Method and device for dynamically updating data set and storage medium

Abstract

The invention discloses a method and a device for dynamically updating a data set and a storage medium, which are used for improving data quality and model performance of model training. Acquiring a training data set of a detection model, wherein the detection model is an artificial intelligence model to be added into the current display screen production line for display screen defect detection work; acquiring a display screen acquisition data set in real time, wherein the display screen acquisition data set is data which is acquired in the current display screen production line and is used for performing display screen defect detection; performing data freshness evaluation on the display screen collection data set; performing data importance evaluation on the display screen collection data set; according to the data freshness evaluation result and the data importance evaluation result, performing update grade analysis on the display screen collection data set; screening out a target data item set from the display screen collection data set according to the update grade analysis result; and updating the screened target data item set into the training data set.

Description

Technical Field

[0001] The present application relates to the field of display screen detection, and in particular to a method, device and storage medium for dynamically updating a data set. Background Art

[0002] In the field of display screen detection, it is usually necessary to detect different types of defects on the display screen, and defect detection and defect compensation are important production links. When different display screen structures generate the same defect, its manifestation is also different. Therefore, on the display screen production line, multiple artificial intelligence models are usually used to detect different display screen defects, such as detection of display screen body defects, detection of display screen display defects, detection of defects on the flat part of the display screen, detection of defects on the curved part of the display screen, etc. However, with the continuous updating and iteration of display screens, the functions on the display screens are gradually increasing and continuously refined, which has also brought more complex changes to the structure of the display screen. For example: in order to increase functionality, a new type of thin-film circuit structure is added under the pixel layer in the display screen. This type of circuit structure has a certain reflective ability. If there is a defect in the pixel layer area on the circuit area, it may cause false detection or missed detection. In order to cope with the defect detection of new display screens, it is often chosen to add a new artificial intelligence defect detection model to the display screen production line, while completing targeted defect detection, reducing the cost of production line changes.

[0003] After the design of the detection model is completed, the pre-deployment training is completed through the training data set (historical display screen collection data). At this time, it is usually not possible to deploy directly because the training data used by the detection model is usually relatively old data, and the new display screen (with a new structure) will have differences in the generation process of each batch. Therefore, before deploying the detection model, it is usually necessary to use data with higher freshness for training. The freshness of the data has an important impact on the performance of the detection model after deployment. However, the existing training data set update method is mostly static, requiring staff to manually select the latest batch of display screen collection data, and because each display screen ID will collect multiple images for defect detection in the same inspection project, the number of display screen collection data is often very large, and staff are required to additionally screen the display screen collection data with higher importance, which makes it time-consuming and inefficient, and the staff cannot accurately screen the display screen collection data with higher fit to the detection model in the current display screen production line, that is, timely reflect the latest changes in the display screen collection data, resulting in poor pre-deployment training effect of the training model, reducing the training data quality and detection model performance. Summary of the invention

[0004] The present application discloses a method, device and storage medium for dynamically updating a data set, which are used to improve data quality and model performance of model training.

[0005] In a first aspect, an embodiment of the present application provides a method for dynamically updating a data set, including: obtaining a training data set of a detection model, where the detection model is an artificial intelligence model that will be added to a current display screen production line to perform display screen defect detection, and the training data set is display screen data used by the detection model during pre-training; obtaining a display screen acquisition data set in real time, where the display screen acquisition data set is data collected in the current display screen production line for display screen defect detection; performing data freshness evaluation on the display screen acquisition data set; performing data importance evaluation on the display screen acquisition data set; performing update level analysis on the display screen acquisition data set based on the data freshness evaluation results and the data importance evaluation results; filtering out a target data item set from the display screen acquisition data set based on the update level analysis results; and updating the filtered target data item set into the training data set.

[0006] Optionally, after the step of updating the screened set of target data items into the training data set, the method further includes: using the updated training data set to pre-train the detection model before it is put into use; evaluating the model performance of the detection model using the updated training data set, and generating performance feedback information; when the performance feedback information shows that the detection model performance meets the standards, deploying the detection model to the display production line; and sending the performance feedback information of the detection model back to the production client for optimization in future model training.

[0007] Optionally, after the step of updating the screened target data item set into the training data set, and before the step of pre-training the detection model before using the updated training data set, the method also includes: pre-training sorting the display screen collection data in the target data item set according to the update level analysis results.

[0008] Optionally, the detection model is an artificial intelligence model for defect detection on the circuit area of ​​the display screen; after the step of acquiring the display screen acquisition data set in real time and before the step of evaluating the data freshness of the display screen acquisition data set, the method also includes: filtering out the display screen acquisition data containing the circuit area from the display screen acquisition data set, and obtaining the generation time data of the display screen acquisition data containing the circuit area; performing defect area and defect type analysis on the display screen acquisition data containing the circuit area, and generating a defect feature label for each display screen acquisition data having a defect in the circuit area; classifying the display screen acquisition data after the defect area and defect type analysis according to the display screen ID and the detection item; generating a circuit structure hierarchy based on the classified display screen acquisition data, the circuit structure hierarchy representing the structural level of the thin-film circuit in the circuit area; binding the acquired generation time data, defect feature label, and circuit structure hierarchy with the classified display screen acquisition data to generate a new display screen acquisition data set.

[0009] Optionally, the step of performing data freshness evaluation on the display screen collection data set includes: selecting a target display screen collection data from a new display screen collection data set; determining a time decay function based on defect detection items of the target display screen collection data, wherein the time decay function has an unknown variable, namely, generation time; generating a business weight based on a display screen model of the target display screen collection data, wherein the business weight refers to the importance of the display screen corresponding to the display screen model to the training of the detection model; determining a data change rate parameter corresponding to the display screen model based on the display screen model of the target display screen collection data and the new display screen collection data set; performing freshness evaluation based on the time decay function, the business weight, the data change rate parameter, and the generation time of the target display screen collection data; performing freshness evaluation on other display screen collection data in the new display screen collection data set, and integrating them to form a data freshness evaluation result.

[0010] Optionally, the step of performing data importance evaluation on the display screen acquisition data set includes: generating a structural weight according to a circuit structure hierarchy of the target display screen acquisition data; determining a target display screen ID of the target display screen acquisition data, and analyzing the defect feature label binding status of all display screen acquisition data of the same detection project under the target display screen ID; generating a detection weight for the target display screen acquisition data according to the binding analysis result; analyzing the defect feature labels of all display screen acquisition data of the same detection project under the target display screen ID, and generating a defect weight for the target display screen acquisition data according to the type analysis result of the defect feature label; performing importance evaluation on the target display screen acquisition data according to the structural weight, the detection weight and the defect weight; performing importance evaluation on other display screen acquisition data in the new display screen acquisition data set, and integrating them to form a data importance evaluation result.

[0011] Optionally, the step of performing update level analysis on the display screen acquisition data set according to the data freshness evaluation result and the data importance evaluation result includes: obtaining a first freshness threshold, a second freshness threshold, a first importance score threshold and a second importance score threshold, the first freshness threshold is less than the second freshness threshold, and the first importance score threshold is less than the second importance score threshold; generating a freshness level for each display screen acquisition data in the display screen acquisition data set according to the data freshness evaluation result, the first freshness threshold and the second freshness threshold; generating an importance level for each display screen acquisition data in the display screen acquisition data set according to the data importance evaluation result, the first importance score threshold and the second importance score threshold; performing update level analysis on each display screen acquisition data in the display screen acquisition data set according to the freshness level and the importance level, and generating an update level analysis result.

[0012] In a second aspect, an embodiment of the present application provides a device for dynamically updating a data set, including: a first acquisition unit, used to acquire a training data set of a detection model, where the detection model is an artificial intelligence model that will be added to the current display screen production line to perform display screen defect detection, and the training data set is the display screen data used by the detection model during pre-training; a second acquisition unit, used to acquire a display screen acquisition data set in real time, where the display screen acquisition data set is data collected in the current display screen production line for display screen defect detection; a first evaluation unit, used to perform data freshness evaluation on the display screen acquisition data set; a second evaluation unit, used to perform data importance evaluation on the display screen acquisition data set; a first analysis unit, used to perform an update level analysis on the display screen acquisition data set based on the data freshness evaluation results and the data importance evaluation results; a screening unit, used to screen out a target data item set from the display screen acquisition data set based on the update level analysis results; and an update unit, used to update the screened target data item set into the training data set.

[0013] Optionally, after the updating unit, the device further includes: a pre-training unit, used to use the updated training data set to pre-train the detection model before it is put into use; a first generation unit, used to evaluate the model performance of the detection model using the updated training data set and generate performance feedback information; a deployment unit, used to deploy the detection model to the display production line when the performance feedback information shows that the detection model performance meets the standards; and a feedback unit, used to send the performance feedback information of the detection model back to the production client for optimization in future model training.

[0014] Optionally, after the updating unit and before the pre-training unit, the device further includes: a sorting unit, configured to perform pre-training sorting for the display screen collected data in the target data item set according to the update level analysis result.

[0015] Optionally, the detection model is an artificial intelligence model for defect detection of the circuit area of ​​the display screen; after the second acquisition unit and before the first evaluation unit, the device also includes: a third acquisition unit, which is used to filter out the display screen acquisition data containing the circuit area from the display screen acquisition data set, and obtain the generation time data of the display screen acquisition data containing the circuit area; a second analysis unit, which is used to perform defect area and defect type analysis on the display screen acquisition data containing the circuit area, and generate a defect feature label for each display screen acquisition data with defects in the circuit area; a classification unit, which is used to classify the display screen acquisition data after the defect area and defect type analysis according to the display screen ID and the detection item; a second generation unit, which is used to generate a circuit structure hierarchy based on the classified display screen acquisition data, and the circuit structure hierarchy represents the structural level of the thin-film circuit in the circuit area; a third generation unit, which is used to bind the acquired generation time data, defect feature label, and circuit structure hierarchy with the classified display screen acquisition data to generate a new display screen acquisition data set.

[0016] Optionally, the steps of the first evaluation unit include: selecting a target display screen collection data from a new display screen collection data set; determining a time decay function based on defect detection items of the target display screen collection data, wherein the time decay function has an unknown variable, namely, generation time; generating a business weight based on a display screen model of the target display screen collection data, wherein the business weight refers to the importance of the display screen corresponding to the display screen model to the training of the detection model; determining a data change rate parameter corresponding to the display screen model based on the display screen model of the target display screen collection data and the new display screen collection data set; performing a freshness evaluation based on the time decay function, the business weight, the data change rate parameter, and the generation time of the target display screen collection data; performing a freshness evaluation on other display screen collection data in the new display screen collection data set, and integrating them to form a data freshness evaluation result.

[0017] Optionally, the steps of the second evaluation unit include: generating a structural weight according to a circuit structure level of the target display screen acquisition data; determining a target display screen ID of the target display screen acquisition data, and analyzing the defect feature label binding status of all display screen acquisition data of the same detection project under the target display screen ID; generating a detection weight for the target display screen acquisition data according to the binding analysis result; analyzing the defect feature labels of all display screen acquisition data of the same detection project under the target display screen ID, and generating a defect weight for the target display screen acquisition data according to the type analysis result of the defect feature label; performing importance evaluation for the target display screen acquisition data according to the structural weight, the detection weight and the defect weight; performing importance evaluation on other display screen acquisition data in the new display screen acquisition data set, and integrating them to form a data importance evaluation result.

[0018] Optionally, the first analysis unit includes: obtaining a first freshness threshold, a second freshness threshold, a first importance score threshold and a second importance score threshold, the first freshness threshold is less than the second freshness threshold, and the first importance score threshold is less than the second importance score threshold; generating a freshness level for each display screen collected data in the display screen collected data set according to the data freshness evaluation result, the first freshness threshold and the second freshness threshold; generating an importance level for each display screen collected data in the display screen collected data set according to the data importance evaluation result, the first importance score threshold and the second importance score threshold; performing an update level analysis on each display screen collected data in the display screen collected data set according to the freshness level and the importance level, and generating an update level analysis result.

[0019] In a third aspect, an embodiment of the present application provides a device for dynamically updating a data set, including: Processor, memory, input-output unit, and bus; The processor is connected to the memory, the input and output unit, and the bus; The memory stores a program, and the processor calls the program to execute the first aspect and any optional method of the first aspect.

[0020] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a program is stored. When the program is executed on a computer, the program executes the first aspect and any optional method of the first aspect.

[0021] It can be seen from the above technical solutions that this application has the following advantages: The present application first obtains a training data set for a detection model, where the detection model is an artificial intelligence model that will be added to the current display screen production line to perform display screen defect detection, and the training data set is the display screen data used by the detection model during pre-training; a display screen acquisition data set is acquired in real time, where the display screen acquisition data set is data collected in the current display screen production line for display screen defect detection; a data freshness assessment is performed on the display screen acquisition data set; a data importance assessment is performed on the display screen acquisition data set; an update level analysis is performed on the display screen acquisition data set based on the data freshness assessment results and the data importance assessment results; a target data item set is screened out from the display screen acquisition data set based on the update level analysis results; and the screened target data item set is updated into the training data set.

[0022] By collecting data from display screens on the production line in real time, we obtain the latest data, conduct freshness and importance assessments on a large amount of display screen data, and use the evaluation results to filter and update the training data set, ensuring that the display screen defect detection model can continuously learn and adapt to new situations on the production line, thereby improving the data quality and model performance of model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solution in the present application, the drawings required for use in the embodiments or the description of the prior art 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 paying any creative labor.

[0024] Figure 1 A schematic diagram of an embodiment of a method for dynamically updating a data set of the present application; Figure 2 A schematic diagram of an embodiment of a method for pre-training a detection model for the present application; Figure 3 A schematic diagram of an embodiment of a method for sorting display screen data collected by the present application; Figure 4 A schematic diagram of an embodiment of a method for adjusting a display screen acquisition data set for the present application; Figure 5 A schematic diagram of an embodiment of a method for evaluating data freshness of the present application; Figure 6 A schematic diagram of an embodiment of a method for evaluating the importance of data in the present application; Figure 7 A schematic diagram of an embodiment of a method for updating grade analysis for this application; Figure 8 A schematic diagram of an embodiment of the screen color correction device of the present application; Fig. 9 This is a schematic diagram of another embodiment of the screen color correction device of the present application. DETAILED DESCRIPTION

[0025] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0026] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0027] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0028] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0029] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0030] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0031] In the prior art, in order to cope with the defect detection of new display screens, a new artificial intelligence defect detection model is often chosen to be added to the production line of the display screen, so as to complete the targeted defect detection and reduce the change cost of the production line. After the design of the detection model is completed and the pre-deployment training is completed through the training data set, it is usually not directly deployed at this time, because the training data used by the detection model is usually relatively old data, and there will be differences in the generation process of each batch of new display screens. Therefore, before deploying the detection model, it is usually necessary to use data with higher freshness for training. The freshness of the data has an important impact on the performance of the detection model after deployment. However, the existing training data set update method is mostly static, requiring staff to manually select the latest batch of display screen collection data, and because each display screen ID will collect multiple images for defect detection in the same detection project, the number of display screen collection data is often very large, and the staff is required to additionally screen the display screen collection data with higher importance, which makes it time-consuming and inefficient, and the staff cannot accurately screen the display screen collection data with higher fit to the detection model in the current display screen production line, that is, timely reflect the latest changes in the display screen collection data, resulting in poor pre-deployment training effect of the training model, reducing the training data quality and detection model performance.

[0032] Based on this, the present application discloses a method, device and storage medium for dynamically updating a data set, which are used to improve the data quality and model performance of model training.

[0033] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0034] The method of the present application can be applied to a server, a device, a terminal or other devices with logic processing capabilities, and the present application does not limit this. For the convenience of description, the following description is made by taking the execution subject as an example of a terminal.

[0035] See also Figure 1 The present application provides an embodiment of a method for dynamically updating a data set, including: 101. Obtain a training data set for a detection model, where the detection model is an artificial intelligence model that will be added to the current display production line to perform display defect detection, and the training data set is display data used by the detection model during pre-training.

[0036] First, obtain the training data set of the detection model. The training data set should come from a variety of display production lines to ensure data diversity, covering display data of different models, different production batches, and different defect types, to ensure that the detection model has certain display detection capabilities in the initial training.

[0037] 102. Acquire a display screen acquisition data set in real time, where the display screen acquisition data set is data collected in a current display screen production line and used for display screen defect detection.

[0038] Next, we acquire the display screen data set in real time, using the high-precision and high-stability data acquisition equipment that has been put into production lines to collect data, which can ensure the quality of the collected data to the greatest extent. We also realize real-time data collection and transmission to ensure that the training model can learn the latest production line conditions in a timely manner.

[0039] Specifically, this step includes producing a client and a data source.

[0040] Production client: the production line where the model is deployed.

[0041] Data source: Equipment or sensors on the production line continuously collect new defect data, which directly reflects the latest status of the production line, including new defect types.

[0042] 103. Evaluate the data freshness of the display screen collected data set.

[0043] The core of data freshness evaluation on the display screen data set is to determine the data collection time through the timestamp, and to evaluate the data freshness through data such as collection time, defect type, and frequency.

[0044] Specifically, this step includes a data freshness evaluation unit.

[0045] Data freshness assessment unit: After new defect data is generated from the data source, it will be sent to the data freshness assessment unit with a timestamp. The responsibility of this unit is to assess the newness of the data, that is, the freshness of the data. The freshness can be measured by the difference between the timestamp of the data generation and the current time (i.e. Age of Information, AoI) to obtain the freshness level.

[0046] 104. Evaluate the data importance of the display screen collected data set.

[0047] Next, it is necessary to evaluate the data importance of the display screen data set and conduct multi-angle analysis on the data collected from each production batch.

[0048] Specifically, this step includes a data importance evaluation unit.

[0049] Data importance assessment unit: Evaluate the importance of each data item in the data source based on the data's contribution to the model performance, and use the model's inference score on the data to measure and obtain the importance level.

[0050] 105. Perform update level analysis on the display screen collected data set based on the data freshness evaluation results and the data importance evaluation results.

[0051] Then, the display screen data set is updated according to the data freshness evaluation results and data importance evaluation results. A comprehensive score needs to be calculated for each display screen data, which is based on data freshness and importance. Then, the data is divided into different levels according to the comprehensive score, such as high priority, medium priority, low priority, etc.

[0052] 106. Filter out a target data item set from the display screen collected data set according to the update level analysis result.

[0053] According to the update level analysis results, the target data item set is filtered out from the display screen collection data set. And high-priority data is filtered out first. When high-priority data is insufficient, medium-priority data is considered. In order to ensure that the amount of filtered data is moderate, it is neither too much to cause low processing efficiency nor too little to cause insufficient model update.

[0054] Specifically, this step includes a data screening unit.

[0055] Data screening unit: Filter out data items that need to be updated based on freshness and importance evaluation results.

[0056] 107. Update the selected target data item set into the training data set.

[0057] Finally, the selected target data item set is updated to the training data set, and the new data is integrated into the existing training data set to ensure the coherence and consistency of the data.

[0058] Specifically, this step includes a data updating unit.

[0059] Data update unit: adds the filtered data items to the data set to update the data set.

[0060] In an embodiment of the present application, a training data set of a detection model is obtained. The detection model is an artificial intelligence model that will be added to the current display screen production line to perform display screen defect detection. The training data set is the display screen data used by the detection model during pre-training. A display screen acquisition data set is acquired in real time. The display screen acquisition data set is data collected in the current display screen production line for display screen defect detection. Data freshness is evaluated on the display screen acquisition data set. Data importance is evaluated on the display screen acquisition data set. An update level analysis is performed on the display screen acquisition data set based on the data freshness evaluation results and the data importance evaluation results. A target data item set is filtered out from the display screen acquisition data set based on the update level analysis results. The filtered target data item set is updated into the training data set.

[0061] By automatically monitoring data changes, evaluating data freshness and importance, and dynamically updating data sets, the data quality and model performance of model training can be improved. Specifically, the latest data is obtained by real-time collection of display screens on the production line, and freshness and importance assessments are performed on a large number of display screen collection data. The training data set is screened and updated based on the evaluation results, ensuring that the display screen defect detection model can continue to learn and adapt to new situations in the production line, improving the data quality and model performance of model training.

[0062] Secondly, the embodiments of the present application also have the following beneficial effects: 1. Automated operation: reduces manual intervention and improves the efficiency and accuracy of data updates.

[0063] 2. Improve data freshness: By dynamically updating the data set, the freshness of the data is improved, so that model training can reflect the latest changes in the data in a timely manner.

[0064] 3. Improve model performance: By training the model with the latest and most relevant data, the accuracy and generalization ability of the model are improved.

[0065] 4. Save resources: Optimize data screening strategies and reduce unnecessary data processing and storage requirements.

[0066] 5. Strong adaptability: This method can adapt to rapid changes in data and is suitable for various scenarios that require real-time updates of data sets.

[0067] See also Figure 2 The present application provides an embodiment of a method for pre-training a detection model, comprising: 201. Use the updated training dataset to pre-train the detection model before putting it into use.

[0068] 202. Evaluate the model performance of the detection model using the updated training data set and generate performance feedback information.

[0069] 203. When the performance feedback information shows that the detection model performance meets the standard, the detection model is deployed to the display production line.

[0070] 204. Send performance feedback information of the detection model back to the production client for optimization in future model training.

[0071] In the embodiment of the present application, the terminal pre-trains the detection model with the updated training data set before putting it into use. The purpose of pre-training is to enable the detection model to adapt to the updated data set, learn new defect characteristics, and improve the generalization ability of the model. According to the complexity of the model and the scale of data, select an appropriate training strategy, such as gradually increasing the learning rate, using transfer learning, etc.

[0072] Next, the terminal evaluates the model performance of the detection model using the updated training data set. Specifically, in terms of evaluation indicators, it is necessary to select appropriate evaluation indicators, such as accuracy, recall, F1 score, AUC, etc., to comprehensively measure the performance of the model. In addition, cross-validation is performed. The terminal uses a cross-validation method to evaluate the stability and generalization ability of the model to avoid performance deviations caused by improper data division. Performance benchmarks are compared with previous model versions to evaluate the extent of performance improvement or decline, as well as the performance of the new model on specific defect types.

[0073] When the performance feedback information shows that the detection model performance meets the standard, the detection model is deployed to the display production line. First, ensure that the model files, dependent libraries, configuration files, etc. have been correctly packaged and deployed to the production environment. Then perform compatibility testing in the production environment to ensure that the model can run normally and integrate seamlessly with other system components. In addition, a monitoring and alarm system is deployed on the production line to track the performance changes of the model in real time, and to promptly discover and deal with potential problems.

[0074] Finally, the terminal sends the performance feedback information of the detection model back to the production client for optimization in future model training. Specifically, the terminal collects performance feedback information of the model from the production client, including accuracy, false alarm rate, missed alarm rate, etc. New defects on the production line are marked and added to the training data set to enrich the training samples of the model. Based on the performance feedback information and the newly added training data, the model is iteratively trained and optimized to continuously improve the performance and accuracy of the model. A closed-loop optimization mechanism is established to ensure that the model can continuously learn and adapt to new situations on the production line, forming a virtuous cycle of continuous improvement.

[0075] Through pre-training, performance evaluation, deployment, and feedback collection, a complete model update and optimization cycle is formed. During the implementation process, the performance changes of the model are analyzed, and the training strategy and parameters are adjusted in time to ensure the stability and accuracy of the model on the production line. At the same time, the data annotation and collection process is continuously optimized to enrich the training samples and improve the generalization ability of the model. In addition, establishing an effective monitoring and alarm system is also the key to ensuring the stable operation of the model. In future model training, new algorithms and technologies can be further explored to improve the performance and efficiency of the model.

[0076] In the embodiment of the present application, the above steps are composed of a model training unit, a model evaluation unit, a model deployment unit and a performance feedback unit.

[0077] Model training unit: Use the updated data set to perform model training; Model evaluation unit: evaluates the performance of the model trained using the updated dataset; Model deployment unit: If the model performance is good, the model is deployed to the production line; Performance Feedback Unit: Sends performance feedback information of the model back to the production client for optimization in future model training. This feedback loop ensures that the system can continuously adapt to new defect types and changes.

[0078] See also Figure 3 The present application provides an embodiment of a method for sorting display screen collected data, comprising: 301. Pre-train and sort the display screen collection data in the target data item set according to the update level analysis result.

[0079] In the embodiment of the present application, a training plan is provided for the detection model by updating the level analysis results, the display screen collection data that meets the conditions is sorted by freshness and importance, and the data with higher rankings are trained first, so as to ensure that the detection model meets the expectations before deployment more quickly.

[0080] See also Figure 4 The present application provides an embodiment of a method for adjusting a display screen acquisition data set, comprising: 401. Screening display screen collected data containing a circuit area is filtered out from a display screen collected data set, and generation time data of the display screen collected data containing the circuit area is obtained.

[0081] In an embodiment of the present application, a plurality of display screen defect detection models are arranged on a traditional display screen production line, mainly a display screen flat area detection model, a display screen curved area detection model, etc., as well as an artificial intelligence model for detecting specific defects, a model for detecting screen body defects, and a model for detecting real defects.

[0082] The display screen acquisition data set is the image data collected on the above-mentioned production line. These image data are input into the display screen defect detection models for detection, and the corresponding defect detection data are obtained as sample annotations.

[0083] With the development of display screens, the functionality of display screens has gradually increased, and its structure has gradually become more complex, which is mainly reflected in the increase in the number of layers of the functional layer of the display screen, the curvature of the screen, the splicing between the screens, and the addition of internal circuits between the functional layers. In the embodiment of the present application, the display screen with an internal circuit is mainly discussed, and it is a thin circuit below the pixel layer (the thin circuit is located inside the display screen). The thin circuit belongs to a circuit structure with a small thickness, which is much thinner than the PCB board below the display screen. It is specifically located below the pixel layer inside the display screen. While providing the corresponding pixel point display function for the display screen, the thinner thin circuit will also minimize the light reflection effect caused when the pixel layer is lit. However, with the complexity of the function, the complexity of some thin circuits becomes higher, and the thickness is inevitably close to the critical value. For some display screens that are very thin in themselves, it may reach the degree of affecting the display of the pixel layer, and the thin circuit When realizing the pixel point function, some metal materials or alloys used, its own reflection ability is strong, and a certain thickness can be reached to reflect a specific light intensity. Although the human eye may not be able to easily perceive the light reflection of the thin circuit inside the display screen, for detection instruments and defect detection models, this type of light reflection can affect the defect detection results.

[0084] Due to the improvement of the display screen, different thin-film circuits are added under the pixel layer of the display screen, forming different circuit areas on the display screen. It is necessary to generate a defect detection model dedicated to the defects in the circuit area on the original production line. At this time, a certain number of inspections have been carried out on the display screen with circuit areas in the production line, that is, there are certain samples that can be used for pre-training. At this time, it is necessary to analyze which of these display screen acquisition data is more suitable for pre-training of the detection model to accelerate the detection accuracy of the detection model and accelerate the deployment of the detection model. Using more important data can complete the training faster, and using more fresh data can make the detection model better adapt to the new display screens in the future (display screen data with more circuit areas in the future or display screen data with more complex circuit areas). In the traditional solution, the display screen containing the circuit area is usually first screened out, and the images with defects in the circuit area are screened out. The conclusion of the existence of defects is generated by the original defect detection model on the production line. The detection accuracy of the original detection model for new defects is lower than that of the new defect detection model that will be added to the production line (the artificial intelligence model for defect detection in the circuit area of ​​the display screen). However, such a simple screening can only screen out the display screen acquisition data with very obvious defects in the circuit area. We believe that this type of data is of low importance as training data for new defect detection models (artificial intelligence models that detect defects in the circuit area of ​​display screens) before they are put into production lines, which will make the pre-training effect of new defect detection models worse, and the training time will be longer and the efficiency will be lower. In response to this, we have designed a new method for screening display screen data collected in display screen production lines to increase the pre-training quality of new defect detection models before they are put into production lines and improve pre-training efficiency.

[0085] In the embodiment of the present application, the terminal selects the display screen collection data containing the circuit area from the display screen collection data set, because the display screen samples without the circuit area exist in the original training data set, have been trained a lot, and such samples are more and relatively fresh. Next, the terminal obtains the generation time data of such display screen collection data containing the circuit area, and the generation time data is used for the subsequent freshness evaluation.

[0086] 402. Perform defect area and defect type analysis on the display screen data collected from the circuit area, and generate a defect feature label for each display screen data collected from the circuit area.

[0087] In the embodiment of the present application, the terminal performs defect area and defect type analysis on the display screen data containing the circuit area, and generates a defect feature label for each display screen data with defects in the circuit area. The purpose is to focus on the display screen data with defects in the circuit area through detection by the model in the existing production line. Because the existing defect detection model currently working on the production line may misdetect or accurately detect defects in the circuit area, this type of data has a higher importance. The terminal will first mark this type of display screen data and add a defect feature label.

[0088] 403. Classify the display screen data collected after defect area and defect type analysis according to the display screen ID and the inspection item.

[0089] The terminal will classify the display screen data collected after defect area and defect type analysis according to the display screen ID and test item. The purpose is to integrate multiple display screen data collected under the same test item for the same display screen ID to form a display screen data collection group. For the same display screen ID in one test item (one defect detection model), several images are usually collected. Even for the same grayscale image, at least two images will be collected for repeated testing.

[0090] 404. Generate a circuit structure hierarchy according to the classified display screen collected data, where the circuit structure hierarchy represents a structural level of the thin-sheet circuit in the circuit area.

[0091] Since the display screen models of each batch are not necessarily the same, and the display screen functions required by each manufacturer are not the same, resulting in different structures in the circuit area. The thickness and reflectivity of the circuit area are the main factors affecting defect detection. The circuit structure level consists of three items: the area of ​​the thin circuit, the thickness data (gradient thickness, uniform thickness), and the metal material. The larger the area, the higher the structure level, the more complex the thickness type, the higher the structure level (gradient type is more complex than uniform type), the larger the thickness, the higher the structure level, and the greater the reflectivity of the metal, the higher the structure level. At present, the structural level is mainly divided artificially, and the circuit structure level is divided accordingly. The circuit structure level of the same model of display screen is the same.

[0092] 405. Bind the acquired generation time data, defect feature labels, circuit structure levels, and the classified display screen acquisition data to generate a new display screen acquisition data set.

[0093] The terminal binds the acquired generation time data, defect feature labels, and circuit structure levels with the classified display screen acquisition data to generate a new display screen acquisition data set, that is, to integrate the data of the same display screen ID, and further organize the data of the same inspection item, and finally bind the generation time data, defect feature labels, and circuit structure levels with the acquisition data of each display screen.

[0094] The beneficial effect of the embodiment of the present application is to organize data for subsequent freshness evaluation and importance evaluation, increase the accuracy of the evaluation, and screen out more important display screen acquisition data.

[0095] See also Figure 5 The present application provides an embodiment of a method for evaluating data freshness, including: 501. Select a target display screen acquisition data from a new display screen acquisition data set.

[0096] 502. Determine a time decay function based on defect detection items of data collected from a target display screen. The time decay function has an unknown variable, namely, generation time.

[0097] In an embodiment of the present application, the terminal selects a target display screen collection data from a new display screen collection data set. Next, the terminal determines the time decay function according to the defect detection items of the target display screen collection data, and the time decay function has an unknown variable, the generation time. The defect detection items are mainly screen defect detection and display defect detection. Screen defect detection refers to the structural defects of the screen, such as scratches, broken edges, etc. These defects are usually illuminated by a light source in an unlit state, and then the display screen is captured to capture the image. Display defect detection refers to defects that appear on pixels, which need to be illuminated by inputting a calibration image, and then the image is captured and then detected. There are slight differences in the time decay functions of screen defect detection and display defect detection. The time decay function needs to be substituted into the generation time of the display screen collection data for calculation.

[0098] In the embodiment of the present application, the time decay function is first based on the characteristics of display screen production, because display screens are produced in batches, and batch defects will occur during the production process of display screens. Batch defects refer to a parameter or device (such as a robotic arm) on the production line. Due to improper settings or accidental improper operation, the entire batch of display screens uniformly have specific defects. The appearance of such defects is usually similar but there are differences in the positions. For example, the robotic arm has a deviation in the grasping position due to strain, so that the entire batch of display screens have small screen scratch defects, but the screen scratch defects on each display screen may be different, in terms of position or degree of scratches. According to the characteristics of batch defects, a piecewise time decay function is usually selected, and the decay of the piecewise function is controlled to be faster, because such batch defects are usually small and not easy to observe, and the causes are usually difficult to be quickly checked and eliminated, but once discovered, they are usually quickly eliminated. Therefore, the embodiment of the present application selects a variety of means to evaluate the freshness of data, in which the time decay function is set.

[0099]

[0100] is the time decay function, is the attenuation coefficient, and The time parameter is set according to the average troubleshooting time of batch defects in the current production line. is the data age, is the current time, The time when the data was generated.

[0101] 503. Generate a business weight according to the display screen model of the target display screen collected data, where the business weight refers to the importance of the display screen corresponding to the display screen model to the training of the detection model.

[0102] Next, the terminal generates a business weight based on the display model of the target display screen data collection. The business weight refers to the importance of the display screen corresponding to the display screen model to the training of the detection model. Specifically, the display screen collection data of some display screen models already exist in the training data set. The business weight of the trained display screen model is lower, and the business weight of the untrained display screen model is higher.

[0103] In the embodiment of the present application, the service weight is set to , only distinguishing the business weights of the trained display screens from the business weights of the untrained ones.

[0104] 504. Determine a data change rate parameter corresponding to the display screen model according to the display screen model of the target display screen data collection and the new display screen collection data set.

[0105] The terminal determines the data change rate parameter corresponding to the display model according to the display model of the target display data collection and the new display data collection set. Specifically, the terminal determines the number of data in the display data collection set that is the same as the "display model" of the target display data collection, and generates a data change rate parameter based on this number and the total amount of the display data collection set. The data change rate refers to the size of the data flow of this type. If the flow is large, more training is required, and if the flow is small, less training is required.

[0106] 505. Perform freshness evaluation based on the time decay function, business weight, data change rate parameters, and generation time of the target display screen collected data.

[0107] In this embodiment, the freshness evaluation formula is as follows:

[0108] in, is the freshness evaluation formula, is the data change rate, , , is the weighted value, the sum of the three is 1, The business weight of the current type of display screen.

[0109] 506. Perform freshness evaluation on other display screen collected data in the new display screen collected data set, and integrate them to form a data freshness evaluation result.

[0110] Finally, the terminal performs a freshness evaluation based on the time decay function, business weight, data change rate parameters, and the generation time of the target display screen data. Finally, the terminal performs a freshness evaluation on the other display screen data in the new display screen data set and integrates them to form the data freshness evaluation result.

[0111] This method utilizes a multi-dimensional comprehensive calculation of the freshness of display screen data. It uses basic freshness calculation functions (time decay functions), business weights, and data traffic to comprehensively calculate the freshness of data, thereby providing greater freshness for required data and improving the quality of display screen data collection and detection model performance.

[0112] See also Figure 6 The present application provides an embodiment of a method for evaluating data importance, comprising: 601. Generate a structural weight according to the circuit structure level of the target display screen acquisition data.

[0113] In an embodiment of the present application, the terminal generates a structural weight based on the circuit structure hierarchy of the target display screen data collected. The purpose is to incorporate complex circuit structures into importance judgments. Because the more complex the circuit area structure, the higher its training value, it is necessary to generate a structural weight based on the reflectivity, thickness and other influencing factors of the circuit structure. The structural weight serves as an important item in importance assessment.

[0114] 602. Determine the target display screen ID for collecting data of the target display screen, and analyze the defect feature label binding status of all display screen collected data of the same inspection item under the target display screen ID.

[0115] 603. Generate detection weights for the target display screen data collection according to the binding analysis results.

[0116] In the embodiment of the present application, the terminal determines the target display screen ID of the target display screen collection data, analyzes the defect feature label binding of all display screen collection data of the same detection project under the target display screen ID, and generates a detection weight for the target display screen collection data according to the binding analysis result. When part of all display screen collection data of the same detection project is bound to defect labels, but part is not bound to defect labels, the importance of the part not bound to defect labels is adjusted at this time.

[0117] For example: Display screen No. 001 (with circuit area) collected 10 images in the screen defect detection project, and defect analysis was performed on these 10 images. The detection results showed that there were defects in 3 images (No. 01, No. 02 and No. 05). After the subsequent defect area and defect type analysis, it was found that among the 3 images, No. 01 detected screen defects in the circuit area, and No. 02 and No. 05 detected screen defects in the non-circuit area. A defect feature label was generated for No. 01, and these 10 images were classified into the same batch. At this time, No. 01 belongs to the regular pre-training sample, and No. 02 to No. 10 are determined as advanced training samples of the "new defect detection model" because these samples are samples that are difficult for models of other production lines to detect defects, but they are actually defective. It should be noted that in the same inspection project of the same display ID, when the above situation also exists between the data corresponding to different display screens, the importance adjustment rules are also applicable.

[0118] Specifically, this part is to analyze all the display screen collection data in the same display screen ID and the same inspection project, and analyze whether there is data bound to defect feature labels in this group of data. If so, detection weights are generated for this group of data or only for the target display screen collection data.

[0119] 604. Analyze defect feature labels of all display screen collected data of the same inspection item under the target display screen ID, and generate defect weights for the target display screen collected data according to the type analysis results of the defect feature labels.

[0120] In an embodiment of the present application, the terminal analyzes the defect feature labels of all display screen collection data of the same detection project under the target display screen ID, and generates defect weights for the target display screen collection data according to the type analysis results of the defect feature labels. Specifically, for example: Display screen No. 001 (there is a circuit area) collected 10 images in the screen defect detection project, and defect analysis was performed on these 10 images. The detection results showed that there were defects in 3 images (No. 02 and No. 05). Both images have defects in the circuit area, and the defect types are different. No. 02 is defect A, and No. 05 is an unknown defect. The terminal generates a defect weight for defect A for image No. 02. The defect weight is generated based on the detection accuracy of the defect in the current production line. If the defect is more difficult to detect, the defect weight is larger. Unknown defects indicate that the detection data of the current production line cannot be identified, and the defect weight is generated manually.

[0121] 605. The importance of collecting data for the target display screen is evaluated based on the structure weight, detection weight and defect weight.

[0122] 606. Perform importance assessment on other display screen collection data in the new display screen collection data set, and integrate them to form a data importance assessment result.

[0123] Finally, the terminal performs an importance assessment for the target display screen data according to the structure weight, detection weight and defect weight, and finally performs an importance assessment for the other display screen data in the new display screen data set, and integrates them to form a data importance assessment result. Specifically, the present application adds the structure weight, detection weight and defect weight to obtain the final importance assessment data.

[0124] The above method can more accurately evaluate the importance of each display screen captured image by comprehensively considering the circuit area structure of the display screen data collected and the production line defect detection results. It can better provide greater importance to the required data and improve the quality of the display screen data collected and the performance of the detection model.

[0125] See also Figure 7 The present application provides an embodiment of a method for updating grade analysis, comprising: 701. Obtain a first freshness threshold, a second freshness threshold, a first importance score threshold, and a second importance score threshold, wherein the first freshness threshold is smaller than the second freshness threshold, and the first importance score threshold is smaller than the second importance score threshold.

[0126] 702. Generate a freshness level for each piece of display acquisition data in the display acquisition data set according to the data freshness evaluation result, the first freshness threshold, and the second freshness threshold.

[0127] 703. Generate an importance level for each piece of display acquisition data in the display acquisition data set according to the data importance evaluation result, the first importance score threshold, and the second importance score threshold.

[0128] 704. Perform an update level analysis on each piece of display acquisition data in the display acquisition data set according to the freshness level and the importance level, and generate an update level analysis result.

[0129] In the embodiments of the present application, the freshness evaluation unit uses the age of information (AoI) as the data freshness. In the present invention, the rules for dividing the data freshness level are given: ① Preset a first freshness threshold (AoI_1) and a second freshness threshold (AoI_2) according to the actual generation situation, where AoI_1 < AoI_2; ② Compare the size relationship between the AoI of the current data item and AoI_1 and AoI_2 respectively; ③ If AoI < AoI_1, the data freshness level is high; if AoI_1 < AoI < AoI_2, the data freshness level is medium; if AoI > AoI_2, the data freshness level is low; The evaluation process of data freshness: ① Data monitoring: The freshness evaluation unit first monitors the changes in the database, records the generation time, reception time, and other relevant parameters of the data; ② Freshness calculation: Calculate the AoI freshness of each data item based on the timestamp of the data.

[0130] ③ Freshness level division: The evaluation unit compares the calculated AoI with the preset threshold to determine the data freshness level. A shorter AoI indicates fresher data, while a longer AoI indicates that the data is not fresh or has expired; ④ Decision support: Use the evaluation result for data update decision-making, and preferentially select data items with a high freshness level for update.

[0131] The data importance evaluation unit is a key component in the data set dynamic update method. Its purpose is to determine which data items in the data set have a significant impact on the performance of the model, so as to preferentially retain or update these data. The contribution degree of a data item to the model performance can be measured by its impact on the model evaluation metrics (such as accuracy, recall rate, F1 score, etc.).

[0132] In the embodiments of the present application, the importance of data can be evaluated through a neural network model based on deep learning (a defect detection model that has been put into production on the production line). The specific level classification rules are as follows: 1. Preset a first importance score threshold Importance_Th1 and a second importance score threshold Importance_Th2 according to the actual generation situation, where 0 < Importance_Th1 < Importance_Th2 < 1; 2. Through the detection results of the neural network model of deep learning on the display screen acquisition data, perform importance evaluation to obtain the corresponding classification score score. For example: if the detection model has learned this type of data, the classification score will be relatively high and the data importance will be relatively small. Otherwise, the classification score will be relatively low and the data importance will be relatively large. Therefore, the model inference score of the data is inversely proportional to the data importance score. It should be noted that the classification score score here is only one of the evaluation schemes and not the final importance evaluation result; 3. Calculate the data importance score Importance_score, which can be calculated according to the model inference score score. For example, according to the case in the previous step, the data importance score can be calculated by the following formula: Importance_score = 1 - score. The final data importance score is directly proportional to the data importance; 4. Compare the size relationship between the importance score Importance_score of the data and the first importance score threshold Importance_Th1 and the second importance score threshold Importance_Th2 respectively; 5. If Importance_score < Importance_Th1, the data importance level is low; if Importance_Th1 < Importance_score < Importance_Th2, the data importance level is medium; if Importance_score > Importance_Th2, the data importance level is high; The process of data importance evaluation: ① Data preprocessing: Before evaluation, preprocess the data to ensure the accuracy of the evaluation; ② Importance evaluation: Based on the deep learning model, evaluate the importance score of each data item; ③ Importance level classification: Analyze the evaluation results and determine the importance level of the data item according to the preset importance threshold; ④ Decision support: Use the evaluation results for data update decisions, and preferentially select data items with a high importance level for update.

[0133] Finally, the terminal performs update level analysis on each display screen collection data in the display screen collection data set according to the freshness level and importance level, and generates an update level analysis result. Specifically: The data update conditions are determined by the data freshness level and the data importance level. The specific joint strategy is as follows: High importance and high freshness: Data is both important and fresh and should be processed first; High importance and medium freshness: The data is important but old, and the processing priority needs to be weighed; High importance and low freshness: The data is important but too fresh, so it has a low processing priority; Medium importance and high freshness: The data is generally important but fresh, and the processing priority needs to be weighed; Medium importance and medium freshness: The data is generally important but relatively old, and the processing priority needs to be weighed; Medium importance and low freshness: The data is generally important but too fresh, so the processing priority is low; Low importance and high freshness: data is not important but fresh, so the processing priority is low; Low importance and medium freshness: data is not important but old, so the processing priority is low; Low importance and low freshness: Data is not important and outdated, so it is not processed; This division method helps to achieve the best balance between value and freshness in data transmission and processing. According to this strategy, the data set can be updated more efficiently, and it also ensures that the data set can timely feedback the latest generation status of the client, ensuring that the deployed model can cope with new defect types that appear in the production line and improve the detection performance of the production line. It should be noted that the above method is only one of the update level analysis methods, which has the beneficial effect of fast speed.

[0134] See also Figure 8 The present application provides an embodiment of a screen color correction device, comprising: The first acquisition unit 801 is used to acquire a training data set for a detection model. The detection model is an artificial intelligence model that will be added to the current display production line to perform display defect detection. The training data set is display data used by the detection model during pre-training.

[0135] The second acquisition unit 802 is used to acquire a display screen acquisition data set in real time. The display screen acquisition data set is data collected in the current display screen production line and used for display screen defect detection.

[0136] The third acquisition unit 803 is used to filter out the display screen acquisition data containing the circuit area from the display screen acquisition data set, and acquire the generation time data of the display screen acquisition data containing the circuit area.

[0137] The second analysis unit 804 is used to perform defect area and defect type analysis on the display screen data containing the circuit area, and generate a defect feature label for each display screen data containing a defect in the circuit area.

[0138] The classification unit 805 is used to classify the display screen collected data after defect area and defect type analysis according to the display screen ID and the detection item.

[0139] The second generating unit 806 is used to generate a circuit structure hierarchy according to the classified display screen collected data, where the circuit structure hierarchy represents the structural level of the thin-sheet circuit in the circuit area.

[0140] The third generating unit 807 is used to bind the acquired generation time data, defect feature label, circuit structure level and the classified display screen acquisition data to generate a new display screen acquisition data set.

[0141] The first evaluation unit 808 is used to evaluate the data freshness of the display screen collected data set.

[0142] Optionally, the steps of the first evaluation unit 808 include: Select a target display screen to collect data from the new display screen collection data set.

[0143] The time decay function is determined according to the defect detection items of the target display screen data collection, and the time decay function has an unknown variable, that is, the generation time.

[0144] The business weight is generated according to the display screen model of the target display screen collected data. The business weight refers to the importance of the display screen corresponding to the display screen model to the training of the detection model.

[0145] The data change rate parameter corresponding to the display screen model is determined according to the display screen model of the target display screen data collection and the new display screen collection data set.

[0146] Freshness evaluation is performed based on the time decay function, business weight, data change rate parameters, and the generation time of the data collected by the target display screen.

[0147] Perform freshness evaluation on other display screen collection data in the new display screen collection data set and integrate them to form a data freshness evaluation result.

[0148] The second evaluation unit 809 is used to perform data importance evaluation on the display screen collected data set.

[0149] Optionally, the steps of the second evaluation unit 809 include: The structure weight is generated according to the circuit structure level of the target display screen data collection.

[0150] Determine the target display screen ID for collecting data from the target display screen, and analyze the defect feature label binding status of all display screen collection data for the same inspection project under the target display screen ID.

[0151] Generate detection weights for target display screen acquisition data based on binding analysis results.

[0152] Analyze the defect feature labels of all display screen collection data of the same inspection project under the target display screen ID, and generate defect weights for the target display screen collection data according to the type analysis results of the defect feature labels.

[0153] The importance of collecting data for the target display screen is evaluated based on the structure weight, detection weight and defect weight.

[0154] Perform importance assessment on other display screen collection data in the new display screen collection data set and integrate them to form a data importance assessment result.

[0155] The first analysis unit 810 is used to perform update level analysis on the display screen collected data set according to the data freshness evaluation result and the data importance evaluation result.

[0156] Optionally, the first analyzing unit 810 includes: A first freshness threshold, a second freshness threshold, a first importance score threshold, and a second importance score threshold are obtained, wherein the first freshness threshold is smaller than the second freshness threshold, and the first importance score threshold is smaller than the second importance score threshold.

[0157] A freshness level is generated for each display screen collected data in the display screen collected data set according to the data freshness evaluation result, the first freshness threshold and the second freshness threshold.

[0158] An importance level is generated for each display screen acquisition data in the display screen acquisition data set according to the data importance evaluation result, the first importance score threshold and the second importance score threshold.

[0159] An update level analysis is performed on each display screen collection data in the display screen collection data set according to the freshness level and the importance level, and an update level analysis result is generated.

[0160] The screening unit 811 is used to screen out a target data item set from the display screen acquisition data set according to the update level analysis result.

[0161] The updating unit 812 is used to update the screened target data item set into the training data set.

[0162] The sorting unit 813 is used to perform pre-training sorting for the display screen collection data in the target data item set according to the update level analysis result.

[0163] The pre-training unit 814 is used to use the updated training data set to pre-train the detection model before it is put into use.

[0164] The first generating unit 815 is used to evaluate the model performance of the detection model using the updated training data set and generate performance feedback information.

[0165] The deployment unit 816 is used to deploy the detection model to the display production line when the performance feedback information shows that the detection model performance meets the requirements.

[0166] The feedback unit 817 is used to send the performance feedback information of the detection model back to the production client for optimization in future model training.

[0167] See also Fig. 9 , the present application provides a device for dynamically updating a data set, comprising: Processor 901 , memory 902 , input-output unit 903 , and bus 904 .

[0168] The processor 901 is connected to the memory 902 , the input and output unit 903 , and the bus 904 .

[0169] The memory 902 stores a program, and the processor 901 calls the program to execute the following steps: Figure 1 , Figure 2 and Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7 The method in .

[0170] The present application provides a computer-readable storage medium, on which a program is stored, and when the program is executed on a computer, the program performs the following steps: Figure 1 , Figure 2 and Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7 The method in .

[0171] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0172] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0173] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the embodiments of the present application.

[0174] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0175] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, read-only memory), random access memory (RAM, random access memory), disk or optical disk and other media that can store program code.

Claims

1. A method for dynamically updating a data set, characterized in that: include: Acquire a training data set of a detection model, wherein the detection model is an artificial intelligence model that will be added to the current display screen production line to perform display screen defect detection, and the training data set is display screen data used by the detection model during pre-training; Acquire a display screen acquisition data set in real time, wherein the display screen acquisition data set is data collected in a current display screen production line for display screen defect detection; Performing data freshness evaluation on the display screen collected data set; Performing data importance evaluation on the display screen collected data set; Performing an update level analysis on the display screen collected data set according to the data freshness evaluation results and the data importance evaluation results; Filtering a target data item set from the display screen acquisition data set according to the update level analysis result; Update the filtered target data item set to the training data set.

2. The method according to claim 1, characterized in that After the step of updating the screened target data item set into the training data set, the method further includes: Using the updated training data set to pre-train the detection model before putting it into use; Evaluating the model performance of the detection model using the updated training data set to generate performance feedback information; When the performance feedback information shows that the performance of the detection model meets the standard, the detection model is deployed to the display production line; The performance feedback information of the detection model is sent back to the production client for optimization in future model training.

3. The method according to claim 2, characterized in that After the step of updating the screened target data item set into the training data set and before the step of using the updated training data set to pre-train the detection model before putting it into use, the method further includes: Pre-training and sorting are performed for the display screen collection data in the target data item set according to the update level analysis result.

4. The method according to claim 1, characterized in that: The detection model is an artificial intelligence model for performing defect detection on the circuit area of ​​the display screen; After the step of acquiring the display screen collected data set in real time and before the step of evaluating the data freshness of the display screen collected data set, the method further includes: Screening out the display screen acquisition data containing the circuit area from the display screen acquisition data set, and acquiring generation time data of the display screen acquisition data containing the circuit area; Perform defect area and defect type analysis on the display screen data collected containing the circuit area, and generate defect feature labels for each display screen data collected containing defects in the circuit area; Classify the collected display data after defect area and defect type analysis according to display ID and test items; Generating a circuit structure hierarchy according to the classified display screen acquisition data, wherein the circuit structure hierarchy represents a structural level of a thin-sheet circuit in a circuit area; The acquired generation time data, defect feature labels, circuit structure levels and the classified display screen acquisition data are bound to generate a new display screen acquisition data set.

5. The method according to claim 4, characterized in that The step of evaluating the data freshness of the display screen collected data set comprises: Select a target display screen acquisition data from the new display screen acquisition data set; Determine a time decay function according to the defect detection items of the target display screen collected data, wherein the time decay function has an unknown number of generation time; Generate a business weight according to the display screen model of the target display screen collection data, wherein the business weight refers to the importance of the display screen corresponding to the display screen model to the training of the detection model; Determine the data change rate parameter corresponding to the display screen model according to the display screen model of the target display screen data collection and the new display screen collection data set; Performing freshness evaluation according to the time decay function, the business weight, the data change rate parameter, and the generation time of the target display screen collected data; Perform freshness evaluation on other display screen collection data in the new display screen collection data set and integrate them to form a data freshness evaluation result.

6. The method according to claim 5, characterized in that The step of performing data importance evaluation on the display screen collected data set comprises: Generate structural weights according to the circuit structure level of the target display screen data collection; Determine the target display screen ID for collecting data of the target display screen, and analyze the defect feature label binding status of all display screen collection data of the same inspection item under the target display screen ID; Generating a detection weight for the target display screen acquisition data according to the binding analysis result; Analyze the defect feature labels of all display screen collection data of the same inspection item under the target display screen ID, and generate defect weights for the target display screen collection data according to the type analysis results of the defect feature labels; Collecting data for the target display screen to perform importance assessment according to the structure weight, the detection weight, and the defect weight; Perform importance assessment on other display screen collection data in the new display screen collection data set and integrate them to form a data importance assessment result.

7. The method according to any one of claims 1 to 6, characterized in that The step of performing update level analysis on the display screen collected data set according to the data freshness evaluation result and the data importance evaluation result comprises: Obtaining a first freshness threshold, a second freshness threshold, a first importance score threshold, and a second importance score threshold, wherein the first freshness threshold is smaller than the second freshness threshold, and the first importance score threshold is smaller than the second importance score threshold; Generating a freshness level for each display screen collected data in the display screen collected data set according to the data freshness evaluation result, the first freshness threshold and the second freshness threshold; Generating an importance level for each display screen acquisition data in the display screen acquisition data set according to the data importance evaluation result, the first importance score threshold, and the second importance score threshold; An update level analysis is performed on each display screen collection data in the display screen collection data set according to the freshness level and the importance level to generate an update level analysis result.

8. A device for dynamically updating a data set, characterized in that: include: A first acquisition unit is used to acquire a training data set of a detection model, wherein the detection model is an artificial intelligence model to be added to the current display screen production line to perform display screen defect detection, and the training data set is display screen data used by the detection model during pre-training; A second acquisition unit is used to acquire a display screen acquisition data set in real time, wherein the display screen acquisition data set is data collected in a current display screen production line and used for display screen defect detection; A first evaluation unit, configured to evaluate the data freshness of the display screen collected data set; A second evaluation unit, used for performing data importance evaluation on the display screen collected data set; A first analysis unit, configured to perform an update level analysis on the display screen collected data set according to a data freshness evaluation result and a data importance evaluation result; A screening unit, configured to screen out a target data item set from the display screen acquisition data set according to an update level analysis result; The updating unit is used to update the filtered target data item set into the training data set.

9. The device according to claim 8, characterized in that The device also includes: A pre-training unit, used to use the updated training data set to pre-train the detection model before it is put into use; A first generating unit, configured to evaluate the model performance of the detection model using the updated training data set and generate performance feedback information; A deployment unit, configured to deploy the detection model to a display production line when the performance feedback information shows that the detection model performance meets the standard; The feedback unit is used to send performance feedback information of the detection model back to the production client for optimization in future model training.

10. A computer-readable storage medium having a program stored thereon, wherein the program, when executed on a computer, performs the method according to any one of claims 1 to 7.

Images