiment of the present invention uses the difference between the median of the true importance distribution and the median of the zero importance distribution to calculate the relative difference between the two importance distributions. The larger the difference, the more important the feature is, and the difference is used as the new relative importance.

[0091] (5) The sites are ranked by their relative importance, and the defect rate of the equipment at each site is calculated step by step based on the ranking results.

[0092] In the embodiment of the present invention, the process of calculating the defect rate of the equipment at each site step by step according to the feature importance ranking result is as follows:

[0093] (5.1) According to the feature importance ranking results, several top-ranked sites are obtained.

[0094] (5.2) Calculate the defect rate of different devices at each site respectively, and output the device with the highest defect rate at each site. The device with the highest defect rate is taken as the root cause device.

[0095] Specifically, after obtaining the top K important sites, the defective rate of the equipment, the defective rate of the processing unit, and the defective rate of the processing sub-unit under each site are calculated step by step. If there are multiple devices with the largest defective rate, the device is directly skipped to calculate the defective rates of these multiple processing units, and the device of the processing unit with the largest defective rate is used as the root cause device; if the defective rates of each processing unit are also the same, the processing unit is directly skipped to calculate the defective rates of these multiple processing sub-units, and so on, to obtain the root cause device with the highest defective rate.

[0096] During the display panel processing, the processed product will be in the form of a glass substrate, a glass half-panel, and finally a display panel. At different stages of the form, there will be a unique ID as a product identification. After inspection, the final display panel will have good and defective products. Defective products are represented by different defect codes according to the defect type they contain. For display panels with the same type of defect code, the historical data and defective rate of these display panels at a certain stage of the form are obtained to perform defect root cause analysis.

[0097] The following is a specific process of the defect root cause analysis method when the product form stage is a glass substrate during the display panel processing process, including the following steps:

[0098] S10: acquiring the history data and the defect rate of the glass substrate, and pre-processing the history data and the defect rate of the glass substrate.

[0099] Specifically, first obtain a batch of display panels containing target defect codes, and then obtain the glass substrates corresponding to the batch of display panels. For these glass substrates, obtain the set of sites that the glass substrates have passed through, obtain all the glass substrates that have passed through these sites within a certain period of time, and obtain the historical data of the batch of glass substrates and the ratio of the product defect quantity to the product input quantity corresponding to each glass substrate. Please refer to Table 1 and Table 2. Table 1 is the historical data of the glass substrates, and Table 2 is the product defect quantity and product input quantity corresponding to the defect rate.

[0100] Glass substrate ID Site Site Type Device ID Processing unit ID Processing subunit ID GLASS_1 step_1 Production EQP_1 UNIT_1 SUBUNIT_1 GLASS_2 step_2 Production EQP_7 UNIT_7 - GLASS_3 step_3 Production EQP_2 UNIT_2 SUBUNIT_2 GLASS_4 step_4 Production EQP_3 UNIT_3 SUBUNIT_3 GLASS_5 step_5 Production EQP_4 UNIT_4 SUBUNIT_4 GLASS_6 step_6 Production EQP_5 UNIT_5 SUBUNIT_5 GLASS_7 step_8 Production EQP_8 - - GLASS_8 step_7 Production EQP_6 UNIT_6 SUBUNIT_6

[0101] Table 1

[0102]

[0103]

[0104] Table 2

[0105] Finally, the data in Table 1 and Table 2 are preprocessed: including converting the input quantity and defect quantity for defect rate comparison into Float type; removing the rows with empty or abnormal equipment ID, processing unit ID, and processing sub-unit ID; and retaining the sites with production or rework type.

[0106] S20: Perform format conversion and data splicing on the pre-processed history data and defect rate to obtain training sample data.

[0107] Specifically, after preprocessing the data, each row of the history data represents the equipment, processing unit, and processing sub-unit that a glass substrate passes through at a site. Each glass substrate will pass through multiple processing or rework sites, so each glass substrate and the equipment ID (or processing sub-unit ID) of the glass substrate at different sites can be regarded as a sample. In order to make the data conform to the data input specifications of the machine learning model (behavior samples, columns as features), the preprocessed history data is converted to a table structure. Please refer to Table 3, which is the data structure after format conversion. Each row in Table 3 represents the equipment that a glass substrate passes through at a given site. If it is empty, it means that the glass substrate has not passed through the site corresponding to this column.

[0108] Glass substrate ID step_1 step_3 step_4 step_5 step_6 step_7 GLASS_1 EQP_1 EQP_9 EQP_10 EQP_12 EQP_6 GLASS_3 EQP_6 EQP_2 EQP_4 EQP_12 EQP_6 GLASS_4 EQP_6 EQP_9 EQP_3 EQP_11 EQP_12 EQP_6 GLASS_5 EQP_6 EQP_9 EQP_3 EQP_4 EQP_6 GLASS_6 EQP_13 EQP_2 EQP_3 EQP_4 EQP_5 EQP_6 GLASS_8 EQP_13 EQP_2 EQP_11 EQP_5 EQP_6

[0109] Table 3

[0110] Specifically, after obtaining the corresponding data in Table 3, it is necessary to numerically encode the equipment under each site to obtain the sample characteristics. Please refer to Table 4 for the sample characteristics.

[0111]

[0112]

[0113] Table 4

[0114] Specifically, the label of each glass substrate is calculated by the relevant data in Table 2. Please refer to Table 5, which is the characteristic label.

[0115] Glass substrate ID Defective rate GLASS_1 0 / 195 GLASS_2 0 / 192 GLASS_3 12 / 198 GLASS_4 3 / 194 GLASS_5 7 / 190 GLASS_6 11 / 196 GLASS_7 0 / 197 GLASS_8 0 / 196

[0116] Table 5

[0117] S30: construct a decision tree model, and input the training sample data into the decision tree model to calculate the true importance distribution and zero importance distribution of the site.

[0118] Specifically, since LightGBM is a gradient boosting tree with stronger fitting ability, the embodiment of the present invention selects LightGBM as the decision tree model.

[0119] Specifically, considering the running time and the calculation of the feature importance distribution, the embodiment of the present invention sets the number of repeated training n to 100, that is, the decision tree model is repeatedly trained 100 times to obtain 100 feature importance values ​​of each site, that is, the real importance distribution of each site is obtained; the decision tree model is then repeatedly trained 100 times with the same parameters, and the labels (that is, the defective rate of each glass substrate) are randomly shuffled during each training to obtain 100 importance values ​​of each site, that is, the zero feature importance distribution of the site is obtained. Please refer to Figure 4 , Figure 4 This is a comparison chart of the true importance distribution and zero feature importance distribution of the two sites step_11 and step_402. Figure 4 It can be seen that the real feature importance of step_11 site is concentrated between 1000 and 1300, the zero feature importance is concentrated near 0, and the two importance values ​​have no intersection; while the real feature importance of step_402 site is concentrated between 50 and 160, and the value of zero feature importance is partially greater than the real feature importance, indicating that the importance of this site has increased after a certain shuffle of the labels. Therefore, it can be considered that the importance of step_11 site is higher than that of step_402 site.

[0120] S40: Calculate the relative importance of the site according to the real importance distribution of the site and the difference in the change of the zero importance distribution.

[0121] That is, after calculating the two importance distributions of each site in step S30, for each site, the difference between the median of the true feature importance and the median of the zero feature importance is calculated as the relative feature importance of the site, see Figure 5 , Figure 5 The figure is a comparison diagram of the relative feature importance of the two sites step_11 and step_402. The difference between the median of the true feature importance of the step_11 site and the median of the zero feature importance is 1144.3, and the difference between the median of the true feature importance of the step_402 site and the median of the zero feature importance is 90.9.

[0122] S50: The sites are ranked by feature importance according to their relative importance, and the defect rate of the equipment at each site is calculated level by level according to the feature importance ranking result.

[0123] After obtaining the relative importance of the site, it is necessary to sort the relative feature importance to obtain the top 5 sites. Please refer to Table 6, which shows the relative feature importance of the top 5 sites.

[0124] Site ID Relative feature importance step_11 1144.3 step_12 611.3 step_13 469.4 step_14 445.6 step_15 404.2

[0125] Table 6

[0126] Furthermore, after obtaining the top 5 sites, the defect rates of different equipment at each site are calculated respectively, and the equipment with the highest defect rate at each site is output. Then, the defect rates of different processing sub-units at each returned equipment are calculated, and the processing sub-unit with the highest defect rate at the equipment is returned. For the final output results, please refer to Table 7, which is the output results of the root cause analysis of the top 5 sites.

[0127]

[0128] Table 7

[0129] In order to illustrate the effectiveness of the embodiment of the present invention, the root cause analysis results and hit situations of other three cases are given here. Please refer to Table 8, which shows the root cause analysis results and hit situations of the other three cases.

[0130]

[0131] Table 8

[0132] In summary, the method of the present invention has achieved the top 5 hit sites and abnormal equipment and processing sub-units under the sites in 4 cases with different data volumes, which is convenient for business experts and on-site staff to quickly locate the root cause of the defect, take treatment measures, and improve product yield.

[0133] In addition, in one embodiment, based on the same inventive concept as the above embodiment, the embodiment of the present invention provides a root cause analysis system based on relative feature importance, see Figure 3 , Figure 3 is a structural block diagram of the defect root cause analysis system, the system corresponds one-to-one to the method of Example 1, and the system includes:

[0134] A raw data acquisition unit, the raw data acquisition unit is used to acquire the product's history data and defect rate, and pre-process the product's history data and defect rate;

[0135] A training sample data acquisition unit, the training sample data acquisition unit is used to perform format conversion and data splicing on the pre-processed history data and defect rate to obtain training sample data;

[0136] An importance distribution unit, wherein the importance distribution unit is used to construct a decision tree model, and input the training sample data into the decision tree model to calculate the true importance distribution and zero importance distribution of the site;

[0137] A relative importance unit, wherein the relative importance unit calculates the relative importance of the site according to the difference between the real importance distribution of the site and the change of the zero importance distribution;

[0138] A bad root cause location unit is used to perform feature importance ranking according to the relative importance of the sites, and to calculate the bad root cause of each site step by step according to the feature importance ranking result.

[0139] It should be noted that each unit in the bad root cause analysis system based on relative feature importance in this embodiment corresponds one-to-one to each step in the bad root cause analysis method based on relative feature importance in the aforementioned embodiment. Therefore, the specific implementation manner and technical effects achieved by this embodiment can refer to the implementation manner of the aforementioned bad root cause analysis method based on relative feature importance, and will not be repeated here.

[0140] In addition, in one embodiment, the present application also provides a computer device, which includes a processor, a memory, and a computer program stored in the memory, and the computer program implements the method in the aforementioned embodiment when executed by the processor.

[0141] In addition, in one embodiment, the present application further provides a computer storage medium, on which a computer program is stored, and the computer program implements the method in the aforementioned embodiment when executed by a processor.

[0142] In some embodiments, the computer readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or various devices including one or any combination of the above memories. The computer may be various computing devices including intelligent terminals and servers.

[0143] In some embodiments, executable instructions may be in the form of a program, software, software module, script or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine or other unit suitable for use in a computing environment.

[0144] As an example, executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files storing one or more modules, subroutines, or code portions).

[0145] By way of example, executable instructions may be deployed to be executed on one computing device, or on multiple computing devices located at one site, or on multiple computing devices distributed across multiple sites and interconnected by a communication network.

[0146] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0147] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0148] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a multimedia terminal device (which can be a mobile phone, a computer, a television receiver, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0149] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for root cause analysis based on relative feature importance, characterized in that: The method comprises the following steps: Obtain product history data and defect rate, and pre-process the product history data and defect rate; Perform format conversion and data splicing on the pre-processed resume data and defect rate to obtain training sample data; A decision tree model is constructed, and the training sample data is input into the decision tree model to calculate the true importance distribution and zero importance distribution of the site; The relative importance of the sites is calculated based on the difference between the true importance distribution of the sites and the change of the zero importance distribution; The feature importance is ranked according to the relative importance of the sites, and the root cause of the badness of each site is calculated level by level based on the feature importance ranking results.

2. The method for root cause analysis based on relative feature importance according to claim 1, characterized in that: The preprocessing includes data type conversion, outlier removal, and site type screening.

3. The method for root cause analysis based on relative feature importance according to claim 1, characterized in that: The process of format conversion and data splicing of the pre-processed resume data and defect rate to obtain training sample data is as follows: De-duplicate the pre-processed historical data according to product ID, site ID, and device ID; The structure of the deduplicated historical data is converted according to the table structure of behavior samples, columns as sites, and values ​​as devices of each product under each site; The structural conversion history data and defect rate are concatenated according to the product ID to obtain the training sample data.

4. The method for root cause analysis based on relative feature importance according to claim 1, characterized in that: The process of calculating the true importance distribution of a site is as follows: Set the number of random trainings of the decision tree model to n, and randomly obtain n training samples; The randomly obtained n training sample data are input into the decision tree model to sort the feature importance to obtain the true importance distribution of the site.

5. The method for root cause analysis based on relative feature importance according to claim 1, characterized in that: The process of calculating the zero importance distribution of a site is as follows: Set the number of random trainings of the decision tree model to m, and randomly obtain m training samples; The defective rates corresponding to the randomly obtained m training sample data are shuffled, and the m training sample data with shuffled labels are input into the decision tree model for feature importance sorting to obtain the zero importance distribution of the site.

6. The method for root cause analysis based on relative feature importance according to claim 1, characterized in that: The relative importance of the site is expressed as the difference between the minimum value of the true importance distribution and the maximum value of the zero importance distribution, the difference between the median of the true importance distribution and the median of the zero importance distribution, or the difference between the upper quartile of the true importance distribution and the upper quartile of the zero importance distribution.

7. The method for root cause analysis based on relative feature importance according to claim 1, characterized in that: The process of calculating the root cause of each site step by step according to the feature importance ranking results is as follows: According to the feature importance ranking results, several top-ranked sites are obtained; Calculate the defect rates of different devices at each site respectively, and output the device with the highest defect rate at each site. The device with the highest defect rate is taken as the root cause device.

8. A root cause analysis system based on relative feature importance, characterized in that: The system comprises: A raw data acquisition unit, the raw data acquisition unit is used to acquire the product's history data and defect rate, and pre-process the product's history data and defect rate; A training sample data acquisition unit, the training sample data acquisition unit is used to perform format conversion and data splicing on the pre-processed history data and defect rate to obtain training sample data; An importance distribution unit, wherein the importance distribution unit is used to construct a decision tree model, and input the training sample data into the decision tree model to calculate the true importance distribution and zero importance distribution of the site; A relative importance unit, wherein the relative importance unit calculates the relative importance of the site according to the difference between the real importance distribution of the site and the change of the zero importance distribution; A bad root cause location unit is used to perform feature importance ranking according to the relative importance of the sites, and to calculate the bad root cause of each site step by step according to the feature importance ranking result.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method for analyzing root causes of defects based on relative feature importance described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for analyzing root causes of defects based on relative feature importance as described in any one of claims 1 to 7 is implemented.

Images