Softmax algorithm-based software design method for classifying silicon-based OLED (Organic Light Emitting Diode) badness

Abstract

The invention relates to the technical field of image classification, in particular to a softmax algorithm-based software design method for bad classification of silicon-based OLEDs (Organic Light Emitting Diodes), which comprises the following steps of: receiving silicon-based OLED production data uploaded by a user, indexing through date and process section information labels, executing grouping and sorting of the data, and generating sorted data groups. According to the method, sorting, processing and model training are carried out on silicon-based OLED production data through a softmax algorithm, the quality of training data is ensured according to noise data removal and damaged sample elimination, the generalization ability and classification accuracy of the model are enhanced, the training process is optimized through normalization processing and reasonable division of a data set, and the training efficiency is improved. According to the method, the algorithm can learn more effectively and reduce the risk of overfitting, the model training speed is accelerated through batch processing and weight updating, and the model performance is monitored in real time, so that the parameter optimization is finer, bad defects are classified quickly and accurately in actual production, and the production efficiency and the product quality are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image classification, and in particular to a software design method for classifying defective silicon-based OLEDs based on a softmax algorithm. Background Art

[0002] The field of image classification technology involves using algorithms to identify and classify the content in an image. It is widely used in the fields of computer vision and machine learning. The core goal is to enable computers to recognize objects, features, or conditions in an image. Image classification includes collecting image data, preprocessing images, extracting features, and using classification algorithms to learn and classify features. In modern applications, deep learning, especially convolutional neural networks (CNNs), is the most commonly used technology because of its excellent performance in image recognition tasks, driving progress in a variety of applications from medical imaging analysis to self-driving vehicles.

[0003] Among them, the software design method for classifying silicon-based OLED defects using the softmax algorithm refers to the use of the softmax algorithm to develop software to automatically classify defective images generated during the production process of silicon-based OLED displays. This method analyzes OLED screen image data to identify different types of defects, such as point defects and line defects, to help improve production quality and efficiency. The use of this technology can reduce the need for manual inspection and improve the automation and accuracy of the production line. It is suitable for electronic display equipment manufacturers that pursue high output and high quality.

[0004] Existing image classification technologies rely on more traditional image processing and classification methods to deal with the problem of poor classification in the production of silicon-based OLED displays. These methods have certain limitations when dealing with large-scale or complex data. For example, the lack of an effective data preprocessing mechanism causes the model to be affected by noise and erroneous data during training, reducing the classification accuracy. The failure to achieve dynamic normalization of data or adaptive parameter optimization results in the model's poor adaptability to new or unseen defect types, affecting the automation and accuracy of the entire production line. These technical limitations not only prolong the time for error detection, but also increase production costs. Especially in the field of electronic display equipment manufacturing with high quality requirements, this technical deficiency leads to reduced production efficiency and increased product defect rates. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings in the prior art and to propose a software design method for classifying silicon-based OLED defects based on a softmax algorithm.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a software design method for classifying silicon-based OLED defects based on a softmax algorithm, comprising the following steps:

[0007] S1: Receive the silicon-based OLED production data uploaded by the user, perform grouping and sorting of the data by date and process section information tag index, and generate sorted data groups;

[0008] S2: by grouping the sorted data, identifying and deleting noise and erroneous data, performing missing value filling and damaged sample removal, and performing normalization processing, dividing the data into a training set, a validation set, and a test set in proportion to obtain a processed data set;

[0009] S3: According to the processed data set, the loss function is set to the cross entropy loss, the gradient descent method is selected as the parameter optimization method, the basic structure of the model is constructed, and the input and output of the model are defined to construct a softmax linear regression model;

[0010] S4: using the training set to perform parameter initialization in the softmax linear regression model, using batch processing to input data batch by batch, calculating the loss and updating the weight, repeating the process until convergence, and generating model training monitoring results;

[0011] S5: Based on the model training monitoring results, the validation set is evaluated for performance, the impact of parameter configuration on model performance is analyzed, and the parameter configuration with the best model performance is determined by iteratively testing differentiated hyperparameter settings, thereby generating an optimized softmax linear regression model;

[0012] S6: deploy the optimized softmax linear regression model into the production environment, monitor the real-time production line data flow, classify the bad defects of silicon-based OLED, and import the data into the C# program to generate a real-time query solution.

[0013] As a further solution of the present invention, the sorted data grouping includes date index, process segment identifier and silicon-based OLED production data records under multiple identifiers, the processed data set includes data records with outliers removed, a complete data set with missing information completed, damaged data samples removed and normalized data ratios, the softmax linear regression model includes the number of input features, category labels, parameter weights and biases for classification, the model training monitoring results include loss values ​​of multiple rounds, parameter optimization progress and convergence speed, the optimized softmax linear regression model includes the selected learning rate, the amount of data processed and the application information of the regularization technology, and the real-time query solution includes the access point of the data stream, the key indicators for real-time monitoring and the format and frequency of data output.

[0014] As a further solution of the present invention, the steps of receiving silicon-based OLED production data uploaded by a user, performing grouping and sorting of the data by date and process section information tag index, and generating sorted data groups are specifically as follows:

[0015] S101: Based on the silicon-based OLED production data uploaded by the user, read the date and process section label of each data, classify the data according to the process section label, and sort each type of data into batches according to the date label to generate group records;

[0016] S102: performing a sorting operation on each group of data using the date field according to the grouping records, arranging each item of data in sequence by comparing the timestamps of the data records, checking that each group of data is arranged in chronological order, and generating a sorted data group;

[0017] S103: using the sorted data group, processing the OLED screen parameter data provided by the sensor and the image acquisition device through a C# program, sorting and storing the data, and generating sorted data groups.

[0018] As a further solution of the present invention, the steps of grouping the sorted data, identifying and deleting noise and erroneous data, performing missing value filling and damaged sample removal, and performing normalization processing, and dividing the data into a training set, a validation set, and a test set in proportion to obtain the processed data set are specifically as follows:

[0019] S201: According to the sorted data grouping, filter and remove records with invalid values ​​and null values, check the target field for each record, including the date and value, verify whether there are null values ​​and data that does not meet the input rules, and generate a clean data set;

[0020] S202: Based on the clean data set, identify abnormal points in the data, perform a range check on each numerical data item, exclude data points that exceed the range of conventional production parameters, determine and delete abnormal values ​​by setting thresholds, and generate a standard data set;

[0021] S203: Using the standard data set, adopting the Z score method, uniformly converting the date format, uniformly retaining two decimal places for numerical data, normalizing the data to standardize the data scale, dividing the data set into a training set, a validation set, and a test set according to the proportion, and generating a processed data set.

[0022] As a further embodiment of the present invention, the formula of the Z score method is as follows:

[0023]

[0024] Among them, z ′is the Z score of numerical data, x is the original value of each data point, μ is the mean of the corresponding data, σ is the standard deviation of the corresponding data, β is the data skewness adjustment coefficient, k is the scale adjustment coefficient, and α is the translation parameter.

[0025] As a further solution of the present invention, according to the processed data set, the loss function is set to cross entropy loss, the gradient descent method is selected as the parameter optimization method, the basic structure of the model is constructed, and the input and output of the model are defined. The steps of constructing the softmax linear regression model are specifically as follows:

[0026] S301: according to the processed data set, setting the number of nodes and activation function of the model input layer, analyzing the model weight and bias value, adjusting the weight range and the initialization scale of the bias, and generating an initialization parameter table;

[0027] S302: Based on the initialization parameter table, read the CSV data, adopt a data parsing strategy, convert the data in the CSV into a numerical type, and standardize the numerical value into a feature vector, control the converted data dimension and format, and generate a feature input table;

[0028] S303: According to the feature input table, the model is trained, the number of iterations and the learning rate are set, the feature vectors are input batch by batch, the learning rate is adjusted to optimize the convergence speed, and a softmax linear regression model is constructed.

[0029] As a further solution of the present invention, the training set is used to perform parameter initialization in the softmax linear regression model, batch processing is used to input data batch by batch, the loss is calculated and the weight is updated, and the process is repeated until convergence. The steps of generating the model training monitoring results are specifically as follows:

[0030] S401: performing parameter initialization in the softmax linear regression model using the training set, initializing the weights and biases of the model using normal distribution random values, and obtaining a parameter initialization configuration;

[0031] S402: according to the parameter initialization configuration, import data from the CSV file, convert the data into a feature vector format accepted by the model, calculate the output probability distribution of each data sample, and obtain a probability distribution result;

[0032] S403: Using the probability distribution result, calculate the loss value and update the parameters through the back propagation method, iterate until the model performance is stable, and output the model training monitoring result.

[0033] As a further solution of the present invention, based on the model training monitoring results, the performance of the validation set is evaluated, the influence of the parameter configuration on the model performance is analyzed, and the parameter configuration with the best model performance is determined by iteratively testing the differentiated hyperparameter settings. The steps of generating an optimized softmax linear regression model are specifically as follows:

[0034] S501: Analyze the impact of differentiated parameter configurations on model performance based on the model training monitoring results, use validation set data to evaluate the model performance, and generate initialization evaluation data;

[0035] S502: Based on the initialization evaluation data, design multiple sets of hyperparameter configurations, including differentiated combinations of learning rate and batch size, and obtain hyperparameter test results by iteratively testing the hyperparameter configurations;

[0036] S503: Using the hyperparameter test results, select the hyperparameter configuration with the best performance, retrain the model on the validation set and the training set, verify that the model achieves the optimal performance, and obtain an optimized softmax linear regression model.

[0037] As a further solution of the present invention, the optimized softmax linear regression model is deployed in a production environment, real-time production line data flow is monitored, and the defect classification of silicon-based OLED is performed, and the data is imported into a C# program. The steps of generating a real-time query solution are specifically as follows:

[0038] S601: configuring and starting the optimized softmax linear regression model on the C# program, setting the model interface to receive the real-time production line data stream, analyzing the silicon-based OLED data monitoring and classifying the bad defects, and obtaining the defect classification data stream;

[0039] S602: parse the defect classification data stream, and write the parsing results into the database in real time through SQL statements. The process includes data insertion and update operations, checking the continuous update of wafer status data, and obtaining an updated wafer status set;

[0040] S603: Using the updated wafer status set, a decision tree algorithm is used to perform data query and output real-time wafer defect classification information, and the C# server code is used to query the situation and provide a response to obtain a real-time query solution.

[0041] As a further solution of the present invention, the formula of the decision tree algorithm is as follows:

[0042]

[0043] Among them, C(i) is the bad classification index of the i-th wafer, x ijRepresents the data of the i-th Wafer at the j-th detection index, μ j represents the average value of the jth detection index, w j is the weight coefficient of the jth detection index, and n is the total number of detection indicators.

[0044] Compared with the prior art, the advantages and positive effects of the present invention are:

[0045] In the present invention, the silicon-based OLED production data is sorted, processed and model trained by the softmax algorithm, which brings significant efficiency improvement and quality control optimization. By automatically receiving and indexing data, the accuracy and ease of data management are improved, and the possibility of human error is reduced. In the processing stage, the quality of training data is ensured by removing noise data and damaging samples, and the generalization ability and classification accuracy of the model are enhanced. Normalization processing and reasonable division of data sets optimize the training process, which helps the algorithm learn more effectively and reduce the risk of overfitting. The speed of model training is accelerated by batch processing and weight updating, and the parameter optimization is made more refined by real-time monitoring of model performance, so that bad defects can be quickly and accurately classified in actual production, and production efficiency and product quality are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the main steps of the present invention;

[0047] Figure 2 It is a schematic diagram of the refinement of S1 of the present invention;

[0048] Figure 3 It is a schematic diagram of the refinement of S2 of the present invention;

[0049] Figure 4 It is a schematic diagram of the refinement of S3 of the present invention;

[0050] Figure 5 It is a schematic diagram of the refinement of S4 of the present invention;

[0051] Figure 6 It is a detailed schematic diagram of S5 of the present invention;

[0052] Figure 7 It is a detailed schematic diagram of S6 of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0054] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.

[0055] See also Figure 1 The present invention provides a technical solution, a software design method for classifying silicon-based OLED defects based on a softmax algorithm, comprising the following steps:

[0056] S1: Receive the silicon-based OLED production data uploaded by the user, perform grouping and sorting of the data by date and process section information tag index, and generate sorted data groups;

[0057] S2: By grouping the sorted data, identifying and deleting noise and erroneous data, filling missing values ​​and removing damaged samples for the remaining data, and normalizing them, the data is divided into training set, validation set and test set in proportion to obtain the processed data set;

[0058] S3: According to the processed data set, the loss function is set to cross entropy loss, the gradient descent method is selected as the parameter optimization method, the basic structure of the model is constructed, the input and output of the model are defined, and the softmax linear regression model is constructed;

[0059] S4: Use the training set to perform parameter initialization in the softmax linear regression model, use batch processing to input data batch by batch, calculate the loss and update the weights, repeat the process until convergence, and generate model training monitoring results;

[0060] S5: Based on the model training monitoring results, the validation set is evaluated for performance, the impact of parameter configuration on model performance is analyzed, and the parameter configuration with the best model performance is determined through iterative testing of differentiated hyperparameter settings, generating an optimized softmax linear regression model.

[0061] S6: Deploy the optimized softmax linear regression model into the production environment, monitor the real-time production line data flow, classify the defects of silicon-based OLEDs, and import the data into the C# program to generate real-time query solutions.

[0062] The sorted data grouping includes date index, process segment identifier and silicon-based OLED production data records under multiple identifiers. The processed data set includes data records with outliers removed, complete data sets with missing information completed, damaged data samples removed and normalized data ratios. The softmax linear regression model includes the number of input features, category labels, parameter weights and biases used for classification. The model training monitoring results include loss values ​​of multiple rounds, parameter optimization progress and convergence speed. The optimized softmax linear regression model includes the selected learning rate, the amount of data processed and the application information of the regularization technology. The real-time query solution includes the access point of the data stream, the key indicators for real-time monitoring and the format and frequency of data output.

[0063] See also Figure 2 , receiving the silicon-based OLED production data uploaded by the user, performing grouping and sorting of the data by date and process section information tag index, and generating the sorted data grouping steps are as follows:

[0064] S101: Based on the silicon-based OLED production data uploaded by the user, the date and process section label of each data is read, the data is classified according to the process section label, and each type of data is sorted into batches according to the date label, and the execution process of generating group records is as follows;

[0065] Based on the silicon-based OLED production data uploaded by users, according to the classification of data date and process segment labels, the date and process segment labels of each data are obtained by disassembling and extracting the contents of the original data files, and the data are labeled and managed separately. These data are classified and stored according to the process segment labels, and data batch processing operations are performed. Through the batch processing program, the data under each category is further organized and arranged according to the date label, ensuring the accuracy of the data and the convenience of subsequent operations, and generating group records.

[0066] S102: According to the group records, a sorting operation is performed on each group of data using the date field, and each data item is arranged in sequence by comparing the timestamps of the data records, and each group of data is checked to be arranged in chronological order. The execution flow of generating the sorted data group is as follows;

[0067] For the silicon-based OLED production data that has been classified and batch processed, the sorting operation of the date field is performed. The sorting program is executed on each group of data, and the data is sorted by comparing the timestamps of each data. The comparison mechanism and sorting algorithm in the sorting program are described in detail to ensure that each group of data is arranged in chronological order. Through this sorting process, it is possible to effectively verify whether the time series of the data group is correct, ensure the timeliness and accuracy of the data, achieve the expected data processing effect, and generate a sorted data group.

[0068] S103: Using the sorted data group, processing the OLED screen parameter data provided by the sensor and the image acquisition device through the C# program, sorting and storing the data, and generating the sorted data grouping execution process is as follows;

[0069] Use C# program to process OLED screen parameter data according to the formula Calculate the data grouping after sorting. In the formula, S o Represents the total data group after sorting, p o Represents the weight of each parameter, x o Represents the value of each data. Formula explanation and formula calculation process: Given multiple parameters of the OLED screen, such as brightness x1, color temperature x2, contrast x3, the corresponding weights are set to p1 = 0.3, p2 = 0.5, p3 = 0.2. Substitute these values ​​into the formula S o =0.3·x1+0.5·x2+0.2·x3. Set brightness x1=200, color temperature x2=6500, contrast x3=800, and calculate S o

[0070] =0.3·200+0.5·6500+0.2·800=3620, indicating that the total score of the sorted data group is 3620, which shows the efficiency and accuracy of the sorting program and ensures the quality and performance of data processing.

[0071] See also Figure 3 , by grouping the sorted data, identifying and deleting noise and erroneous data, performing missing value filling and damaged sample removal, and performing normalization, dividing it into training set, validation set and test set in proportion, the steps to obtain the processed data set are as follows:

[0072] S201: According to the sorted data grouping, filter and remove records with invalid values ​​and null values, check the target fields for each record, including date and value, and verify whether there are null values ​​and data that does not meet the input rules. The execution process of generating a clean data set is as follows;

[0073] According to the sorted data grouping, the operation of removing invalid values ​​and empty values ​​is performed, and the target fields of each record are checked, including date and numeric fields. The program automatically detects empty values ​​and data that does not meet the input rules in the records to ensure the detailed operations and execution standards of data purification processing, which directly affects the accuracy of subsequent data processing and analysis, and generates a pure data set.

[0074] S202: Based on the clean data set, identify the abnormal points in the data, perform a range check on each numerical data, exclude data points that exceed the range of regular production parameters, determine and delete abnormal values ​​by setting thresholds, and generate the standard data set. The execution process is as follows;

[0075] Based on the pure data set, the process of identifying outliers in the data is refined, a range check is performed on each numerical data, and the threshold is set according to the standard range of the production parameters. Data points that exceed the range are automatically excluded to ensure that each data point is within an acceptable range. By deleting these outliers, it is more in line with the requirements of production analysis, improves data processing efficiency and data quality, and generates a standardized data set.

[0076] S203: Using the standardized data set, using the Z score method, uniformly converting the date format, uniformly retaining two decimal places for numerical data, normalizing the data to standardize the data scale, dividing the data set into a training set, a validation set, and a test set according to the proportion, and generating the processed data set. The execution process is as follows;

[0077] The formula for the Z-score method is as follows:

[0078]

[0079] Among them, z ′ is the Z score of numerical data, x is the original value of each data point, μ is the mean of the corresponding data, σ is the standard deviation of the corresponding data, β is the data skewness adjustment coefficient, k is the scale adjustment coefficient, and α is the translation parameter.

[0080] Detailed explanation of the formula and the process of formula calculation: Consider a specific data set in which the numerical data is obtained through measurement. To ensure the reality of the example, the following steps and calculations are used:

[0081] The data set contains the following values: 18, 22, 15, 20, 16, 19, 21, 17, which reflect the measured values ​​of a specific monitoring indicator.

[0082] Calculate the mean μ and standard deviation σ of the data set, where the mean μ is the sum of all data values ​​divided by the number of data points, and the standard deviation σ is the square root of the sum of the squares of the differences between each data value and the mean divided by the number of data points.

[0083] Specifically calculate the mean and standard deviation:

[0084]

[0085] and

[0086]

[0087] Set the parameter β, which represents the adjustment of the impact of data skewness, which can be obtained by measuring the skewness of the data. Setting β = 0.5 means adjusting the skewness of the data distribution to reduce the impact of the skewness of the distribution on the standardized results.

[0088] k is the scale adjustment coefficient. The optimal value is determined by experiments or cross-validation methods. Here it is set to 1.5, which is a choice based on previous experience in similar data processing to enhance the impact of standardization.

[0089] α is a translation parameter, which is set to 0.5 and is used to adjust the center location of the data, which is determined based on the characteristics of the dataset and the expected data distribution location.

[0090] Substituting the values ​​into the calculation, we can calculate the improved Z score, considering x = 18:

[0091]

[0092] This result shows that the z calculated by the Z score method ′ The value is -0.17, which means that after the adjustment coefficient and translation parameter processing, the standardized value of data point 18 is closer to the center of the data set, demonstrating the precise control of the standardization process by adjusting the skewness and data scale.

[0093] See also Figure 4 , according to the processed data set, set the loss function to cross entropy loss, select gradient descent as the parameter optimization method, build the basic structure of the model, and define the input and output of the model. The specific steps of building a softmax linear regression model are:

[0094] S301: According to the processed data set, the number of nodes and activation function of the model input layer are set, the model weight and bias value are analyzed, the weight range and the initialization scale of the bias are adjusted, and the execution process of generating the initialization parameter table is as follows;

[0095] Set the number of nodes and activation function of the model input layer, and adjust the weight range and bias initialization scale according to the formula Calculate the initial weight W. Where W init represents the initial weight, n w Represents the number of input nodes. Formula explanation and formula calculation derivation process: Set the initial weight W init =0.01, input node number n w =100, then the weight adjustment formula This shows that each weight item is adjusted to 0.001, which ensures the stability of the model in the early stage of training and improves the learning efficiency.

[0096] S302: Based on the initialization parameter table, read the CSV data, adopt the data parsing strategy, convert the data in the CSV into a numerical type, and standardize the numerical value into a feature vector, control the converted data dimension and format, and generate the feature input table. The execution process is as follows;

[0097] Based on the initialization parameter table, a data parsing strategy is used to process the data in the CSV file. By writing a data conversion script, the string data in the CSV file is converted into a numerical type, and these numerical values ​​are standardized to control the converted data dimension and format. The process covers data cleaning, type conversion and normalization, ensuring the consistency and applicability of the data, providing an accurate input format for the model, and generating a feature input table.

[0098] S303: According to the feature input table, the model is trained, the number of iterations and the learning rate are set, the feature vectors are input batch by batch, the learning rate is adjusted to optimize the convergence speed, and the execution process of constructing the softmax linear regression model is as follows;

[0099] For the feature input table, perform model training operations, set the number of iterations and learning rate, input feature vectors batch by batch, and optimize the convergence speed of the model by dynamically adjusting the learning rate. The process involves the application of the learning rate adjustment strategy and the management of the iterative process to ensure the effectiveness of model training and the accuracy of model prediction, and to build a softmax linear regression model.

[0100] See also Figure 5 , use the training set to perform parameter initialization in the softmax linear regression model, use batch processing to input data batch by batch, calculate the loss and update the weights, repeat the process until convergence, and generate the model training monitoring results in the following steps:

[0101] S401: Perform parameter initialization in the softmax linear regression model using the training set, and use normal distribution random values ​​to initialize the weights and biases of the model. The execution flow of obtaining the parameter initialization configuration is as follows;

[0102] The training set is used to perform parameter initialization in the softmax linear regression model, and the weights and biases of the model are initialized using normal distributed random values. This process generates random normally distributed values ​​for each parameter by writing an initialization script to ensure that the initial values ​​of the parameters are not too concentrated, avoiding local minimum problems during training, laying the foundation for subsequent model training, enabling the model to start learning from a variety of starting points, increasing the comprehensiveness and efficiency of model training, and obtaining parameter initialization configuration.

[0103] S402: According to the parameter initialization configuration, data is imported from the CSV file, the data is converted into a feature vector format accepted by the model, and the output probability distribution of each data sample is calculated. The execution flow of obtaining the probability distribution result is as follows;

[0104] According to the parameter initialization configuration, data is imported from the CSV file, data processing is performed, and the raw data is converted into a feature vector format accepted by the model. The process includes data cleaning, format conversion, and normalization to ensure data standardization and consistency. The output probability distribution is calculated for each data sample. The key is to effectively convert the input data into a format that can be directly used by the model, thereby ensuring that the model can accurately calculate the predicted probability of each category and obtain the probability distribution result.

[0105] S403: Using the probability distribution results, calculating the loss value and updating the parameters through the back propagation method, iterating repeatedly until the model performance is stable, and outputting the model training monitoring results. The execution process is as follows;

[0106] The loss value is calculated using the probability distribution results, and the parameters are updated through the back propagation method according to the formula Calculate the loss L. In the formula, y u represents the one-hot encoding of the true label, p u Represents the predicted probability. Detailed explanation of the formula and the process of formula calculation: Set the training sample to m = 100. For a certain sample, the true category is the first category, y1 = 1 and y2 to y k All are 0, set the probability of the model predicting the first category to 0.8, and the remaining category probabilities are evenly divided into 0.2. Insert the loss formula to get L = -log (0.8) ≈ 0.223, which shows the uncertainty of the model's current prediction for this sample. This process is repeated until the model performance is stable, ensuring that the loss value gradually decreases with each iteration, and the effectiveness of model training is continuously monitored and improved.

[0107] See also Figure 6 Based on the model training monitoring results, the performance of the validation set is evaluated, the impact of parameter configuration on model performance is analyzed, and the parameter configuration with the best model performance is determined by iteratively testing differentiated hyperparameter settings. The specific steps for generating an optimized softmax linear regression model are as follows:

[0108] S501: According to the model training monitoring results, analyze the impact of differentiated parameter configurations on model performance, use the validation set data to evaluate the model performance, and generate the execution process of initialization evaluation data as follows;

[0109] Based on the model training monitoring results, differentiated parameter configuration analysis is performed. This process involves using validation set data to evaluate the model's performance. By writing data analysis scripts, various parameter configurations are tested to evaluate the specific impact on model performance. The key is to ensure the stability and accuracy of the model, provide scientific data support for further model optimization and parameter adjustment, and generate initial evaluation data.

[0110] S502: Based on the initial evaluation data, multiple sets of hyperparameter configurations are designed, including differentiated combinations of learning rate and batch size, and the hyperparameter configurations are iteratively tested to obtain the hyperparameter test results. The execution process is as follows;

[0111] Based on the initial evaluation data, we design multiple sets of hyperparameter configurations, including differentiated combinations of learning rate and batch size. We perform iterative tests on these hyperparameters and conduct multiple rounds of experiments to determine the best hyperparameter combination. The process includes not only the adjustment of hyperparameters, but also the recording and analysis of the results of each iteration. Through these iterative tests, we ensure that the hyperparameter settings that best suit the current model and data set are selected, and obtain the hyperparameter test results.

[0112] S503: Using the hyperparameter test results, select the hyperparameter configuration with the best performance, retrain the model on the validation set and the training set, verify that the model achieves the optimal performance, and obtain the optimized softmax linear regression model. The execution process is as follows;

[0113] Use the hyperparameter test results to select the best performing hyperparameter configuration, according to the formula Update the model parameters θ. In the formula, θ represents the model parameters, η represents the learning rate, Represents the gradient of the loss function. Detailed explanation of the formula and the derivation process of the formula calculation: Set the initial parameter θ old =0.05, learning rate η = 0.01, loss function gradient According to the update formula θ new

[0114] =0.05-0.01×0.1=0.049. In this way, the parameters are gradually optimized until the model performs best on the validation set and the training set. The optimal performance of the model is verified. The optimized softmax linear regression model has higher accuracy and generalization ability.

[0115] See also Figure 7 , deploy the optimized softmax linear regression model into the production environment, monitor the real-time production line data flow, classify the defects of silicon-based OLEDs, and import the data into the C# program. The specific steps to generate a real-time query solution are as follows:

[0116] S601: Configure and start the optimized softmax linear regression model on the C# program, set the model interface to receive the real-time production line data stream, analyze the silicon-based OLED data monitoring and classify the bad defects, and obtain the execution process of the defect classification data stream as follows;

[0117] Configure and start the optimized softmax linear regression model on the C# program, set the model interface to receive the real-time production line data stream. The process includes model deployment, interface configuration and data stream access. The program monitors and analyzes silicon-based OLED data, detects and classifies defects. In this way, the real-time processing and classification accuracy of the data are ensured, the monitoring efficiency of the production line and product quality control are improved, and the defect classification data stream is obtained.

[0118] S602: parse the defect classification data stream and write the parsing results to the database in real time through SQL statements. The process includes data insertion and update operations, and checks the continuous update of the wafer status data. The execution flow of the updated wafer status set is as follows;

[0119] Parse the defect classification data stream, and write the parsing results to the database in real time by writing SQL statements. This operation involves inserting and updating data to ensure the continuous updating of Wafer status data. The process not only improves the automation and real-time performance of data processing, but also ensures the accuracy and timeliness of data in the database, making the update of Wafer status sets more efficient and accurate, supporting subsequent production decisions and quality management, and obtaining updated Wafer status sets.

[0120] S603: Using the updated Wafer status set, a decision tree algorithm is used to query data and output real-time Wafer bad classification information. The C# server code is used to query the situation and provide a response. The execution process of the real-time query solution is as follows;

[0121] The formula for the decision tree algorithm is as follows:

[0122]

[0123] Among them, C(i) is the bad classification index of the i-th wafer, x ij Represents the data of the i-th Wafer at the j-th detection index, μ j represents the average value of the jth detection index, w j is the weight coefficient of the jth detection index, and n is the total number of detection indicators.

[0124] The meaning of each parameter of the formula and how to obtain it:

[0125] C(i) represents the defective classification index of the i-th wafer, which is obtained through the following calculation and reflects the degree of deviation of the wafer from the conventional performance indicators.

[0126] x ij It is the actual measured value of the jth inspection index of the i-th wafer, which is directly measured by the wafer inspection equipment.

[0127] μ j It is the historical average value of the j-th detection indicator, which is calculated based on the batch wafer data in a certain period in the past.

[0128] w j is the weight coefficient of the jth detection index, which is set according to the degree of influence of this index on product quality in historical data. The specific value is quantified by analyzing the correlation between historical failure data and product return rate.

[0129] n is the total number of detection indicators, which is a fixed parameter determined during the wafer production and detection process.

[0130] Let's set up a practical example for calculation:

[0131] A wafer is set with three detection indicators, and the historical average values ​​are μ1=100, μ2=150, and μ3=200. The actual measured value x i1 =102,x i2 =148,x i3 = 203. The weight coefficients are set as w1 = 0.5, w2 = 1.5, and w3 = 1.0. These weight coefficients reflect the influence of different indicators on the failure judgment and are coefficients obtained based on past failure analysis.

[0132] Calculation process:

[0133] Calculate the deviation and absolute value of each indicator:

[0134] |x i1 -μ1|=|102-100|=2

[0135] |x i2 -μ2|=|148-150|=2

[0136] |x i3 -μ3|=|203-200|=3

[0137] Compute the weighted sum and sum of squares of the deviations:

[0138]

[0139] and

[0140]

[0141] The bad classification index is calculated using the formula:

[0142] The results show that the wafer's bad classification index is 1.698, indicating the extent to which the wafer deviates from the average performance in conventional performance testing. This value is low, indicating that the performance of the wafer is closer to the expected quality standard and can continue to the subsequent production steps.

[0143] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A software design method for classifying silicon-based OLED defects based on a softmax algorithm, characterized in that: The following steps are involved: Receive the silicon-based OLED production data uploaded by the user, perform grouping and sorting of the data by date and process section information tag index, and generate sorted data groups; By grouping the sorted data, identifying and deleting noise and erroneous data, performing missing value filling and damaged sample removal, and performing normalization processing, the data is divided into a training set, a validation set, and a test set in proportion to obtain a processed data set; According to the processed data set, the loss function is set to cross entropy loss, the gradient descent method is selected as the parameter optimization method, the basic structure of the model is constructed, the input and output of the model are defined, and a softmax linear regression model is constructed; Using the training set to perform parameter initialization in the softmax linear regression model, using batch processing to input data batch by batch, calculating the loss and updating the weights, repeating the process until convergence, and generating model training monitoring results; Based on the model training monitoring results, the validation set is evaluated for performance, the impact of parameter configuration on model performance is analyzed, and the parameter configuration with the best model performance is determined by iteratively testing differentiated hyperparameter settings to generate an optimized softmax linear regression model; The optimized softmax linear regression model is deployed in the production environment to monitor the real-time production line data flow, classify the defects of silicon-based OLEDs, and import the data into the C# program to generate a real-time query solution.

2. The software design method for classifying defective silicon-based OLEDs based on the softmax algorithm according to claim 1, characterized in that: The sorted data grouping includes date index, process segment identifier and silicon-based OLED production data records under multiple identifiers; the processed data set includes data records with outliers removed, a complete data set with missing information completed, damaged data samples removed and normalized data ratios; the softmax linear regression model includes the number of input features, category labels, parameter weights and biases used for classification; the model training monitoring results include loss values ​​of multiple rounds, parameter optimization progress and convergence speed; the optimized softmax linear regression model includes the selected learning rate, the amount of data processed and the application information of the regularization technology; the real-time query solution includes the access point of the data stream, the key indicators for real-time monitoring and the format and frequency of data output.

3. The software design method for classifying defective silicon-based OLEDs based on the softmax algorithm according to claim 1, characterized in that: The steps of receiving the silicon-based OLED production data uploaded by the user, performing grouping and sorting of the data by date and process section information tag index, and generating the sorted data grouping are as follows: Based on the silicon-based OLED production data uploaded by users, the date and process section label of each data item is read, the data is classified according to the process section label, and each type of data is sorted into batches according to the date label to generate group records; According to the grouping records, a sorting operation is performed on each group of data using the date field, and each data item is arranged in sequence by comparing the timestamps of the data records, and each group of data is checked to be arranged in chronological order, thereby generating a sorted data group; The sorted data group is used to process the OLED screen parameter data provided by the sensor and the image acquisition device through a C# program, sort and store the data, and generate sorted data groups.

4. The software design method for classifying defective silicon-based OLEDs based on the softmax algorithm according to claim 1, characterized in that: The steps of grouping the sorted data, identifying and deleting noise and erroneous data, performing missing value filling and damaged sample removal, and performing normalization processing, and dividing the data into training set, validation set and test set in proportion, to obtain the processed data set are as follows: According to the sorted data grouping, filter and remove records with invalid values ​​and null values, check the target fields for each record, including dates and values, verify whether there are null values ​​and data that does not meet the input rules, and generate a clean data set; Based on the clean data set, outliers in the data are identified, a range check is performed on each numerical data item, data points that exceed the range of conventional production parameters are excluded, outliers are determined and deleted by setting thresholds, and a standard data set is generated; The standardized data set is adopted, the Z score method is used, the date format is uniformly converted, the numerical data is uniformly retained to two decimal places, the data is normalized to standardize the data scale, the data set is divided into a training set, a validation set and a test set according to the proportion, and a processed data set is generated.

5. According to the software design method for classifying defective silicon-based OLEDs based on the softmax algorithm according to claim 4, the formula of the Z score method is as follows: in, z ′ is the Z score of numerical data, x is the original value of each data point, μ is the mean of the corresponding data, σ is the standard deviation of the corresponding data, β is the data skewness adjustment coefficient, k is the scale adjustment coefficient, and α is the translation parameter.

6. The software design method for classifying defective silicon-based OLEDs based on the softmax algorithm according to claim 1, characterized in that: According to the processed data set, the loss function is set to cross entropy loss, the gradient descent method is selected as the parameter optimization method, the basic structure of the model is constructed, and the input and output of the model are defined. The specific steps of constructing the softmax linear regression model are: According to the processed data set, the number of nodes and activation function of the model input layer are set, the model weight and bias value are analyzed, the weight range and the initialization scale of the bias are adjusted, and an initialization parameter table is generated; Based on the initialization parameter table, read the CSV data, adopt the data parsing strategy, convert the data in the CSV into a numerical type, and standardize the numerical value into a feature vector, control the converted data dimension and format, and generate a feature input table; According to the feature input table, the model is trained, the number of iterations and the learning rate are set, the feature vectors are input batch by batch, the learning rate is adjusted to optimize the convergence speed, and a softmax linear regression model is constructed.

7. The software design method for classifying defective silicon-based OLEDs based on the softmax algorithm according to claim 1, characterized in that: The training set is used to perform parameter initialization in the softmax linear regression model, and the data is input batch by batch in a batch processing manner. The loss is calculated and the weight is updated. The process is repeated until convergence. The steps for generating the model training monitoring results are as follows: Perform parameter initialization in the softmax linear regression model using the training set, initialize the weights and biases of the model using normal distribution random values, and obtain a parameter initialization configuration; According to the parameter initialization configuration, import data from the CSV file, convert the data into a feature vector format accepted by the model, calculate the output probability distribution of each data sample, and obtain the probability distribution result; The probability distribution result is used to calculate the loss value and update the parameters through the back propagation method, and the cycle is iterated until the model performance is stable, and the model training monitoring result is output.

8. The software design method for classifying defective silicon-based OLEDs based on the softmax algorithm according to claim 1, characterized in that: Based on the model training monitoring results, the performance of the validation set is evaluated, the impact of parameter configuration on model performance is analyzed, and the parameter configuration with the best model performance is determined by iteratively testing differentiated hyperparameter settings. The specific steps for generating an optimized softmax linear regression model are as follows: According to the model training monitoring results, analyze the impact of differentiated parameter configurations on model performance, use validation set data to evaluate the model performance, and generate initialization evaluation data; Based on the initialization evaluation data, design multiple sets of hyperparameter configurations, including differentiated combinations of learning rates and batch sizes, and obtain hyperparameter test results by iteratively testing the hyperparameter configurations; Using the hyperparameter test results, the best performing hyperparameter configuration is selected, and the model is retrained on the validation set and the training set to verify that the model achieves the optimal performance, thereby obtaining an optimized softmax linear regression model.

9. The software design method for classifying defective silicon-based OLEDs based on the softmax algorithm according to claim 1, characterized in that: The optimized softmax linear regression model is deployed in the production environment to monitor the real-time production line data flow, classify the defects of silicon-based OLEDs, and import the data into the C# program. The specific steps of generating a real-time query solution are as follows: The optimized softmax linear regression model is configured and started on the C# program, a model interface is set to receive the real-time production line data stream, silicon-based OLED data monitoring is analyzed and bad defects are classified to obtain a defect classification data stream; Parse the defect classification data stream and write the parsing results into the database in real time through SQL statements. The process includes data insertion and update operations, checking the continuous update of wafer status data, and obtaining the updated wafer status set; The updated Wafer status set is used to query data and output bad classification information of real-time Wafer by using decision tree algorithm. The C# server code is used to query the situation and provide response to obtain a real-time query solution.

10. The software design method for classifying defective silicon-based OLEDs based on the softmax algorithm according to claim 9, characterized in that: The formula of the decision tree algorithm is as follows: Among them, C(i) is the bad classification index of the i-th wafer, x ij Represents the data of the i-th Wafer at the j-th detection index, μ j represents the average value of the jth detection index, w j is the weight coefficient of the jth detection index, and n is the total number of detection indicators.

Images