Yield prediction model training method and device, computer equipment and storage medium

Abstract

The invention provides a yield prediction model training method and device, computer equipment and a storage medium, and belongs to the technical field of computers. In the method, through an encoder and a decoder in a yield prediction model, equipment data of production equipment through which a sample product passes is processed, and the model is adjusted according to a prediction result output by the model and an actual result corresponding to the sample product. The encoder in the model can compress the high-dimensional data so as to extract the effective features from the high-dimensional data, so that the trained encoder can process the high-dimensional equipment data so as to extract the effective feature data from the high-dimensional equipment data. And a trained decoder in the model can process the effective feature data to obtain an accurate prediction result. According to the accurate prediction result, the production process is adjusted before the product is produced, the efficiency of adjusting the production process can be improved, and the production efficiency of the product is further improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a training method, apparatus, computer equipment and storage medium for a yield prediction model. Background Art

[0002] In the semiconductor industry, the production process of products is complex and the production equipment is numerous, resulting in a variety of defective products produced in the production process. Different defective types have their corresponding yields, so the production process of the product can be analyzed based on the defective type and the corresponding yield.

[0003] At present, defect analysis for product production processes mainly focuses on testing products in the production process or products that have been produced, and analyzing the subsequent production process based on the test results, so as to gradually adjust the production process in the subsequent production process. The efficiency of adjusting the production process is low. Therefore, there is a lack of a yield prediction model training method to improve the efficiency of adjusting the production process, thereby improving the production efficiency of the product. Summary of the invention

[0004] The present application embodiment provides a yield prediction model training method, device, computer equipment and storage medium, which are used to improve the efficiency of adjusting the production process. The technical solution is as follows:

[0005] In a first aspect, a method for training a yield prediction model is provided, the method comprising:

[0006] Inputting the equipment data of the production equipment used in the production of the sample product into the yield prediction model, the equipment data includes equipment parameters of multiple dimensions, which are the parameters used by the production equipment when producing the sample product. The yield prediction model includes an encoder and a decoder;

[0007] The device data is processed by an encoder to obtain feature data corresponding to the device data, where the feature data includes features of multiple dimensions;

[0008] Processing the feature data through a decoder, and outputting a prediction result corresponding to the sample product, the prediction result including at least one defect type corresponding to the sample product and a yield rate corresponding to the defect type;

[0009] Based on the prediction results and actual results corresponding to the sample products, the yield prediction model is adjusted to obtain a trained yield prediction model.

[0010] The training method of the yield prediction model provided in the embodiment of the present application processes the equipment data of the production equipment through which the sample product is produced through the encoder and decoder in the yield prediction model, and adjusts the model according to the prediction results output by the model and the actual results corresponding to the sample product. Since the encoder in the model can compress high-dimensional data, thereby extracting effective features of the high-dimensional data from the high-dimensional data, the trained encoder can process the high-dimensional equipment data to extract effective feature data therefrom. The decoder trained in the model can process the above-mentioned effective feature data to obtain accurate prediction results. According to the accurate prediction results, the production process is adjusted before the product is produced, which can improve the efficiency of adjusting the production process, thereby improving the production efficiency of the product.

[0011] Optionally, before inputting the equipment data of the production equipment through which the sample product is produced into the yield prediction model, the method further includes:

[0012] Preprocess the device data of the sample product, and the preprocessing includes at least one of the following:

[0013] If it is detected that the yield of the sample product is less than the yield threshold, the device data of the sample product is discarded;

[0014] If it is detected that the number of missing device parameters in the device data of the sample product is greater than or equal to the missing threshold, the device data of the sample product is discarded;

[0015] Supplement the missing equipment parameters in the equipment data of sample products;

[0016] Normalize the device data of the sample products.

[0017] Optionally, the yield prediction model includes a first network module and a second network module, both of which include an encoder, and the feature data corresponding to the device data obtained by processing the device data through the encoder includes:

[0018] Processing the device data through an encoder in the first network module to obtain first characteristic data corresponding to the device data;

[0019] The device data is processed by an encoder in the second network module to obtain second characteristic data corresponding to the device data, and the model parameters in the second network module have opposite signs to the corresponding model parameters in the first network module.

[0020] Optionally, both the first network module and the second network module include a decoder, and the processing of the feature data by the decoder to output a prediction result corresponding to the sample product includes:

[0021] Processing the first feature data through a decoder in the first network module to obtain a first prediction result corresponding to the sample product;

[0022] Processing the second feature data through a decoder in the second network module to obtain a second prediction result corresponding to the sample product;

[0023] The sum of the first prediction result and the second prediction result is output as the prediction result corresponding to the sample product.

[0024] Optionally, the training process of the yield prediction model includes multiple rounds of training. The adjustment of the yield prediction model based on the prediction results and actual results corresponding to the sample products includes:

[0025] Based on the prediction results and actual results corresponding to the preset number of sample products, a loss value is obtained, where the loss value indicates the accuracy of the yield prediction model prediction;

[0026] Based on the loss value, the yield prediction model is adjusted for the next round of training.

[0027] Optionally, the above-mentioned prediction results and actual results corresponding to the preset number of sample products, obtaining the loss value includes:

[0028] The loss value is calculated based on the repetition penalty parameter, the predicted results corresponding to the preset number of sample products, and the actual results. The repetition penalty parameter indicates the number of times the preset number of sample products are repeatedly used for model training.

[0029] Optionally, the above-mentioned prediction results and actual results corresponding to the preset number of sample products, obtaining the loss value includes:

[0030] If it is detected that the number of training rounds of the yield prediction model is less than the preset number of rounds, a loss value is obtained based on the prediction results and actual results corresponding to the first number of sample products;

[0031] If it is detected that the number of training rounds of the yield prediction model is greater than or equal to the preset number of rounds, the loss value is obtained based on the predicted results and actual results corresponding to the second sample number of sample products, and the second sample number is less than the first sample number.

[0032] In a second aspect, a training device for a yield prediction model is provided, the device comprising:

[0033] An input module, used to input the equipment data of the production equipment used in the production of the sample product into the yield prediction model, the equipment data includes equipment parameters of multiple dimensions, and the equipment parameters are parameters used by the production equipment when producing the sample product. The yield prediction model includes an encoder and a decoder;

[0034] A first processing module, configured to process the device data through an encoder to obtain feature data corresponding to the device data, where the feature data includes features of multiple dimensions;

[0035] A second processing module is used to process the feature data through a decoder and output a prediction result corresponding to the sample product, where the prediction result includes at least one defect type corresponding to the sample product and a yield rate corresponding to the defect type;

[0036] The adjustment module is used to adjust the yield prediction model based on the prediction results and actual results corresponding to the sample products to obtain the trained yield prediction model.

[0037] Optionally, the above device further includes:

[0038] The preprocessing module is used to preprocess the device data of the sample product, and the preprocessing includes at least one of the following:

[0039] If it is detected that the yield of the sample product is less than the yield threshold, the device data of the sample product is discarded;

[0040] If it is detected that the number of missing device parameters in the device data of the sample product is greater than or equal to the missing threshold, the device data of the sample product is discarded;

[0041] Supplement the missing equipment parameters in the equipment data of sample products;

[0042] Normalize the device data of the sample products.

[0043] Optionally, the yield prediction model includes a first network module and a second network module, both the first network module and the second network module include an encoder, and the first processing module includes:

[0044] A first feature data acquisition unit, configured to process the device data through an encoder in the first network module to obtain first feature data corresponding to the device data;

[0045] The second characteristic data acquisition unit is used to process the device data through the encoder in the second network module to obtain second characteristic data corresponding to the device data. The model parameters in the second network module have opposite positive and negative signs to the corresponding model parameters in the first network module.

[0046] Optionally, both the first network module and the second network module include a decoder, and the second processing module is used to:

[0047] Processing the first feature data through a decoder in the first network module to obtain a first prediction result corresponding to the sample product;

[0048] Processing the second feature data through a decoder in the second network module to obtain a second prediction result corresponding to the sample product;

[0049] The sum of the first prediction result and the second prediction result is output as the prediction result corresponding to the sample product.

[0050] Optionally, the training process of the yield prediction model includes multiple rounds of training, and the above adjustment module includes:

[0051] A loss value acquisition unit, used to acquire a loss value based on the prediction results and actual results corresponding to a preset number of sample products, wherein the loss value indicates the accuracy of the yield prediction model prediction;

[0052] The adjustment unit is used to adjust the yield prediction model based on the loss value to perform a next round of training.

[0053] Optionally, the loss value obtaining unit is used to:

[0054] The loss value is calculated based on the repetition penalty parameter, the predicted results corresponding to the preset number of sample products, and the actual results. The repetition penalty parameter indicates the number of times the preset number of sample products are repeatedly used for model training.

[0055] Optionally, the loss value obtaining unit is used to:

[0056] If it is detected that the number of training rounds of the yield prediction model is less than the preset number of rounds, a loss value is obtained based on the prediction results and actual results corresponding to the first number of sample products;

[0057] If it is detected that the number of training rounds of the yield prediction model is greater than or equal to the preset number of rounds, the loss value is obtained based on the predicted results and actual results corresponding to the second sample number of sample products, and the second sample number is less than the first sample number.

[0058] In a third aspect, a yield prediction method is provided, the method comprising:

[0059] Inputting the equipment data of the production equipment used in the production of the product into the yield prediction model, the equipment data includes equipment parameters of multiple dimensions, which are the parameters used by the production equipment when producing the product. The yield prediction model includes an encoder and a decoder;

[0060] The device data is processed by an encoder to obtain feature data corresponding to the device data, where the feature data includes features of multiple dimensions;

[0061] The feature data is processed by a decoder to output a prediction result corresponding to the product, where the prediction result includes at least one defect type corresponding to the product and the yield rate corresponding to the defect type.

[0062] Optionally, before inputting the equipment data of the production equipment through which the product is produced into the yield prediction model, the method further includes:

[0063] Normalize the device data of the product.

[0064] Optionally, the yield prediction model includes a first network module and a second network module, both of which include an encoder, and the feature data corresponding to the device data obtained by processing the device data through the encoder includes:

[0065] Processing the device data through an encoder in the first network module to obtain first characteristic data corresponding to the device data;

[0066] The device data is processed by an encoder in the second network module to obtain second characteristic data corresponding to the device data, and the model parameters in the second network module have opposite signs to the corresponding model parameters in the first network module.

[0067] Optionally, both the first network module and the second network module include a decoder, and the processing of the feature data by the decoder to output a prediction result corresponding to the product includes:

[0068] Processing the first feature data through a decoder in the first network module to obtain a first prediction result corresponding to the product;

[0069] Processing the second feature data through a decoder in the second network module to obtain a second prediction result corresponding to the product;

[0070] The sum of the first prediction result and the second prediction result is output as the prediction result corresponding to the product.

[0071] In a fourth aspect, a computer device is provided, comprising a processor and a memory, wherein the memory is used to store at least one computer program, and the at least one computer program is loaded and executed by the processor to implement the operations performed by the yield prediction model training method provided in the above-mentioned first aspect or various optional implementations of the first aspect.

[0072] In a fifth aspect, a computer device is provided, comprising a processor and a memory, wherein the memory is used to store at least one computer program, and the at least one computer program is loaded and executed by the processor to implement the operations performed by the yield prediction method provided in the third aspect or various optional implementations of the third aspect.

[0073] In a sixth aspect, a computer-readable storage medium is provided, wherein at least one computer program is stored in the computer-readable storage medium, and the at least one computer program is loaded and executed by a processor to implement the operations performed by the yield prediction model training method provided in the above-mentioned first aspect or various optional implementations of the first aspect.

[0074] In the seventh aspect, a computer-readable storage medium is provided, in which at least one computer program is stored. The at least one computer program is loaded and executed by a processor to implement the operations performed by the yield prediction method provided in the third aspect or various optional implementations of the third aspect.

[0075] In an eighth aspect, a computer program product or a computer program is provided, which includes a computer program code, and the computer program code is stored in a computer-readable storage medium. A processor of a computer device reads the computer program code from the computer-readable storage medium, and the processor executes the computer program code, so that the computer device performs the operations performed by the yield prediction model training method provided in the above-mentioned first aspect or various optional implementations of the first aspect.

[0076] In the ninth aspect, a computer program product or a computer program is provided, which includes a computer program code, and the computer program code is stored in a computer-readable storage medium. A processor of a computer device reads the computer program code from the computer-readable storage medium, and the processor executes the computer program code, so that the computer device performs the operations performed by the yield prediction method provided in the third aspect or various optional implementations of the third aspect.

[0077] Based on the implementations provided in the above aspects, the present application can also be further combined to provide more implementations. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0079] Figure 1 This is a schematic diagram of an application environment of a yield prediction model provided in an embodiment of the present application;

[0080] Figure 2 It is a model structure diagram of a yield prediction model provided in an embodiment of the present application;

[0081] Figure 3 It is a flow chart of a method for training a yield prediction model provided in an embodiment of the present application;

[0082] Figure 4 It is a training flow chart of a yield prediction model provided in an embodiment of the present application;

[0083] Figure 5 is a flow chart of a method for obtaining training data provided in an embodiment of the present application;

[0084] Figure 6 This is an example diagram of a client interface provided by an embodiment of the present application;

[0085] Figure 7 This is an example diagram of a training data import panel provided in an embodiment of the present application;

[0086] Figure 8 This is an example diagram of a user login interface provided in an embodiment of the present application;

[0087] Fig. 9 This is an example diagram of a bad type import panel provided in an embodiment of the present application;

[0088] Fig.10 is a flowchart of another yield prediction model training method provided in an embodiment of the present application;

[0089] Fig.11 This is an example diagram of a test product level provided in an embodiment of the present application;

[0090] Fig.12 This is an example diagram of a defect threshold import panel provided in an embodiment of the present application;

[0091] Fig.13 This is an example diagram of the impact of a device parameter on yield provided by an embodiment of the present application;

[0092] Fig.14 It is a structural block diagram of a training device for a yield prediction model provided in an embodiment of the present application;

[0093] Fig.15 It is a structural diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0094] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below in conjunction with the accompanying drawings.

[0095] In this application, the terms "first", "second", etc. are used to distinguish identical or similar items with basically the same effects and functions. It should be understood that there is no logical or temporal dependency between "first", "second", and "nth", nor is there any limitation on quantity and execution order.

[0096] In the present application, the term "at least one" means one or more, and the term "plurality" means two or more.

[0097] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions. For example, the device parameters involved in this application are obtained with full authorization.

[0098] Figure 1 Schematic diagram of an application environment of a yield prediction model provided in an embodiment of the present application. Figure 1 As shown, the implementation environment includes a terminal 101 and a server 102, which are collectively referred to as computer devices. The terminal 101 is connected to the server 102 via a wired network or a wireless network. Figure 1 In the embodiment, the number of terminals and servers is only an example. For example, the number of terminals and servers can be multiple, and this is not limited in the embodiments of the present application.

[0099] Among them, terminal 101 is a device used by operation and maintenance personnel, management personnel or on-site workers in the production workshop, such as smart phones, tablet computers, laptops, desktop computers, smart speakers, smart watches, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, etc., but are not limited to this. A production analysis system is installed on the terminal 101, through which the terminal 101 can receive data of production equipment sent by other computer devices or input by users, and can also obtain data of production equipment from a database, which is not limited in the embodiment of the present application. Through the above-mentioned production analysis system, the terminal 101 can analyze the production situation of products in the production workshop based on the data of the above-mentioned production equipment, such as analyzing the yield of the product.

[0100] The server 102 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), big data and artificial intelligence platforms. The server 102 is used to provide background services for the production analysis system installed on the terminal 101, that is, the above-mentioned production analysis system is deployed with a yield prediction model, and the yield prediction model is deployed on the server 102. The server 102 can predict the yield of the product produced by the production equipment according to the data of the production equipment sent by the terminal, and return the prediction result to the terminal.

[0101] In some embodiments, the above-mentioned production analysis system is deployed on terminal 101 and the yield prediction model deployed in the production analysis system is also deployed on terminal 101. Terminal 101 can independently predict the yield of the product based on the data of the production equipment. The embodiments of the present application are not limited to this.

[0102] The yield prediction model involved in this application is briefly described below. Figure 2 is a model structure diagram of a yield prediction model provided in an embodiment of the present application, such as Figure 2 As shown, the yield prediction model includes a first network module and a second network module. Different network modules are used to process the input data differently to extract as many features as possible from the input data. Each network module includes an encoder and a decoder. The encoder is used to map the input high-dimensional data to low-dimensional feature data to extract effective features from the high-dimensional data, and the decoder is used to map the low-dimensional feature data to the prediction result, thereby outputting the prediction result. The training process of the yield prediction model is described below.

[0103] Figure 3 is a flow chart of a method for training a yield prediction model provided in an embodiment of the present application, such as Figure 3 As shown, taking the method for training the yield prediction model executed by a computer device as an example, the method includes the following steps:

[0104] 301. The computer device inputs the equipment data of the production equipment used in the production of the sample product into the yield prediction model. The equipment data includes equipment parameters of multiple dimensions. The equipment parameters are the parameters used by the production equipment when producing the sample product. The yield prediction model includes an encoder and a decoder.

[0105] Among them, the sample product is a product belonging to a preset product type, such as a certain type of TFT-LCD (Thin film transistor liquid crystal display) display panel. The production equipment used in the production of the above sample products refers to the production equipment used in the production of the sample products. In the production process of the sample products, the production equipment used to produce the sample products includes multiple equipment, and different production equipment is used to perform different treatments on the sample products. Taking the display panel as an example, in the production process of the display panel, the production equipment used to produce the display panel includes a cutting machine, a cleaning machine, a hot press machine and a testing equipment, etc. The cutting machine is used to cut the display panel, the cleaning machine is used to clean the display panel, the hot press machine is used to perform hot pressing operations on the display panel to bond the various components in the display panel, and the testing equipment is used to test the performance and quality of the display panel. The above equipment data includes equipment parameters of multiple dimensions, and the equipment parameters of multiple dimensions include the parameters used by the above multiple production equipment when producing the sample products. For example, the equipment data of the display panel includes the cutting thickness of the cutting machine, the water pressure of the cleaning machine, and the temperature of the hot press machine. In the above-mentioned multiple dimensions of equipment parameters, the dimension indicates the parameter type, and the equipment parameter indicates the parameter value. For example, a certain dimension indicates the parameter type of the cutting thickness of the cutting machine, and the equipment parameter of this dimension indicates the value of the cutting thickness. The yield prediction model is a neural network model used to predict the yield of the product based on the above-mentioned equipment data.

[0106] The encoder in the yield prediction model includes multiple neural network layers, each of which includes multiple model parameters, which are used to extract features from the input device data. In the above device data, each piece of data contains a large number of device parameters of different dimensions. The encoder can map the input high-dimensional device data to low-dimensional feature data to extract the effective features corresponding to the device data from the high-dimensional device data, that is, convert the important features contained in the input data into an effective internal representation in the yield prediction model.

[0107] The decoder in the yield prediction model also includes multiple neural network layers, each of which includes multiple model parameters, which are used to reconstruct the feature data obtained by the encoder to map the above feature data to the following prediction results, that is, to convert the above internal representation into an output of a specified dimension.

[0108] In addition, the sample product may be a product that has been produced in a production workshop, and accordingly, the equipment data corresponding to the sample product includes the equipment parameters actually used by the production equipment when producing the sample product. The above-mentioned sample product may also be a fictitious product, and accordingly, the equipment data corresponding to the sample product includes the equipment parameters set by relevant staff based on historical experience, which is not limited in the embodiments of the present application.

[0109] In the embodiment of the present application, the computer device obtains the equipment data of the production equipment through which the sample product is produced from a local or database, or the computer device receives the equipment data sent by other computer devices or input by the user, which is not limited in the embodiment of the present application. The computer device inputs the equipment data into the yield prediction model.

[0110] 302. The computer device processes the device data through an encoder to obtain feature data corresponding to the device data, where the feature data includes features in multiple dimensions.

[0111] The feature data corresponding to the above-mentioned equipment data includes features in multiple dimensions, and the features in multiple dimensions indicate information related to the yield of the sample product.

[0112] In an embodiment of the present application, the computer device processes the input device data in sequence through multiple neural network layers in the encoder to obtain feature data corresponding to the device data.

[0113] 303. The computer device processes the feature data through a decoder and outputs a prediction result corresponding to the sample product, where the prediction result includes at least one defect type corresponding to the sample product and a yield rate corresponding to the defect type.

[0114] Among them, the prediction result includes at least one defective type corresponding to the sample product and the yield corresponding to the defective type, and each defective type indicates a reason why the sample product is judged to be defective. Taking the display panel as an example, the defective type can be uneven brightness, dark spots on the display panel, foreign matter such as dust adsorbed on the display panel, or scratches on the display panel. The sample product includes multiple sub-sample products, and the yield corresponding to the defective type refers to the proportion of sub-sample products that do not have the corresponding problem of the defective type in the multiple sub-sample products included in the sample product. Taking the display panel as an example, a display panel includes multiple sub-display panels, and the multiple sub-display panels indicate different areas on the display panel, or the display panel is spliced ​​by the multiple sub-display panels, which is not limited in the embodiment of the present application. The yield corresponding to the presence of dark spots on the display panel refers to: the proportion of sub-display panels that do not have dark spots in the display panel among the above-mentioned multiple sub-display panels.

[0115] In an embodiment of the present application, the computer device processes the above-mentioned feature data in sequence through multiple neural network layers in the decoder, and outputs the processing results as prediction results corresponding to the above-mentioned sample products.

[0116] 304. The computer device adjusts the yield prediction model based on the prediction results and actual results corresponding to the sample products to obtain a trained yield prediction model.

[0117] Among them, the actual results corresponding to the sample products are pre-set. If the sample products are products that have been produced in the production workshop, the actual results include the actual defective types corresponding to the sample products and the yield corresponding to the defective types. If the sample products are fictitious products, the actual results include the defective types set by relevant staff based on historical experience and the yield corresponding to the defective types.

[0118] In an embodiment of the present application, a computer device obtains a loss value corresponding to a yield prediction model based on the predicted results and actual results corresponding to a sample product, and performs back-propagation adjustment on the yield prediction model based on the loss value to obtain a trained yield prediction model.

[0119] The training method of the yield prediction model provided in the embodiment of the present application processes the equipment data of the production equipment through which the sample product is produced through the encoder and decoder in the yield prediction model, and adjusts the model according to the prediction results output by the model and the actual results corresponding to the sample product. Since the encoder in the model can compress high-dimensional data, thereby extracting effective features of the high-dimensional data from the high-dimensional data, the trained encoder can process the high-dimensional equipment data to extract effective feature data therefrom. The decoder trained in the model can process the above-mentioned effective feature data to obtain accurate prediction results. According to the accurate prediction results, the production process is adjusted before the product is produced, which can improve the efficiency of adjusting the production process, thereby improving the production efficiency of the product.

[0120] Above Figure 3 The content shown in FIG. 1 briefly describes the training process of the yield prediction model. Before training the yield prediction model, Figure 4 As shown, data collection is also required, that is, obtaining training data corresponding to the yield prediction model. The process of obtaining training data can be performed by the computer device in the embodiment of the present application, or by other computer devices, and the embodiment of the present application does not limit this. The following takes the computer device in the embodiment of the present application to obtain training data as an example to illustrate the process of obtaining training data. Figure 5 is a flow chart of a method for obtaining training data provided in an embodiment of the present application. Figure 5 As shown, the method includes the following steps.

[0121] 501. The computer device obtains product transit data corresponding to the preset product type, equipment parameter data and detection data of the production workshop.

[0122] Among them, product transit data, equipment parameter data and detection data are all data exported by the manufacturing system corresponding to the production workshop, which can be data in CSV (Comma-Separated Values) format or TXT (Text Document) format, which is not limited in the embodiments of the present application. The production workshop can be used to produce products of different product types, each product type corresponds to a production process, each production process includes multiple production sites, and different production processes include different production sites. Each production site will use at least one production equipment, and each production equipment corresponds to at least one equipment parameter. The above-mentioned production site can be a detection site or other types of sites, which is not limited in the embodiments of the present application.

[0123] The above-mentioned product transit data is used to record the products passing through each production site in the production process corresponding to the above-mentioned preset product type and the transit information of these products, such as the production equipment used by the product at the production site, etc. The following Table 1 is an example of product transit data provided in an embodiment of the present application, and Table 1 is also called a product transit resume table. As shown in Table 1, the product transit data includes the batch ID (lot id) of the batch to which the product belongs, the product ID (glassid) of the product, the site ID (stepid) of the production site passed by the product, the equipment ID (eqpid) of the production equipment used by the product at the production site, and the exit time (end_time) of the product.

[0124] Table 1

[0125] lotid glassid step eqp end_time lot1 glass1.1 step1 eqp1 time1 lot1 glass1.2 step1 eqp2 time2 lot2 glass2.1 step2 eqp1 time3 … … … … …

[0126] The above-mentioned equipment parameter data is used to record the parameters used by each production equipment in the production workshop when producing different products. The following Table 2 is an example of equipment parameter data provided in an embodiment of the present application. As shown in Table 2, the equipment parameter data includes the batch identifier (lotid) of the batch to which the product belongs, the product identifier (glassid) of the product, the equipment identifier (eqpid) of the production equipment used by the product, the parameter type (param_type) used by the production equipment, the equipment parameter (param_value) used by the production equipment, and the production time (p_time) of the product.

[0127] Table 2

[0128] lotid glassid eqp param_type param_value p_time lot1 glass1.1 eqp1 param1 value1 time4 lot1 glass1.2 eqp2 param2 value2 time5 lot2 glass2.1 eqp3 param3 value3 time6 … … … … … …

[0129] The above-mentioned detection data is used to record the detection results of each product belonging to the above-mentioned preset product type, and the detection results include the site identification of the production site through which the product passes, the defective type corresponding to the product at the production site, the yield corresponding to the defective type, and the grade of the product, and the grade of the product includes defective products and good products. The following table 3 is an example of a detection data provided by an embodiment of the present application. As shown in Table 3, the detection data includes the batch identification (lotid) of the batch to which the product belongs, the product identification (glassid) of the product, the site identification (stepid) of the production site through which the product passes, the defective type (defect) corresponding to the product at the production site, the yield corresponding to the defective type, and the product grade (grade) of the product. Wherein, the yield is also referred to as the detection rate, and the product grade of the product includes defective products (N) and good products (G). In some embodiments, the detection data also includes the yield corresponding to the batch product, that is, the proportion of good products in all products in the same batch of products. If the computer device detects that the yield corresponding to the batch product is less than the yield threshold, the data corresponding to the batch product is deleted in the product transit data and the equipment parameter data.

[0130] Table 3

[0131] lotid glassid step grade defect ratio lot1 glass1.1 step1 N defect1 ratio1 lot1 glass1.2 step1 G defect2 ratio2 lot2 glass2.1 step3 G defect3 ratio3 … … … … … …

[0132] In an embodiment of the present application, the computer device obtains the above-mentioned production data and test data locally, or the computer device receives the production data and test data sent by other computer devices.

[0133] In some embodiments, a model training system is installed on the computer device, and the production data and test data are obtained through the model training system. In the process of obtaining the production data and test data, the computer device displays the client interface corresponding to the model training system. Figure 6 is an example diagram of a client interface provided in an embodiment of the present application, such as Figure 6 As shown, a training data import control 601 is displayed on the client interface 600 , and the computer device displays a training data import panel in response to a triggering operation on the training data import control 601 . Figure 7 is an example diagram of a training data import panel provided in an embodiment of the present application, such as Figure 7 As shown, a file selection control 701 is displayed on the training data import panel 700. The computer device transmits the locally stored production data and test data to the model training system in response to the triggering operation of the file selection control 701. The computer device then extracts training data from the production data and test data through the model training system, and trains the yield detection model according to the training data.

[0134] In some embodiments, the computer device displays a user login interface before displaying the above-mentioned client device, such as Figure 8 As shown, the user login interface 800 displays a user name input box 801 and a password input box 802. The computer device responds to the input operation of the user name input box and the input operation of the password input box, detects that the input user name is correct and the input password is the correct password corresponding to the input user name, and then the computer device displays the above-mentioned client interface. The user is verified through the user login interface, which can ensure the security of model training. In some embodiments, the above-mentioned login interface is a network page, and the computer device responds to the input operation of the login interface address and displays the login interface.

[0135] The above process is explained by taking a computer device as both a client device and a backend service device of a model training system as an example. In some embodiments, a model training system is installed on the above-mentioned computer device, which serves as a backend service device of the model training system. Other computer devices serve as client devices of the model training system, and obtain production data and test data through the above-mentioned client interface, and send the production data and test data to the above-mentioned computer device serving as a backend service device.

[0136] 502. The computer device extracts the equipment parameter data of the production workshop based on the product passing data corresponding to the preset product type, and obtains the sample input data in the training data. The sample input data includes the equipment data of the production equipment passed by multiple sample products belonging to the preset product type.

[0137] Among them, the data corresponding to a sample product in the sample input data is the equipment data of the sample product, including the equipment parameters used by multiple production equipment in the production process when producing the sample product. The following Table 4 is an example of sample input data provided by an embodiment of the present application. As shown in Table 4, the sample input data includes the batch identifier (lotid) of the batch to which the sample product belongs, the product identifier (glassid) of the sample product, and the equipment parameters (value) under multiple parameter types (param).

[0138] Table 4

[0139] lotid glassid param1 param2 … lot1 glass1.1 value1 value4 … lot1 glass1.2 value3 value2 … … … … … …

[0140] In an embodiment of the present application, the computer device takes the products included in the product transit data as sample products, and based on the product identification of each sample product, extracts the equipment parameters used by multiple production equipment in the production process when producing the sample product from the equipment parameter data, and adds the equipment parameters used when producing the sample product to the sample input data of the training data according to the correspondence between the equipment parameters and the parameter types.

[0141] 503. The computer device extracts the above-mentioned inspection data based on the preset defect type and the above-mentioned product passing data to obtain sample output data in the training data, and the sample output data includes actual results corresponding to multiple sample products.

[0142] The preset defect type is preset and used as a benchmark for the prediction result output by the yield prediction model, that is, it is used to limit the prediction result output by the yield prediction model so that the defect types in the prediction result all belong to the preset defect type. In some embodiments, the preset defect type is manually maintained by relevant staff, and the setting process of the preset defect type includes: Figure 6 As shown, the client interface 600 also displays an unhealthy type import control 602. In response to the triggering operation of the unhealthy type import control, the computer device displays an unhealthy type import panel. Fig. 9 is an example diagram of a bad type import panel provided in an embodiment of the present application, such as Fig. 9 As shown, a file selection control 901 is displayed on the bad type import panel 900. In response to a trigger operation on the file selection control 901, the computer device transmits a file containing a preset bad type to the above-mentioned model training system, so that the model training system can obtain training data through the preset bad type.

[0143] The data corresponding to a sample product in the above sample output data includes the defect type corresponding to the sample product and the yield corresponding to the defect type, and the defect type belongs to a preset defect type. The following Table 5 is an example of sample output data provided by an embodiment of the present application. As shown in Table 5, the sample output data includes the batch identifier (lotid) of the batch to which the sample product belongs, the product identifier (glassid) of the sample product, the defect type (defect) corresponding to the sample product, and the yield (ratio) corresponding to the defect type.

[0144] Table 5

[0145] lotid glassid defect ratio lot1 glass1.1 defect1 ratio1 lot1 glass1.2 defect2 ratio2 … … … …

[0146] In an embodiment of the present application, the computer device takes the products included in the product passing data as sample products, and based on the preset defective type and the product identification of each sample product, extracts the defective type and the yield corresponding to the defective type that belong to the above-mentioned preset defective type and corresponds to the sample product from the inspection data, and adds the defective type and the corresponding yield to the sample output data of the training data.

[0147] The above steps 502 to 503 are the process of computer equipment integrating product transit data, equipment parameter data and detection data to obtain training data. This process uses the product identifier in the product transit data as the associated primary key. Based on the associated primary key, the data corresponding to the products in the above preset product types can be found from the equipment parameter data and the detection data, and the data integration efficiency is relatively high.

[0148] In some embodiments, Figure 4 and Figure 2 As shown, after the computer device obtains the above training data, it first pre-processes the training data, and then trains the yield prediction model based on the training data. In some embodiments, the process of pre-processing the training data by the computer device includes: for any sample product, in the sample output data corresponding to the sample product, the computer device detects that the yield of the sample product is less than the yield threshold, then discards the device data of the sample product, that is, deletes the data corresponding to the sample product in the sample input data. Among them, the yield of the sample product is less than the yield threshold, which can be that there are a preset number of yields less than the yield threshold in the yield corresponding to the sample product, or the yield corresponding to the sample product is less than the yield threshold, which is not limited in the embodiment of the present application. The yield thresholds corresponding to different defective types can be the same or different, which is not limited in the embodiment of the present application. In some embodiments, when the above detection data also includes the yield corresponding to the batch product, the computer device detects that the yield corresponding to the batch product is less than the yield threshold, then deletes the data corresponding to the batch product in the sample input data.

[0149] In some embodiments, the process of preprocessing the training data by the computer device includes: for any sample product, in the device data of the sample product, if the computer device detects that the number of missing device parameters in the device data of the sample product is greater than or equal to the missing threshold, the device data of the sample product is discarded. For example, if the missing threshold is 50%, and the computer device detects that more than half of the device parameters in the device data of product glass1.1 are null values, the device data of product glass1.1 is discarded.

[0150] The above process is explained by taking the example of discarding the device data that does not meet the conditions after the computer device obtains the training data. In some embodiments, the computer device first deletes the data that does not meet the conditions from the product passing data, device parameter data and detection data in the above manner, and then obtains the training data based on the product passing data, device parameter data and detection data. The embodiments of the present application do not limit this. For example, if the computer device detects a product with a yield less than the yield threshold in the detection data, the data corresponding to the product is deleted from the device parameter data and the product passing data. The training data is obtained based on the deleted product passing data, device parameter data and detection data.

[0151] The above process can improve the data quality of training data and avoid overfitting of the yield prediction model by filtering out data that does not meet the conditions (invalid data).

[0152] In some embodiments, the process of preprocessing the training data by the computer device includes: taking device parameters belonging to the same parameter type as an example, that is, taking a column of device parameters in the sample input data as an example, the computer device supplements the missing device parameters in the column of device data, for example, filling the missing device parameters with zero, that is, zero filling processing. For another example, the computer device calculates the device parameter mean of the device parameters in the column of data, and fills the missing device parameters with the corresponding device parameter mean. The embodiment of the present application does not limit the above supplementation method.

[0153] In some embodiments, the process of preprocessing the training data by the computer device includes: the computer device normalizes the device data of the sample product. The method of normalizing the device data includes: taking the normalization of device parameters belonging to the same parameter type as an example, that is, taking the normalization of a column of device parameters in the sample input data as an example, the computer device can process the device parameters in the column of device data through the following formula (1) to obtain the normalized device parameters.

[0154]

[0155] Among them, x ′ is the normalized device parameter, x is the device parameter before normalization, and x min is the minimum value in the device data of this column, x max is the maximum value in the device data of this column. The minimum value in the device data of this column is the minimum value before the missing device parameters are supplemented.

[0156] In some embodiments, for device parameters that have been zero-padded, the process of normalizing the device parameters by a computer device includes: the computer device subtracts the device parameter from the mean of the above-mentioned device parameters to obtain the difference between the device parameter and the mean of the device parameters, and divides the difference by the standard deviation of the device parameters under the corresponding parameter type to obtain the normalized device parameters mentioned above.

[0157] In some embodiments, the computer device can also normalize the device data of the sample product through other calculation methods, which is not limited in this embodiment of the present application.

[0158] Through the above Figure 5 After obtaining the training data in the manner shown, the computer device can train the yield prediction model through the training data. The training process of the yield prediction model includes multiple rounds of training. Figure 3 The content shown briefly describes the training process of the yield prediction model. Fig.10 This is a flowchart of another method for training a yield prediction model provided in an embodiment of the present application. Fig.10 , a round of training process of the yield prediction model is described in detail.

[0159] 1001. The computer device obtains the equipment data of the production equipment used in the production of the sample product. The equipment data includes equipment parameters of multiple dimensions. The equipment parameters are the parameters used by the production equipment when producing the sample product.

[0160] In the embodiment of the present application, the computer device obtains the equipment data of the production equipment through which the sample product is produced from the above sample input data. The process is described by taking the example that the computer device obtains the equipment data corresponding to a sample product from the above sample input data during a round of training. During the round of training, the computer device trains the model through the equipment data corresponding to the sample product. The training process consumes less computing power and the training speed is faster.

[0161] In some embodiments, during a round of training, the computer device obtains device data corresponding to a preset number of sample products from the above sample input data. During the above round of training, the computer device trains the model through the device data corresponding to the preset number of sample products, and the device data corresponding to the preset number of sample products jointly determine the direction of adjusting the model, which can reduce the randomness of model training, improve the accuracy of model training, and thus improve the efficiency of model training. The above preset number of sample products can be multiple sample products from the same batch, or multiple sample products from different batches, which is not limited in the embodiments of the present application.

[0162] 1002. The computer device inputs the equipment data of the production equipment through which the sample product is produced into the yield prediction model, and the yield prediction model includes an encoder and a decoder.

[0163] In some embodiments, Figure 4 and Figure 2 As shown, the computer device converts the above device data into a two-dimensional matrix (matrixization), and inputs the matrix corresponding to the device data into the yield prediction model. In a round of training, if the training process uses the device data corresponding to a sample product for training, the matrix corresponding to the device data is an m×1 matrix, where m is an integer greater than 0, indicating m device parameters. If the training process uses the device data corresponding to multiple sample products for training, the matrix corresponding to the device data is an m×n matrix, where n is an integer greater than 1, indicating n sample products.

[0164] 1003. The computer device processes the device data through an encoder to obtain feature data corresponding to the device data, where the feature data includes features in multiple dimensions.

[0165] Among them, Figure 2 As shown, the yield prediction model includes a first network module and a second network module. The first network module and the second network module have the same structure and both include the above encoder and decoder, which can reduce the complexity of model training. The encoder and decoder can be called a type of AE (Autoencoder) network architecture. The model parameters in the second network module have opposite signs to the corresponding model parameters in the first network module. For example, Figure 2 The model parameters of a neural network layer in the first network module are "w11, w12, ..., w1k", and the model parameters of the corresponding neural network layer in the second network module are "-w11, -w12, ..., -w1k". Accordingly, the above-mentioned feature data includes the feature data obtained by the first network module and the feature data obtained by the second network module.

[0166] In the embodiment of the present application, the computer device processes the device data through the encoder in the first network module to obtain the first feature data corresponding to the device data, and processes the device data through the encoder in the second network module to obtain the second feature data corresponding to the device data. The process of the computer device processing the device data through the encoder is the same as the process of processing the device data through the encoder in the above step 302. For example, Figure 2As shown, in the first network module and the second network module, "a1, a2, a3, ..., am" is the device data corresponding to a sample product in the input yield prediction model. The computer device obtains the feature data corresponding to the device data by multiplying the device data with the model parameters of the neural network layer in the encoder in sequence.

[0167] In the above process, processing the device data through the first network module is called forward processing, and processing the device data through the second network module is called reverse processing. The computer device performs forward processing and reverse processing on the device data respectively through the encoder in the first network module and the encoder in the second network module, which is equivalent to adding an additional network module for reverse processing on the basis of the original first network module, thereby enhancing the model's ability to extract features and being able to extract as many features as possible from the device data. Subsequently, model training is performed based on as many features as possible, which can improve the accuracy of model training and thus improve the efficiency of model training. The above-mentioned first network module and second network module are also set up to provide two input terminals for the yield prediction model, and these two input terminals correspond to different output contents respectively, so as to facilitate the following bilateral ReLU activation function to process the different output contents.

[0168] 1004. The computer device processes the feature data output by the encoder through the activation function in the yield prediction model to obtain processed feature data.

[0169] Among them, Figure 4 As shown, the activation function in the first network module and the activation function in the second network module in the yield prediction model are both bilateral ReLU functions, as shown in the following formula (2).

[0170] f(x)=max(0,x)-max(0,-x)(2)

[0171] Among them, x is the feature in the feature data. By processing the feature data through this activation function, the problem of unilateral response of the general ReLU function f(x)=max(0,x) can be avoided, and the convergence speed of model training can be improved. The unilateral response refers to the process of processing the feature data through the general ReLU function. The general ReLU function will map the features with negative values ​​to zero. When the loss function is subsequently derived and the model gradient is calculated based on the feature and the derived loss function, the obtained gradient is zero, making it impossible for the computer device to adjust the model according to the feature, that is, some features are lost during the model training process. By processing the feature data through the above-mentioned bilateral ReLU function, the traditional Sigmoid function (S-type function) f(x)=1 / (1+e -x) The problem of gradient vanishing in the model training process is solved to improve the effectiveness of model training. Gradient vanishing means that during the back propagation process, the gradient is transmitted back from the output layer to the input layer layer by layer. The gradient may gradually decrease due to the multiplication effect, resulting in a smaller gradient received by the front neural network layer. The adjustment based on the smaller gradient is limited, resulting in slow training of the front neural network layer.

[0172] In the present application embodiment, Figure 2 As shown, the computer device processes the first feature data and the second feature data respectively through the activation function in the first network module and the activation function in the second network module in the yield prediction model to obtain the processed first feature data and the processed second feature data.

[0173] 1005. The computer device processes the processed feature data through a decoder and outputs a prediction result corresponding to the sample product. The prediction result includes at least one defective type corresponding to the sample product and the yield corresponding to the defective type.

[0174] It can be seen from the above description that the processed feature data includes processed first feature data and processed second feature data.

[0175] In the embodiment of the present application, the computer device processes the first feature data through the decoder in the first network module to obtain the first prediction result corresponding to the sample product, and processes the second feature data through the decoder in the second network module to obtain the second prediction result corresponding to the sample product. Among them, the process of the computer device processing the feature data through the decoder is the same as the process of the computer device processing the feature data through the decoder in the above step 303, and the embodiment of the present application is not repeated here. The computer device outputs the sum of the first prediction result and the second prediction result as the prediction result corresponding to the sample product, that is, the computer device adds the yield corresponding to the same defective type in the first prediction result and the second prediction result to obtain the prediction result corresponding to the sample product, and outputs the prediction result.

[0176] The above process is explained by taking the example that the computer device outputs the sum of the first prediction result and the second prediction result as the prediction result corresponding to the sample product. In some embodiments, the computer device can also perform a weighted sum of the first prediction result and the second prediction result, and output the result of the weighted sum as the prediction result corresponding to the sample product. The embodiments of the present application are not limited to this.

[0177] In some embodiments, when a round of training is performed using device data corresponding to multiple sample products, the prediction results output by the decoder include prediction results corresponding to each of the multiple sample products.

[0178] 1006. The computer device obtains a loss value corresponding to this round of training process based on the prediction result output by the above decoder and the actual result corresponding to the sample product, and the loss value indicates the accuracy of the yield prediction model.

[0179] In the embodiment of the present application, the computer device calculates the predicted result output by the decoder and the actual result corresponding to the sample product through the loss function to obtain the loss value corresponding to the current round of training process. The loss function is preset, for example, the loss function can be the following formula (3):

[0180]

[0181] Among them, MSE1 is the loss value, m is the number of defective types corresponding to the sample product, and y j is the actual yield corresponding to the jth defect type in the actual result, It is the predicted yield corresponding to the j-th defect type in the prediction result.

[0182] The above process is explained by taking the example of obtaining the loss value based on the prediction result corresponding to a sample product. In some embodiments, the computer device obtains the loss value based on the prediction results and actual results corresponding to the preset sample number of sample products. That is, the computer device uses the small batch gradient descent method to train the model, which can speed up the model convergence speed. In this process, the computer device calculates the prediction results corresponding to the preset sample number of sample products and the actual results corresponding to the preset sample number of sample products through the loss function to obtain the loss value corresponding to this round of training process. Among them, the loss function is pre-set, for example, the loss function can be the following formula (4):

[0183]

[0184] Among them, MSE2 is the loss value, m is the number of defective types corresponding to the sample product, n is the preset number of samples, and y ij is the actual yield corresponding to the jth defect type under the i-th sample product in the actual result, It is the predicted yield corresponding to the j-th defect type under the i-th sample product in the prediction result.

[0185] In some embodiments, if the computer device detects that the number of training rounds of the yield prediction model is less than the preset number of rounds, then in one round of training, the training is performed based on the device data corresponding to the first sample number of sample products, and accordingly, the loss value is obtained based on the prediction results and actual results corresponding to the first sample number of sample products. If the computer device detects that the number of training rounds of the yield prediction model is greater than or equal to the preset number of rounds, the training is performed based on the device data corresponding to the second sample number of sample products, and accordingly, the loss value is obtained based on the prediction results and actual results corresponding to the second sample number of sample products. Wherein, the second sample number is less than the first sample number. For example, in the initial stage of training, the batch is 64, that is, the device data corresponding to 64 sample products are used for training in one round of training. As the training proceeds, the batch gradually decreases by 2 to the power of n until the batch is reduced to 8. In the above process, in the early stage of model training, the training is performed by using the device data corresponding to a large number of sample products, which can ensure that the computer device adjusts the model in the correct gradient descent direction. In the process of model training, gradually reducing the number of sample products corresponding to one round of training can reduce the amount of calculation of model training and improve the speed of model training.

[0186] In some embodiments, it is considered that reusing the data corresponding to some sample products to train the model will reduce the accuracy of the model prediction, such as Figure 4 As shown, the computer device introduces a repeated penalty term in the loss function, and calculates the loss value based on the repeated penalty parameter, the predicted results corresponding to the preset sample number of sample products, and the actual results, so as to reduce the impact of the repeated import of the device data corresponding to the same sample product on the model prediction accuracy. Among them, the repeated penalty parameter indicates the number of times the preset sample number of sample products are repeatedly used for model training. The following formula (5) is an example of a loss function introducing a repeated penalty term provided in an embodiment of the present application:

[0187]

[0188] Among them, MSE3 is the loss value, k is the repeated penalty term, m is the number of defective types corresponding to the sample product, n is the preset number of samples, and y ij is the actual yield corresponding to the jth defect type under the i-th sample product in the actual result, It is the predicted yield corresponding to the j-th defect type under the i-th sample product in the prediction result.

[0189] 1007. The computer device adjusts the yield prediction model based on the loss value for the next round of training.

[0190] In an embodiment of the present application, the computer device performs back-propagation adjustment on the yield prediction model based on the loss value, and performs the next round of training using the adjusted yield prediction model.

[0191] The above steps 1006 to 1007 are a possible implementation method for the computer device to adjust the yield prediction model based on the prediction results and actual results corresponding to the sample product. Fig.10 The process of training the model shown can be performed by a computer device through the above-mentioned model training system, and the embodiment of the present application does not limit this. The computer device obtains a trained yield prediction model through multiple rounds of training.

[0192] The training method of the yield prediction model provided in the embodiment of the present application processes the equipment data of the production equipment through which the sample product is produced through the encoder and decoder in the yield prediction model, and adjusts the model according to the prediction results output by the model and the actual results corresponding to the sample product. Since the encoder in the model can compress the high-dimensional data, thereby extracting the effective features of the high-dimensional data from the high-dimensional data, the trained encoder can process the high-dimensional equipment data to extract effective feature data therefrom. The decoder trained in the model can process the above-mentioned effective feature data to obtain accurate prediction results. According to the accurate prediction results, the production process is adjusted before the product is produced, which can improve the efficiency of adjusting the production process, thereby improving the production efficiency of the product. Yield prediction through the above-mentioned yield prediction model can also reduce the labor cost consumed in the process of analyzing the production process, saving the time consumed by the user to analyze the production process.

[0193] The above process describes in detail the training process of the yield prediction model. In addition to training the above yield prediction model, in some embodiments, the computer device can also test the trained yield prediction model through the test data corresponding to the test product. The process of acquiring the test data is the same as the process of acquiring the training data mentioned above, and the embodiments of the present application will not be repeated here. The computer device processes the test data through the yield prediction model to obtain the prediction results corresponding to the test product, and then determines the grade of the test product, such as good or defective products, according to the pre-set defect threshold. For example, the computer device uses the smallest yield in the prediction results of the test product as the yield of the test product. If the computer device detects that the yield of the test product is less than the defect threshold, the grade of the test product is determined to be a defective product. If the computer device detects that the yield of the test product is greater than or equal to the defect threshold, the grade of the test product is determined to be a good product. The following Fig.11 is an example diagram of a test product level provided in an embodiment of the present application. Fig.11In the figure, the ordinate is the yield rate, and each dot represents the yield rate of a tested product. Fig.11 As shown, the defect threshold is 0.7, the solid dots represent the test products determined to be good products (positive samples), and the hollow dots represent the test products determined to be defective products (negative samples).

[0194] In some embodiments, the computer device can obtain the accuracy of the yield prediction model's prediction by comparing the actual grade of the test product with the grade of the test product determined according to the prediction results. If the accuracy does not meet the preset standard, the model continues to be trained through the model training system. If the accuracy does not meet the preset standard, the yield prediction model is output.

[0195] In some embodiments, the above defect threshold setting process includes: Figure 6 As shown, the above client interface 600 also displays a defect threshold import control 603. In response to the triggering operation of the defect threshold import control, the computer device displays a defect threshold import panel. Fig.12 is an example diagram of a defect threshold import panel provided by an embodiment of the present application, such as Fig.12 As shown, an input control 1201 is displayed on the defect threshold import panel 1200 , and the computer device obtains the defect threshold in response to an input operation on the input control 1201 .

[0196] In some embodiments, after obtaining a trained yield prediction model, the computer device can use the yield prediction model to extract features of equipment parameters of different dimensions in the equipment data of the production equipment before product production, and further predict the yield of the product. According to the predicted yield, the production equipment is adjusted to improve the yield, reduce the probability of producing defective products, and improve production efficiency.

[0197] Accordingly, the computer device inputs the equipment data of the production equipment through which the product is produced into the yield prediction model. The equipment data includes equipment parameters of multiple dimensions. The equipment parameters are parameters used by the production equipment when producing the product. The yield prediction model includes an encoder and a decoder. The computer device processes the equipment data through the encoder to obtain feature data corresponding to the equipment data. The feature data includes features of multiple dimensions. The computer device processes the feature data through the decoder and outputs the prediction result corresponding to the product. The prediction result includes at least one defective type corresponding to the product and the yield corresponding to the defective type.

[0198] In some embodiments, before inputting the equipment data of the production equipment through which the product is produced into the yield prediction model, the computer device normalizes the equipment data of the product to reduce the amount of computing power consumed by the computer device in processing the equipment data and improve prediction efficiency.

[0199] In some embodiments, the yield prediction model includes a first network module and a second network module. In the process of processing the device data through the encoder to obtain feature data corresponding to the device data, the computer device processes the device data through the encoder in the first network module to obtain first feature data corresponding to the device data, and processes the device data through the encoder in the second network module to obtain second feature data corresponding to the device data. The model parameters in the second network module have opposite signs to the corresponding model parameters in the first network module.

[0200] In some embodiments, when the yield prediction model includes a first network module and a second network module, in the process of processing the feature data through a decoder and outputting a prediction result corresponding to the product, the computer device processes the first feature data through the decoder in the first network module to obtain a first prediction result corresponding to the product, processes the second feature data through the decoder in the second network module to obtain a second prediction result corresponding to the product, and outputs the sum of the first prediction result and the second prediction result as the prediction result corresponding to the product.

[0201] In the process of adjusting the production equipment according to the predicted yield, for any sample product, the computer equipment can obtain the impact value of each equipment parameter on the yield based on the equipment data and prediction results corresponding to the sample product and historical experience, so that the computer equipment can locate the root cause of the defect, improve the efficiency of analyzing the production process, and improve the detection efficiency and maintenance efficiency. Fig.13 is an example diagram of the impact of a device parameter on yield provided by an embodiment of the present application, Fig.13 In the figure, the horizontal axis is the yield of the sample product, the vertical axis is the equipment parameters of different parameter types, and a broken line corresponds to a sample product. Fig.13 As shown, on a broken line, the difference in yields corresponding to two adjacent device parameters indicates the impact value of the upper device parameter among the two device parameters on the yield.

[0202] Above Fig.11 and Fig.13 It is a visual display of the prediction results output by the yield prediction model. In some embodiments, the computer device can provide a visualization interface for the user to view the prediction results output by the yield prediction model.

[0203] Fig.14 is a structural block diagram of a yield prediction model training device provided in an embodiment of the present application. The device is used to execute the steps of the yield prediction model training method described above, see Fig.14 , the device comprises:

[0204] An input module 1401 is used to input the equipment data of the production equipment used in the production of the sample product into the yield prediction model, where the equipment data includes equipment parameters of multiple dimensions, and the equipment parameters are parameters used by the production equipment when producing the sample product. The yield prediction model includes an encoder and a decoder;

[0205] A first processing module 1402 is used to process the device data through an encoder to obtain feature data corresponding to the device data, where the feature data includes features of multiple dimensions;

[0206] The second processing module 1403 is used to process the feature data through a decoder and output a prediction result corresponding to the sample product, where the prediction result includes at least one defect type corresponding to the sample product and a yield rate corresponding to the defect type;

[0207] The adjustment module 1404 is used to adjust the yield prediction model based on the prediction results and actual results corresponding to the sample products to obtain a trained yield prediction model.

[0208] In some embodiments, the above apparatus further comprises:

[0209] The preprocessing module is used to preprocess the device data of the sample product, and the preprocessing includes at least one of the following:

[0210] If it is detected that the yield of the sample product is less than the yield threshold, the device data of the sample product is discarded;

[0211] If it is detected that the number of missing device parameters in the device data of the sample product is greater than or equal to the missing threshold, the device data of the sample product is discarded;

[0212] Supplement the missing equipment parameters in the equipment data of sample products;

[0213] Normalize the device data of the sample products.

[0214] In some embodiments, the yield prediction model includes a first network module and a second network module, both of which include encoders, and the first processing module 1402 includes:

[0215] A first feature data acquisition unit, configured to process the device data through an encoder in the first network module to obtain first feature data corresponding to the device data;

[0216] The second characteristic data acquisition unit is used to process the device data through the encoder in the second network module to obtain second characteristic data corresponding to the device data. The model parameters in the second network module have opposite positive and negative signs to the corresponding model parameters in the first network module.

[0217] In some embodiments, the first network module and the second network module both include a decoder, and the second processing module 1403 is used to:

[0218] Processing the first feature data through a decoder in the first network module to obtain a first prediction result corresponding to the sample product;

[0219] Processing the second feature data through a decoder in the second network module to obtain a second prediction result corresponding to the sample product;

[0220] The sum of the first prediction result and the second prediction result is output as the prediction result corresponding to the sample product.

[0221] In some embodiments, the training process of the yield prediction model includes multiple rounds of training, and the adjustment module 1404 includes:

[0222] A loss value acquisition unit, used to acquire a loss value based on the prediction results and actual results corresponding to a preset number of sample products, wherein the loss value indicates the accuracy of the yield prediction model prediction;

[0223] The adjustment unit is used to adjust the yield prediction model based on the loss value to perform a next round of training.

[0224] In some embodiments, the loss value obtaining unit is used to:

[0225] The loss value is calculated based on the repetition penalty parameter, the predicted results corresponding to the preset number of sample products, and the actual results. The repetition penalty parameter indicates the number of times the preset number of sample products are repeatedly used for model training.

[0226] In some embodiments, the loss value obtaining unit is used to:

[0227] If it is detected that the number of training rounds of the yield prediction model is less than the preset number of rounds, a loss value is obtained based on the prediction results and actual results corresponding to the first number of sample products;

[0228] If it is detected that the number of training rounds of the yield prediction model is greater than or equal to the preset number of rounds, the loss value is obtained based on the predicted results and actual results corresponding to the second sample number of sample products, and the second sample number is less than the first sample number.

[0229] It should be noted that: when the device provided in the above embodiment performs model training, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiment belong to the same concept, and the specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0230] Fig.15 It is a structural diagram of a computer device provided in an embodiment of the present application. The computer device 1500 may have relatively large differences due to different configurations or performances, and may include one or more CPUs (Central Processing Units, processors) 1501 and one or more memories 1502, wherein the memory 1502 stores at least one computer program, and the at least one computer program is loaded and executed by the processor 1501 to implement the yield prediction model training method or yield prediction method provided in the above-mentioned various method embodiments. Of course, the computer device may also have components such as a wired or wireless network interface, a keyboard, and an input and output interface for input and output. The computer device may also include other components for realizing the functions of the device, which will not be repeated here.

[0231] The embodiment of the present application also provides a computer-readable storage medium, in which at least one computer program is stored, and the at least one computer program is loaded and executed by a processor of a computer device to implement the training method of the yield prediction model of the above embodiment or the operation performed by the computer device in the yield prediction method. For example, the computer-readable storage medium can be ROM (Read-Only Memory), RAM (Random Access Memory), CD-ROM (Compact Disc Read-Only Memory), magnetic tape, floppy disk and optical data storage device, etc.

[0232] The embodiment of the present application also provides a computer program product or a computer program, which includes a computer program code, and the computer program code is stored in a computer-readable storage medium. The processor of the computer device reads the computer program code from the computer-readable storage medium, and the processor executes the computer program code, so that the computer device executes the yield prediction model training method or yield prediction method provided in the above various optional implementations.

[0233] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0234] The above description is only an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for training a yield prediction model, characterized in that: The method comprises: Inputting equipment data of the production equipment used in the production of the sample product into the yield prediction model, wherein the equipment data includes equipment parameters of multiple dimensions, and the equipment parameters are parameters used by the production equipment when producing the sample product, and the yield prediction model includes an encoder and a decoder; Processing the device data by the encoder to obtain feature data corresponding to the device data, wherein the feature data includes features of multiple dimensions; Processing the feature data through the decoder to output a prediction result corresponding to the sample product, wherein the prediction result includes at least one defect type corresponding to the sample product and a yield rate corresponding to the defect type; Based on the prediction results and actual results corresponding to the sample products, the yield prediction model is adjusted to obtain a trained yield prediction model.

2. The method according to claim 1, characterized in that Before inputting the equipment data of the production equipment through which the sample product is produced into the yield prediction model, the method further includes: Preprocessing the device data of the sample product, wherein the preprocessing includes at least one of the following: If it is detected that the yield of the sample product is less than the yield threshold, the device data of the sample product is discarded; If it is detected that the number of missing device parameters in the device data of the sample product is greater than or equal to a missing threshold, the device data of the sample product is discarded; Supplement the missing equipment parameters in the equipment data of the sample product; The device data of the sample product is normalized.

3. The method according to claim 1, characterized in that The yield prediction model includes a first network module and a second network module, the first network module and the second network module both include an encoder, and the encoder is used to process the device data to obtain feature data corresponding to the device data, including: Processing the device data through an encoder in the first network module to obtain first feature data corresponding to the device data; The device data is processed by an encoder in the second network module to obtain second characteristic data corresponding to the device data, and the model parameters in the second network module have opposite signs to the corresponding model parameters in the first network module.

4. The method according to claim 3, characterized in that The first network module and the second network module both include a decoder, and the processing of the feature data by the decoder to output a prediction result corresponding to the sample product includes: Processing the first feature data through a decoder in the first network module to obtain a first prediction result corresponding to the sample product; Processing the second feature data through a decoder in the second network module to obtain a second prediction result corresponding to the sample product; The sum of the first prediction result and the second prediction result is output as the prediction result corresponding to the sample product.

5. The method according to claim 1, characterized in that The training process of the yield prediction model includes multiple rounds of training, and the adjustment of the yield prediction model based on the prediction results and actual results corresponding to the sample products includes: Based on the prediction results and actual results corresponding to a preset number of sample products, a loss value is obtained, wherein the loss value indicates the accuracy of the yield prediction model prediction; Based on the loss value, the yield prediction model is adjusted to perform the next round of training.

6. The method according to claim 5, characterized in that The obtaining of the loss value based on the predicted results and actual results corresponding to the preset number of sample products includes: The loss value is calculated based on a repetition penalty parameter, the predicted results and actual results corresponding to the preset number of sample products, and the repetition penalty parameter indicates the number of times the preset number of sample products are repeatedly used for model training.

7. The method according to claim 5, characterized in that The obtaining of the loss value based on the predicted results and actual results corresponding to the preset number of sample products includes: If it is detected that the number of training rounds of the yield prediction model is less than the preset number of rounds, the loss value is obtained based on the prediction results and actual results corresponding to the first sample number of sample products; If it is detected that the number of training rounds of the yield prediction model is greater than or equal to the preset number of rounds, the loss value is obtained based on the prediction results and actual results corresponding to the second sample number of sample products, and the second sample number is less than the first sample number.

8. A training device for a yield prediction model, characterized in that: The device comprises: An input module, used to input the equipment data of the production equipment used in the production of the sample product into the yield prediction model, wherein the equipment data includes equipment parameters of multiple dimensions, and the equipment parameters are parameters used by the production equipment when producing the sample product. The yield prediction model includes an encoder and a decoder; A first processing module, configured to process the device data through the encoder to obtain feature data corresponding to the device data, wherein the feature data includes features of multiple dimensions; A second processing module, configured to process the feature data through the decoder and output a prediction result corresponding to the sample product, wherein the prediction result includes at least one defect type corresponding to the sample product and a yield rate corresponding to the defect type; The adjustment module is used to adjust the yield prediction model based on the prediction results and actual results corresponding to the sample products to obtain a trained yield prediction model.

9. A yield prediction method, characterized in that: The method comprises: Inputting equipment data of the production equipment used in the production of the product into the yield prediction model, wherein the equipment data includes equipment parameters of multiple dimensions, and the equipment parameters are parameters used by the production equipment when producing the product, and the yield prediction model includes an encoder and a decoder; Processing the device data by the encoder to obtain feature data corresponding to the device data, wherein the feature data includes features of multiple dimensions; The characteristic data is processed by the decoder to output a prediction result corresponding to the product, wherein the prediction result includes at least one defective type corresponding to the product and a yield corresponding to the defective type.

10. A computer device, characterized in that: The computer device comprises a processor and a memory, wherein the memory is used to store at least one computer program, and the at least one computer program is loaded by the processor to execute the method according to any one of claims 1 to 7.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store at least one computer program, and the at least one computer program is used to execute the method according to any one of claims 1 to 7.

12. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images