Product performance prediction model construction method, device and computer equipment

Abstract

The present application relates to a method, apparatus, computer device, and storage medium for constructing a product performance prediction model. The method comprises: for each first set of data to be processed, selecting data from the first set of data to be processed as model processing data, dividing the model processing data into training data and verification data; performing model training based on the training data in all the model processing data, and determining the importance ranking of each product performance parameter based on the training results; verifying the first model based on the test data in all the model processing data to obtain a first model performance score; if the first model performance score is greater than a preset threshold, then based on the importance ranking result, removing a preset number of product performance parameters from the model processing data before the ranking, returning to the loop execution steps until the first model performance score is less than the preset threshold; and using the remaining product performance parameters of the final model processing data as identically distributed variables.

Description

Technical Field

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

[0002] With the development of science and technology, product performance prediction is becoming more and more common. Product performance prediction refers to predicting the performance of a product based on its various performance parameters. Currently, product performance prediction is often carried out through model construction.

[0003] The product has a corresponding version of the product control system. When the version number of the product control system is different, the control parameters of the variables related to the product performance will change, such as the control timing and control duration. As a result, when the version number of the product control system is different, the value range of the performance parameters will change. Existing solutions often build a model for predicting product performance for a single product control system version. When the version number changes, the value range of the performance parameters changes. At this time, when the performance prediction of the performance parameters corresponding to other versions of the product control system is still based on the model, the prediction result accuracy is very low. Summary of the Invention

[0004] Based on this, it is necessary to provide a product performance prediction model construction method, device, computer equipment, computer-readable storage medium and computer program product that can improve the model prediction accuracy in response to the above technical problems.

[0005] In a first aspect, the present application provides a method for constructing a product performance prediction model. The method comprises:

[0006] Acquire a complete data set, where the complete data set is a training set in a collection of data, and the collection of data includes product performance parameters corresponding to multiple versions of product control systems;

[0007] Based on the full data set, at least one data subset is obtained, and the full data set and each data subset are respectively merged with a test set in the collected data set to obtain a first data set to be processed;

[0008] For each first set of data to be processed, selecting data from the first set of data to be processed as model processing data and dividing the model processing data into training data and verification data;

[0009] Perform model training based on the training data in all model processing data to obtain a first model, and determine the importance ranking of each product performance parameter based on the training results. The importance is used to represent the degree of influence of the product performance parameter on the model prediction result;

[0010] Validating the first model based on validation data in all model-processed data to obtain a first model performance score, where the first model performance score is used to characterize model prediction accuracy;

[0011] If the performance score of the first model is greater than a preset threshold, then based on the importance sorting result, a preset number of product performance parameters before sorting are removed from the model-processed data, the model-processed data after removal is used as the model-processed data again, and the model training based on the training data in all the model-processed data is returned to obtain the first model step and the execution is continued until the performance score of the first model is less than the preset threshold;

[0012] The remaining product performance parameters of the final model-processed data are used as identically distributed variables, and a product performance prediction model is constructed based on the identically distributed variables to perform product performance prediction.

[0013] In one embodiment, constructing a product performance prediction model based on the identically distributed variables to perform product performance prediction includes:

[0014] Selecting data from the first set of data to be processed as model testing data;

[0015] Determining product performance parameters corresponding to the identically distributed variables from the model test data;

[0016] Testing the first model based on product performance parameters corresponding to the identically distributed variables to obtain a second model performance score;

[0017] determining spatial distribution information of product performance parameters corresponding to the identically distributed variables by a principal component analysis method, and determining a model performance score of the first model based on the identically distributed information;

[0018] If the second model performance score satisfies the corresponding conditions and / or the model performance score determined based on the spatial distribution information satisfies the corresponding conditions, a product performance prediction model is constructed based on the identically distributed variables to perform product performance prediction.

[0019] In one embodiment, before performing model training based on training data in all model processing data to obtain the first model, the method further includes:

[0020] Delete product performance parameters whose variable correlation coefficients in the first data set to be processed are lower than a threshold value, where the variable correlation coefficients are used to characterize the correlation between each product performance parameter and the product performance prediction result.

[0021] In one embodiment, obtaining at least one data subset based on the entire data set includes:

[0022] Based on the entire data set, obtaining multiple candidate subsets;

[0023] A target candidate subset that meets the screening conditions is determined from multiple candidate subsets as the data subset.

[0024] In one embodiment, the screening condition includes at least one of the following:

[0025] The proportion of the number of product control system versions corresponding to the target candidate subset in the number of product control system versions corresponding to the full data set exceeds a corresponding preset proportion;

[0026] The number of samples included in the target candidate subset accounts for a proportion of the number of samples included in the full data set that exceeds a corresponding preset proportion; the full data set includes samples corresponding to each version of the product control system, different samples corresponding to the same version of the product control system are distinguished based on the time of collection, and each sample includes product performance parameters corresponding to the corresponding version of the product control system;

[0027] The difference between the maximum value of the variable correlation coefficient of the target candidate subset and the maximum value of the variable correlation coefficient of the entire data set is greater than the corresponding preset difference, and the variable correlation coefficient is used to represent the correlation between the product performance parameter and the product performance prediction result;

[0028] The difference between the number of product control software versions corresponding to the target candidate subset and the number of product control system versions corresponding to other candidate subsets is greater than the corresponding preset difference.

[0029] In one embodiment, constructing a product performance prediction model based on the identically distributed variables to perform product performance prediction includes:

[0030] For each first data set to be processed, determining identically distributed variables from the first data set to be processed to obtain a second data set to be processed, determining a training set from the second data set to be processed, and dividing the training set into training data and validation data;

[0031] Performing model training based on the training data in the second set of data to be processed to obtain a second model, and determining an importance ranking of each product performance parameter based on the training results of the second model;

[0032] Validating the second model based on the validation data in the second set of data to be processed to obtain a third model performance score;

[0033] If the change in the performance score of the third model exceeds a preset range, then based on the importance ranking result of the product performance parameters in the second model, a preset number of product performance parameters with lower rankings are removed from the second set of data to be processed, and the removed second set of data to be processed is used as the second set of data to be processed again. The process returns to the step of training the model based on the training data in the second set of data to be processed to obtain the second model and continues until the change in the performance score of the third model does not exceed the preset range;

[0034] A product performance prediction model is constructed based on the remaining product performance parameters in the final second data set to be processed to predict product performance.

[0035] In one embodiment, the method further comprises:

[0036] The samples to be predicted are respectively input into the product performance prediction models constructed for the entire data set and the data subset to obtain multiple prediction results, and the optimal value or average value of the multiple prediction results is used as the product performance prediction result.

[0037] In a second aspect, the present application also provides a device for constructing a product performance prediction model. The device comprises:

[0038] A full set acquisition module is used to acquire a full set of data, where the full set of data is a training set in a collection of data, and the collection of data includes product performance parameters corresponding to multiple versions of product control systems;

[0039] a subset acquisition module, configured to acquire at least one data subset based on the full data set, and merge the full data set and each data subset with a test set in the collected data set to obtain a first data set to be processed;

[0040] a selection module, configured to select data from each first set of data to be processed as model processing data, and divide the model processing data into training data and validation data;

[0041] The first module is a construction module, which is used to perform model training based on the training data in all model processing data to obtain a first model, and determine the importance ranking of each product performance parameter based on the training results. The importance is used to represent the degree of influence of the product performance parameter on the model prediction result;

[0042] a model validation module, configured to validate the first model based on validation data in all model-processed data to obtain a first model performance score, wherein the first model performance score is used to characterize the prediction accuracy of the model;

[0043] a loop module configured to, if the first model performance score is greater than a preset threshold, remove a preset number of product performance parameters from the model-processed data based on the importance sorting result, use the removed model-processed data as the model-processed data again, return to the step of performing model training based on the training data in all the model-processed data, and continue executing the step of obtaining the first model until the first model performance score is less than the preset threshold;

[0044] The second module is a construction module, which is used to use the remaining product performance parameters of the final model processing data as identically distributed variables, and to construct a product performance prediction model based on the identically distributed variables to perform product performance prediction.

[0045] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned product performance prediction model construction method when executing the computer program.

[0046] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-mentioned method for constructing a product performance prediction model.

[0047] In a fifth aspect, the present application further provides a computer program product, comprising a computer program that, when executed by a processor, implements the steps of the above-mentioned method for constructing a product performance prediction model.

[0048] In this embodiment, the collected data set includes a training set and a test set. Since the collected data set includes product performance parameters corresponding to multiple versions of product control systems, the version numbers of the product control systems corresponding to the training set and the test set are not completely consistent. The full data set is the training set. The full data set is obtained, and at least one data subset is obtained based on the full data set. The first full set and each data subset are respectively merged with the test set in the collected data set to obtain a first data set to be processed, that is, the first data set to be processed includes both data from the training set and data from the test set. For each first data set to be processed, data is selected from the first data set to be processed as model processing data, and the model processing data is divided into training data and verification data, so that model training is performed based on the training data in all model processing data to obtain a first model. Based on the training results, the importance ranking of each product performance parameter is determined. The importance is used to characterize the degree of influence of the product performance parameter on the model prediction result. At the same time, based on the verification data in all model processing data, the first model is verified to obtain a first model performance score. The first model performance score is used to characterize the accuracy of model prediction. Accuracy. If the performance score of the first model is greater than the preset threshold, it means that the prediction accuracy of the first model is high and it can well distinguish the data of the training set and the data of the test set. Based on the importance sorting result, the product performance parameters of a preset number of items before sorting are removed from the model processing data, and the removed model processing data is used again as the model processing data. Return to the training data based on all the model processing data for model training, obtain the step of the first model and continue to execute until the performance score of the first model is less than the preset threshold. It can be seen that each iterative cycle of the present application can remove the product performance parameters that have a greater impact on the model prediction results until the performance score of the first model is less than the preset threshold, that is, the first model cannot well distinguish the training set and the test set. The remaining product performance parameters of the final model processing data are used as identically distributed variables, and a product performance prediction model is constructed based on the identically distributed variables to perform product performance prediction. Since the product performance prediction model is constructed by identically distributed variables, the product performance prediction model cannot well distinguish the product performance data corresponding to different versions of the product control system. Therefore, when the version number of the product control system changes, the prediction result is more accurate when the product performance prediction model is used for prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A flowchart of a method for constructing a product performance prediction model in one embodiment is shown;

[0050] Figure 2 A flowchart of a method for constructing a product performance prediction model in another embodiment;

[0051] Figure 3 A flowchart of a method for constructing a product performance prediction model in another embodiment;

[0052] Figure 4 A flowchart of a method for constructing a product performance prediction model in another embodiment;

[0053] Figure 5 Schematic diagram of model fusion;

[0054] Figure 6 A flowchart of a method for constructing a product performance prediction model in another embodiment;

[0055] Figure 7 A flowchart of a method for constructing a product performance prediction model in another embodiment;

[0056] Figure 8 This is a structural diagram of the product performance prediction and analysis module;

[0057] Figure 9 A structural block diagram of a device for building a product performance prediction model in one embodiment;

[0058] Figure 10 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0060] Since the product performance parameters have different value ranges for different product control system versions in the existing solution, for example, for the V1 version of the product control system, the value range of the product performance parameters is 0 to 10, and for the V2 version of the product control system, the value range of the product performance parameters is 10 to 20. A model is constructed for the V1 version of the product control system. The model can only accurately predict the performance of product performance parameters with a value range of 0 to 10, while the accuracy of performance prediction for product performance parameters corresponding to the V2 version of the product control system is very low. In order to improve the accuracy of product performance prediction, the present application solution is proposed.

[0061] In one embodiment, Figure 1 As shown, a method for constructing a product performance prediction model is provided. This embodiment takes the method applied to a gateway server as an example for explanation. The gateway server can be a product performance prediction and analysis module deployed in an edge computing gateway. The method includes the following steps:

[0062] Step S101: Acquire a complete data set, where the complete data set is a training set in a collection of data, and the collection of data includes product performance parameters corresponding to multiple versions of product control systems.

[0063] The product control system refers to the control program that controls the product to realize its functions. The control program is burned into the controller and has a corresponding number. The number of each control program is the version number of the product control system.

[0064] The gateway server collects data to obtain a collection data set, which includes samples corresponding to multiple versions of the product control system. For the same version of the product control system, samples at multiple times can be collected. Different samples corresponding to the same version of the product control system are distinguished based on the collection time, and each sample includes product performance parameters corresponding to the corresponding version of the product control system. For example, the collection data set includes samples corresponding to version A of the product control system, samples corresponding to version B of the product control system, and samples corresponding to version C of the product control system. Version A of the product control system includes samples collected at collection time 1 and samples collected at collection time 2. The samples at each collection time also include product performance parameter 1 and product performance parameter 2. The sample situation corresponding to version B or version C of the product control system is similar to that of version A of the product control system.

[0065] In this embodiment, the dimensions of the product performance parameters in each sample may be the same or different. For example, all samples may include X product performance parameters, or one sample may include X product performance parameters and another sample may include Y product performance parameters, where X is different from Y. Both X and Y are positive integers.

[0066] The collected data set includes samples that have already been labeled with a performance mean. These samples are used as the training set and are labeled S1. The collected data set also includes samples for which the performance mean needs to be predicted. These samples are used as the test set and are labeled S2. The gateway server preprocesses the training and test sets separately and uses the preprocessed training set as the full data set, which includes the samples labeled S1. The version numbers of the product control systems corresponding to the training and test sets are not completely consistent.

[0067] Step S102: Based on the full data set, at least one data subset is obtained, and the full data set and each data subset are respectively merged with a test set in the collected data set to obtain a first data set to be processed;

[0068] The gateway server obtains at least one data subset based on the full data set;

[0069] For the entire data set, the gateway server merges the entire data set with the test set (that is, the samples marked as S1 and the samples marked as S2 are merged) to obtain the first set of data to be processed corresponding to the entire data set; for the data subset, the gateway server also merges the data subset with the test set to obtain the first set of data to be processed corresponding to the data subset, that is, the entire data set has a corresponding first set of data to be processed, and the data subset also has a corresponding set of data to be processed.

[0070] For each data set to be processed, the following steps S103 to S107 are executed.

[0071] In this embodiment, each first to-be-processed data set includes samples corresponding to multiple versions of product control systems, and different samples corresponding to the same version of the product control system are distinguished based on the collection time.

[0072] Step S103: for each first data set to be processed, selecting data from the first data set to be processed as model processing data, and dividing the model processing data into training data and verification data;

[0073] The gateway server randomly selects data from the first set of data to be processed as model processing data and reserves the unselected data as model testing data. For example, 70% of the data may be selected as model processing data and 30% of the data may be reserved as model testing data.

[0074] The gateway server divides the model processing data into training data and verification data. The training data includes samples marked as S1 and samples marked as S2, and the verification data also includes samples marked as S1 and samples marked as S2.

[0075] Step S104: Perform model training based on the training data in all model-processed data to obtain a first model, and determine the importance ranking of each product performance parameter based on the training results. The importance is used to represent the degree of influence of the product performance parameter on the model prediction result.

[0076] The gateway server performs model training and optimization based on the training data in the model processing data to obtain the first model. At the same time, it determines the degree of influence of each product performance parameter on the model prediction result from the training results, obtains the weight of each product performance parameter based on the degree of influence, and uses the weight as the importance of the product performance parameter to rank the importance of each product performance parameter.

[0077] The first model is a classification model. The model training process is to divide the samples in the training data into the category corresponding to S1 or the category corresponding to S2. The model prediction result refers to the classification result of the sample by the first model. The performance of the first model is the accuracy of the model classification. If the model can accurately divide the sample marked as S1 into the category corresponding to S1, and accurately divide the sample marked as S2 into the category corresponding to S2, it proves that the model performs well. In any case, if the classification is wrong, it means that the model performs poorly.

[0078] Step S105: Validate the first model based on the validation data in all model-processed data to obtain a first model performance score, where the first model performance score is used to characterize the prediction accuracy of the model.

[0079] The gateway server verifies the first model based on the verification data in all model processing data to obtain a first model performance score. The first model performance score is used to characterize the model prediction accuracy. The higher the model prediction accuracy, the more the first model can distinguish between the training set marked as S1 and the test set marked as S2, that is, the more it can distinguish between product performance data collected for different versions of product control systems.

[0080] Step S106: If the performance score of the first model is greater than a preset threshold, then based on the importance ranking result, a preset number of product performance parameters are removed from the model processing data before sorting, and the model processing data after removal is used as the new model processing data. The model training is then returned to the training data in all the model processing data to obtain the first model and the steps are continued until the performance score of the first model is less than the preset threshold.

[0081] If the first model performance score is greater than a preset threshold, the gateway server removes a preset number of product performance data items before sorting from the model processing data based on the importance sorting result. The preset number is generally one, that is, the product performance data at the top of the sorting is removed.

[0082] The gateway server uses the removed model processing data as the model processing data in step S104 again, and returns to the loop to execute steps S104 to S106 until the first model performance score is less than the preset threshold. In this way, product performance parameters that have a greater impact on the model prediction results can be removed, thereby reducing the first model performance score, and making it impossible for the first model to distinguish well between the training set marked as S1 and the test set marked as S2.

[0083] Step S107 : Use the remaining product performance parameters of the final model-processed data as identically distributed variables, and construct a product performance prediction model based on the identically distributed variables to perform product performance prediction.

[0084] The gateway server uses the remaining product performance parameters of the final model-processed data as identically distributed variables. For each sample in the first set of data to be processed, the product performance parameters other than the identically distributed variables in the sample are eliminated, and a product performance prediction model is constructed based on the first set of data to be processed after the data is eliminated. That is, the corresponding identically distributed variables are finally determined for the entire data set and the corresponding product performance prediction model is constructed. The corresponding identically distributed variables are also determined for each data subset and the corresponding product performance prediction model is constructed.

[0085] For example, each sample of the first data set to be processed initially includes product performance parameter A, product performance parameter B, and product performance parameter C, and the identically distributed variables are product performance parameters B and C. Then, product performance parameter A is eliminated, and product performance parameters B and product performance parameters C are included in each sample of the first data set to be processed after the elimination of the data.

[0086] The gateway server jointly performs product performance prediction based on the product performance prediction model corresponding to the entire data set and the product performance prediction models corresponding to each data subset.

[0087] In this embodiment, the collected data set includes a training set and a test set. Since the collected data set includes product performance parameters corresponding to multiple versions of product control systems, the version numbers of the product control systems corresponding to the training set and the test set are not completely consistent. The full data set is the training set. The full data set is obtained, and at least one data subset is obtained based on the full data set. The first full set and each data subset are respectively merged with the test set in the collected data set to obtain a first data set to be processed, that is, the first data set to be processed includes both data from the training set and data from the test set. For each first data set to be processed, data is selected from the first data set to be processed as model processing data, and the model processing data is divided into training data and verification data, so that model training is performed based on the training data in all model processing data to obtain a first model. Based on the training results, the importance ranking of each product performance parameter is determined. The importance is used to characterize the degree of influence of the product performance parameter on the model prediction result. At the same time, based on the verification data in all model processing data, the first model is verified to obtain a first model performance score. The first model performance score is used to characterize the accuracy of model prediction. Accuracy. If the performance score of the first model is greater than the preset threshold, it means that the prediction accuracy of the first model is high and it can well distinguish the data of the training set and the data of the test set. Based on the importance sorting result, the product performance parameters of a preset number of items before sorting are removed from the model processing data, and the removed model processing data is used again as the model processing data. Return to the training data based on all the model processing data for model training, obtain the step of the first model and continue to execute until the performance score of the first model is less than the preset threshold. It can be seen that each iterative cycle of the present application can remove the product performance parameters that have a greater impact on the model prediction results until the performance score of the first model is less than the preset threshold, that is, the first model cannot well distinguish the training set and the test set. The remaining product performance parameters of the final model processing data are used as identically distributed variables, and a product performance prediction model is constructed based on the identically distributed variables to perform product performance prediction. Since the product performance prediction model is constructed by identically distributed variables, the product performance prediction model cannot well distinguish the product performance data corresponding to different versions of the product control system. Therefore, when the version number of the product control system changes, the prediction result is more accurate when the product performance prediction model is used for prediction.

[0088] In one embodiment, the selected identically distributed variables need to be verified, such as Figure 2 As shown, step S107 includes:

[0089] Step S201, selecting data from the first data set to be processed as model test data;

[0090] The model test data also includes samples labeled S1 and samples labeled S2.

[0091] Step S202, determining product performance parameters corresponding to identically distributed variables from the model test data;

[0092] For each sample in the model test data, the product performance parameters except the identically distributed variables in the sample are eliminated. The performance parameters contained in the model test data after eliminating the data are the product performance parameters corresponding to the identically distributed variables.

[0093] Step S203, testing the first model based on product performance parameters corresponding to the identically distributed variables to obtain a second model performance score;

[0094] In this embodiment, the second model performance score is the same as the first model performance score, and is also used to characterize the model prediction accuracy.

[0095] In this embodiment, the first model is tested to detect whether the first model is overfitting.

[0096] Step S204: determining spatial distribution information of product performance parameters corresponding to identically distributed variables using a principal component analysis method, and determining a model performance score of the first model based on the identically distributed information;

[0097] The spatial distribution information of product performance parameters corresponding to the same distribution variables refers to the similarity in data space distribution between the product performance parameters corresponding to the same distribution variables in the training data and the product performance parameters corresponding to the same distribution variables in the test data.

[0098] The gateway server can determine the similarity through the principal component analysis (PCA) method, and thus can measure the model performance of the first model based on the similarity. The higher the similarity, the less the model can distinguish between the sample marked as S1 and the sample marked as S2, indicating that the model performance is worse. It is necessary to ensure that the similarity is greater than a certain similarity threshold.

[0099] Step S205 : If the second model performance score satisfies the corresponding condition and / or the model performance score determined based on the spatial distribution information satisfies the corresponding condition, a product performance prediction model is constructed based on the identically distributed variables to perform product performance prediction.

[0100] The second model performance score meeting the corresponding condition may be a second model performance score model overfitting score threshold, which indicates that the first model is not overfitting.

[0101] In this embodiment, the server tests whether the first model is overfitting through model test data, and determines the spatial distribution information through principal component analysis to verify whether the product performance data corresponding to the same distribution variables in the training data and the product performance data corresponding to the same distribution variables in the test data are similar in the data space. If the similarity is greater than the similarity threshold, it means that the first model cannot distinguish the training set and the test set well. If the first model is not overfitting and cannot distinguish the training set and the test set well, it means that the determined same distribution variables are accurate, and the product performance prediction model is constructed based on the same distribution variables. The present application can verify the selected same distribution variables to avoid the wrong selection of the same distribution variables.

[0102] In one embodiment, since the number of samples in the first data set to be processed is huge, in order to reduce the data processing pressure of the gateway server, product performance parameters with variable correlation coefficients lower than a threshold in the first data set to be processed can be deleted in advance. In step S105, model training is performed based on the training data in all model processing data. Before obtaining the first model, the method further includes:

[0103] The product performance parameters whose variable correlation coefficients in the first to-be-processed data set are lower than a threshold are deleted. The variable correlation coefficients are used to characterize the correlation between each product performance parameter and the product performance prediction result.

[0104] In this embodiment, the variable correlation coefficient of each product performance parameter can be the maximum mutual information coefficient (MIC) between the product performance parameter and the product performance prediction result, or it can be the linear correlation coefficient between the product performance parameter and the product performance prediction result, such as the Pearson correlation coefficient, or the variable correlation coefficient can include both the MIC and the Pearson correlation coefficient. The variable correlation index is a comprehensive index that comprehensively considers the linear correlation relationship (Pearson) and the nonlinear correlation relationship mic, and its expression is: max(abs(pearsonr),mic).

[0105] Since the number of product performance parameters in the first data set to be processed is large, the variable correlation coefficient for a particular product performance parameter can be the average of the variable correlation coefficients of multiple product performance parameters. For example, if the first data set to be processed includes 1000 product performance parameters A, the variable correlation coefficient for product performance parameter A is the average of the variable correlation coefficients of the 1000 product performance parameters A.

[0106] In this embodiment, product performance parameters with variable correlation coefficients lower than a threshold value may be deleted in advance to reduce the amount of data processing, thereby alleviating the data processing pressure on the gateway server.

[0107] Reference Figure 3 In one embodiment, in step S102, obtaining at least one data subset based on the entire data set includes:

[0108] Based on the entire data set, multiple candidate subsets are obtained;

[0109] A target candidate subset that meets the screening conditions is determined from multiple candidate subsets as the data subset.

[0110] In this embodiment, the gateway server can obtain multiple candidate subsets from the entire data set. The multiple candidate subsets may contain overlapping samples or different samples. For example, if the entire data set only includes samples marked as S1, then the data subset also only includes samples marked as S1.

[0111] Reference Figure 3 , this embodiment specifically includes: step S1: based on the entire data set, obtaining multiple candidate subsets, and taking multiple candidate subset groups as a candidate subset set; S2: initializing the candidate subset set; S3: judging whether there is a candidate subset in the candidate subset set, and if there is no candidate subset, ending the loop;

[0112] If a candidate subset exists, execute step SS4: take the last ranked candidate subset in the candidate subset set as the target candidate subset; S5: determine whether the target candidate subset meets the screening condition. If the screening condition is met, execute step S6: add the target candidate subset as a data subset to the data subset set, remove the target candidate subset from the candidate subset set as a new candidate subset set, and return to the step of determining whether there are still candidate subsets in the candidate subset set to continue executing. If the screening condition is not met, execute step S7: remove the target candidate subset from the candidate subset set as a new candidate subset set, and return to the step of determining whether there are still candidate subsets in the candidate subset set to continue executing.

[0113] In this embodiment, a data subset that meets the filtering conditions can be filtered out. By setting the filtering conditions in the following embodiments, product performance parameters that weaken each other can be separated. Ultimately, when product performance prediction is performed for product performance parameters corresponding to a certain version of a product control system, the product performance prediction model constructed based on the data subset has higher prediction accuracy than the product performance prediction model constructed based on the entire data set.

[0114] In one embodiment, the screening condition includes at least one of the following:

[0115] The proportion of the number of versions of the product control system corresponding to the target candidate subset to the number of versions of the product control system corresponding to the entire data set exceeds the corresponding preset proportion;

[0116] The number of samples in the target candidate subset exceeds a preset ratio in the number of samples in the full data set; the full data set includes samples corresponding to each version of the product control system. Different samples corresponding to the same version of the product control system are distinguished based on the time of collection, and each sample includes product performance parameters corresponding to the corresponding version of the product control system.

[0117] The difference between the maximum value of the variable correlation coefficient of the target candidate subset and the maximum value of the variable correlation coefficient of the entire data set is greater than the corresponding preset difference. The variable correlation coefficient is used to represent the correlation between the product performance parameter and the product performance prediction result;

[0118] The difference between the number of product control software versions corresponding to the target candidate subset and the number of product control system versions corresponding to other candidate subsets is greater than the corresponding preset difference.

[0119] The exact method to obtain the maximum value of the variable correlation coefficient is to determine the variable correlation coefficient of each product performance parameter separately. For example, the correlation corresponding to any of the above-mentioned product performance parameters can be the maximum mutual information coefficient and / or Pearson correlation coefficient, and the maximum correlation among the variable correlation coefficients of multiple product performance parameters is taken as the maximum value of the variable correlation coefficient.

[0120] When predicting product performance based on the model constructed based on the product performance parameters corresponding to multiple versions of product control software, the impact of changes in certain product performance parameters on the performance mean will be numerically weakened. However, in the present application, the screening condition is set to the difference between the maximum value of the variable correlation coefficient of the target candidate subset and the maximum value of the variable correlation coefficient of the entire data set is greater than the corresponding preset difference, that is, relative to the entire data set, the maximum value of the variable correlation coefficient in the target candidate subset is larger, and the difference between the maximum value of the variable correlation coefficient in the target candidate subset and the maximum value of the variable correlation coefficient in the entire data set is greater than the corresponding preset difference. In this way, after using the target candidate subset as the data subset, the product performance parameters of different product control system versions that weaken each other can be separated. When predicting product performance for the product performance parameters corresponding to a certain version of the product control system, the product performance prediction model constructed based on the data subset is obviously more accurate than the product performance prediction model constructed based on the entire data set.

[0121] In one embodiment, after ensuring that the product performance prediction model cannot well distinguish the product performance parameters corresponding to different versions of the product control system, it is also necessary to improve the accuracy of the product performance prediction model. Therefore, in order to improve the accuracy of the product performance prediction model, this application also proposes the following solution: Figure 1The goal of the embodiment scheme is to distinguish between the training set and the test set. Therefore, the product performance parameters with higher importance ranking results need to be removed. The purpose of this embodiment is to screen effective modeling variables (modeling variables refer to product performance parameters). Therefore, the product performance parameters with the lowest importance ranking results need to be removed in each cycle. Figure 1 The termination condition of each iteration of the loop of the embodiment scheme is that the performance score of the first model is greater than the preset threshold. The termination condition of the iteration loop of this scheme is that the performance of the model no longer improves. The details of this scheme are as follows:

[0122] In one embodiment, referring to Figure 4 Step S108 constructs a product performance prediction model based on identically distributed variables to perform product performance prediction, specifically including:

[0123] Step S401: for each first data set to be processed, determine identically distributed variables from the first data set to be processed to obtain a second data set to be processed, determine a training set from the second data set to be processed, and divide the training set into training data and validation data;

[0124] In this embodiment, for each first data set to be processed, the gateway server removes product performance parameters other than identically distributed variables from each sample of the first data set to be processed to obtain a second data set to be processed.

[0125] Since screening valid variables only requires variable screening based on samples marked as S1, the gateway server determines a training set from the second set of data to be processed, that is, determines samples marked as S1, and divides the samples marked as S1 into training data and verification data.

[0126] Step S402: performing model training based on the training data in the second data set to be processed to obtain a second model, and determining the importance ranking of each product performance parameter based on the training result of the second model;

[0127] The gateway server performs model training and optimization based on the training data of the second model to obtain the second model. At the same time, it determines the degree of influence of each product performance parameter on the model prediction results from the training results, obtains the weight of each product performance parameter based on the degree of influence, and uses the weight as the importance of the product performance parameter to rank the importance of each product performance parameter.

[0128] In this embodiment, the second model includes a regression model.

[0129] Step S403: verifying the second model based on the verification data in the second data set to be processed to obtain a third model performance score;

[0130] The gateway server verifies the second model based on the verification data to obtain a third model performance score, where the third model performance score is used to characterize the prediction accuracy of the model.

[0131] Step S404: If the change in the performance score of the third model exceeds a preset range, then based on the importance ranking of the product performance parameters in the second model, a preset number of product performance parameters with the lowest ranking are removed from the second set of data to be processed, and the removed second set of data to be processed is used as the second set of data to be processed. The process returns to the step of training the model based on the training data in the second set of data to be processed to obtain the second model and continues until the change in the performance score of the third model does not exceed the preset range.

[0132] If the change in the performance score of the third model exceeds the preset range of change, it means that the performance of the second model is still improving and can be further optimized. Based on the importance ranking results of the product performance parameters in the second model, a preset number of product performance parameters with low ranking are removed from each sample of the training data. The preset number of items is generally one, that is, the product performance parameter with the lowest ranking can be removed.

[0133] The gateway server uses the removed model-processed data as the second set of unprocessed data for the second model in step S401, and loops back through steps S402 to S404 until the change in the third model performance score does not exceed a preset range. This removes product performance parameters that have a minor impact on the model's prediction results, making the second prediction model more accurate. Furthermore, the product performance prediction model constructed based on the remaining product performance parameters from the final model-processed data is more accurate.

[0134] Step S405: constructing a product performance prediction model based on the remaining product performance parameters in the final second data set to be processed to predict product performance.

[0135] The gateway server performs modeling and optimization based on the remaining product performance parameters in the final second set of data to be processed, and constructs a corresponding product performance prediction model to predict product performance.

[0136] In this embodiment, by eliminating product performance parameters with less impact on product performance prediction results in the second data set to be processed and selecting product performance parameters with greater impact on model prediction results for modeling, the accuracy of model prediction can be improved.

[0137] In one embodiment, referring to Figure 5 After integrating the product performance prediction models, the product performance is predicted together. The method further includes:

[0138] The samples to be predicted are input into the product performance prediction models constructed for the entire data set and the data subset respectively to obtain multiple prediction results, and the optimal value or average value among the multiple prediction results is used as the product performance prediction result.

[0139] As mentioned above, a product performance prediction model is constructed for the entire data set, and a product performance prediction model is also constructed for a subset of the data (e.g. Figure 5 As shown in the figure, a product performance prediction model is constructed for the entire data set, and a product performance prediction model is constructed for each subset of data 1 to n, where n is an integer greater than 1). When performing product performance prediction, the samples to be predicted are input into each product performance prediction model, and multiple product performance prediction results can be obtained (for example Figure 5 As shown, product performance prediction results 1 to n+1 are obtained. If the model construction data used to construct the product performance prediction model (i.e., the final second data to be processed mentioned above) includes product performance parameters of the same version of the product control system corresponding to the sample to be predicted, the optimal value of the multiple product performance prediction results is used as the product performance prediction result; otherwise, the average value of the multiple product performance prediction results is used as the product performance prediction result.

[0140] Reference Figure 6 The overall application plan is summarized as follows:

[0141] Step S601: Data collection.

[0142] Collect samples corresponding to multiple versions of product control systems. Different samples corresponding to the same version of the product control system are distinguished based on the time of collection, and each sample includes at least one product performance parameter. For example, hydrogen fuel cells collect control and controllable quantity data for the battery stack, hydrogen circuit, air circuit, cooling circuit, and electrical circuit of different battery systems, as well as monitoring heartbeat data, anomaly detection data, battery flag data, and other data. These data are product performance parameters. Information is extracted from the collected data (information extraction uses time-frequency domain statistical information extraction) to obtain candidate data.

[0143] The candidate data includes samples that have obtained performance mean labels. These samples are used as training sets and marked as S1. The candidate data also includes samples that need to predict performance mean. These samples are used as test sets and marked as S2.

[0144] Step S602: data preprocessing.

[0145] The training set in the candidate data is preprocessed to obtain the full data set. The test set is also preprocessed.

[0146] Step S603: data subset selection.

[0147] Select a data subset from the entire data set that meets the following conditions:

[0148] The number of product control system versions corresponding to the data subset accounts for 60% of the number of control system versions corresponding to the full data set;

[0149] The number of samples in the data subset accounts for 60% of the number of samples in the full data set;

[0150] The difference between the number of versions of the corresponding product control software and the number of versions of the product control systems corresponding to other candidate subsets is greater than 2;

[0151] The difference between the maximum value of the variable correlation coefficient of the data subset and the maximum value of the variable correlation coefficient of the entire data set is greater than the corresponding preset difference value greater than 0.05;

[0152] The entire data set only includes samples marked as S1, so the selected data subset also only includes samples marked as S1.

[0153] Step S604: extract the corresponding identically distributed variables for the entire data set, and extract the corresponding identically distributed variables for each data subset.

[0154] Reference Figure 7 , the specific implementation is as follows:

[0155] 1) Remove product performance parameters with variable correlation coefficients lower than 1 in the entire data set, and remove product performance parameters with variable correlation coefficients lower than 1 in the data subset (corresponding to the correlation coefficient threshold screening in the figure);

[0156] 2) For the entire data set, merge the test set (samples marked as S2) with the entire data set (the entire data set only includes samples marked as S1) to obtain the first set of data to be processed for the entire data set;

[0157] For the data subset, the test set (samples marked as S2) is merged with the data subset (the data subset only includes samples marked as S1) to obtain a first set of data to be processed for the data subset;

[0158] For each first set of data to be processed, execute the following steps 3) to 6) as follows:

[0159] 3) Randomly extract model processing data from the first set of data to be processed, for example, extract 70% of the data as model processing data, and reserve the remaining data as model prediction data, for example, reserve 30% of the data as model prediction data (corresponding to the randomization in the figure), and perform the following operations on the model processing data:

[0160] a) Divide the model-processed data (70% of the model-processed data includes samples labeled S1 and S2) into training data (the training data includes samples labeled S1 and S2) and validation data (the validation data includes samples labeled S1 and S2);

[0161] b) using the training data (the training data corresponds to samples 1 to k in the figure) to perform model training and optimization to obtain a first model;

[0162] c) Determine the importance ranking of each product performance parameter;

[0163] d) validating the first model using the validation data to obtain a performance score of the first model;

[0164] The first model is a classification model. If the first model can accurately classify samples into the categories corresponding to S1 or S2, for example, it can classify samples marked as S1 into the category of S1 and samples marked as S2 into the category of S2, then it means that the first model performs well.

[0165] e) When the mean of the first model performance scores of the five samplings is greater than 2, remove the top-ranked product performance parameters from the model-processed data and return to step b) until the first model performance score is less than 2;

[0166] f) The remaining product performance parameters of the model-processed data are used as identically distributed variables (steps c to e in this embodiment). Cross-validation and feature screening are performed on each sample in the corresponding figure. Feature screening refers to screening identically distributed variables. Cross-validation is to verify the first model with test data. Subsequently, based on the determined identically distributed variables, product performance parameters other than the identically distributed variables in the first data set to be processed are eliminated to obtain a new first data set to be processed, that is, the new first data set to be processed corresponds to candidate subset 1 in the figure.

[0167] 4) Use the model test data (i.e., the 30% of the predicted data) to calculate the second model performance score of the first model (corresponding to the candidate subset for the test set evaluation in the figure);

[0168] 5) The data marked as S1 in the model test data is used as training data, and the data marked as S2 is used as validation data. The principal component analysis method is used to determine the similarity of the data space distribution between the training data containing only identically distributed variables and the validation data containing only identically distributed variables (corresponding to the PCA inspection in the figure);

[0169] 6) If the performance score and similarity of the second model both meet the corresponding conditions, the identically distributed variables determined in step f are used as the final identically distributed variables (the final subset is determined in the corresponding figure).

[0170] The samples that do not belong to the same distribution variables in each sample of the first data set to be processed are eliminated to obtain the second data set to be processed.

[0171] Step S605 , screening product performance parameters for the entire data set, and also screening product performance parameters for each data subset.

[0172] For the entire data set, product performance parameters with higher importance ranking are screened out from the second set of data to be processed corresponding to the entire data set. For the data subset, product performance parameters with higher importance ranking are screened out from the second set of data to be processed corresponding to the data subset. The specific screening method is described in the above embodiment and will not be repeated here.

[0173] Step S606: construct product performance prediction models for the entire data set and each data subset.

[0174] The remaining product performance parameters in the second to-be-processed data set corresponding to the entire data set are used as training data for the product performance prediction model. The model is trained and optimized using the random forest algorithm to construct a product performance prediction model corresponding to the entire data set.

[0175] For each data subset, the remaining product performance parameters in the second to-be-processed data set corresponding to the data subset are used as training data for the product performance prediction model. The random forest algorithm is used to train and optimize the model to construct the product performance prediction model corresponding to the data subset.

[0176] Step S607: Model fusion to determine product performance prediction results.

[0177] The sample to be predicted is input into the product performance prediction model corresponding to the entire data set, and at the same time, it is input into the product performance prediction model corresponding to each data subset. If the version number of the product control system corresponding to the training data of one or more product performance prediction models is the same as the version number of the product control system corresponding to the sample to be predicted, then the optimal value of the multiple product performance prediction results is used as the product performance prediction result; otherwise, the average value of the multiple product performance prediction results is used as the product performance prediction result.

[0178] It should be noted that traditional predictions of product performance are often performed by servers. This method is relatively inefficient and requires relatively high network bandwidth, which can easily cause problems such as jamming, high latency, and low efficiency. However, this application can deploy the product performance analysis module on the edge computing gateway, and migrate the massive data processing and calculations that were originally handled by the server to the edge computing gateway for processing, which can avoid the jamming, high latency, low efficiency, and other problems caused by the server processing massive data. Of course, this application does not exclude the deployment of the product performance analysis module on the server, but there are problems with processing jams and delays compared to deploying it on the edge computing gateway. Specifically:

[0179] like Figure 8 The figure shows the structure of the product performance prediction and analysis module, which protects key data and ensures data security through authentication, encrypted transmission, access control, trusted computing, etc.; it collects time series data through communication protocol, interface module and driver protocol adaptation, and stores data in the relational library and time series library. The data processing includes data preprocessing, version subset selection, full set subset data extraction, feature variable extraction, and edge reasoning after data processing. Edge reasoning includes building algorithm models, model tuning and model fusion, and then visual display. The product performance prediction and analysis module also includes hardware modules, which include wifi, 4G and / or 5G modules, Ethernet modules and serial port modules. The serial port module connects to serial port devices, the Ethernet module connects to Ethernet interface devices, and the wifi, 4G and / or 5G modules communicate with the early warning terminal. The early warning terminal includes a mobile terminal and a personal computer (Personal Computer, PC) terminal.

[0180] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0181] Based on the same inventive concept, the present application also provides a device for constructing a product performance prediction model for implementing the aforementioned method for constructing a product performance prediction model. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations in one or more embodiments of the device for constructing a product performance prediction model provided below can be found in the aforementioned limitations on the method for constructing a product performance prediction model, and will not be further elaborated here.

[0182] In one embodiment, Figure 9 As shown, a product performance prediction model building device 900 is provided, comprising:

[0183] A full set acquisition module 901 is used to acquire a full set of data, where the full set of data is a training set in a collection of data, and the collection of data includes product performance parameters corresponding to multiple versions of product control systems.

[0184] A subset acquisition module 902 is configured to acquire at least one data subset based on the full data set, and merge the full data set and each data subset with a test set in the collected data set to obtain a first data set to be processed;

[0185] A selection module 903 is configured to select data from each first set of data to be processed as model processing data, and divide the model processing data into training data and validation data;

[0186] A first module construction module 904 is configured to perform model training based on the training data in all model processing data to obtain a first model, and to determine the importance ranking of each product performance parameter based on the training results, where the importance is used to represent the degree of influence of the product performance parameter on the model prediction result;

[0187] A model validation module 905 is configured to validate the first model based on validation data in all model-processed data to obtain a first model performance score, where the first model performance score is used to characterize the prediction accuracy of the model.

[0188] A loop module 906 is configured to, if the first model performance score is greater than a preset threshold, remove a preset number of product performance parameters from the model-processed data based on the importance sorting result, use the removed model-processed data as the new model-processed data, and return to the step of training the model based on the training data in all the model-processed data to obtain the first model, and continue executing the step until the first model performance score is less than the preset threshold;

[0189] The second module constructs a module 907, which is used to use the remaining product performance parameters of the final model-processed data as identically distributed variables, and to construct a product performance prediction model based on the identically distributed variables to perform product performance prediction.

[0190] In one embodiment, the second module constructs a module, specifically used to select data from the first set of data to be processed as model test data; determine the product performance parameters corresponding to the identically distributed variables from the model test data; test the first model based on the product performance parameters corresponding to the identically distributed variables to obtain a second model performance score; determine the spatial distribution information of the product performance parameters corresponding to the identically distributed variables through a principal component analysis method, and determine the model performance score of the first model based on the identically distributed information; if the second model performance score meets the corresponding conditions and / or the model performance score determined based on the spatial distribution information meets the corresponding conditions, then, based on the identically distributed variables, construct a product performance prediction model to perform product performance prediction.

[0191] In one embodiment, the device also includes: a deletion module, which is used to perform model training in the first model building module 905 based on the training data in all model processing data, and before obtaining the first model, delete the product performance parameters in the first set of data to be processed whose variable correlation coefficient is lower than a threshold, and the variable correlation coefficient is used to characterize the correlation between each product performance parameter and the product performance prediction result.

[0192] In one embodiment, the data subset acquisition module is specifically configured to acquire multiple candidate subsets based on the entire data set; and determine a target candidate subset that meets the screening conditions from the multiple candidate subsets as the data subset.

[0193] In one embodiment, the screening conditions include at least one of the following: the proportion of the number of versions of the product control system corresponding to the target candidate subset in the number of versions of the product control system corresponding to the entire data set exceeds the corresponding preset proportion; the proportion of the number of samples contained in the target candidate subset in the number of samples contained in the entire data set exceeds the corresponding preset proportion; the entire data set includes samples corresponding to each version of the product control system, and different samples corresponding to the same version of the product control system are distinguished based on the collection time, and each sample includes product performance parameters corresponding to the corresponding version of the product control system; the difference between the maximum value of the variable correlation coefficient of the target candidate subset and the maximum value of the variable correlation coefficient of the entire data set is greater than the corresponding preset difference, and the variable correlation coefficient is used to characterize the correlation between product performance parameters and product performance prediction results; the difference between the number of product control software versions corresponding to the target candidate subset and the number of product control system versions corresponding to other candidate subsets is greater than the corresponding preset difference.

[0194] In one embodiment, the second model building module is specifically configured to: for each first data set to be processed, determine identically distributed variables from the first data set to be processed to obtain a second data set to be processed, determine a training set in the second data set to be processed, and divide the training set into training data and verification data; perform model training based on the training data in the second data set to be processed to obtain a second model, and determine the importance ranking of each product performance parameter based on the training result of the second model; verify the second model based on the verification data in the second data set to be processed to obtain a third model performance score; if the change in the third model performance score exceeds a preset range of change, remove a preset number of product performance parameters with lower rankings from the second data set to be processed based on the importance ranking result of the product performance parameters in the second model, use the removed second data set as the second data set, return to the step of performing model training based on the training data in the second data set to obtain the second model and continue executing until the change in the third model performance score does not exceed the preset range of change; and construct a product performance prediction model based on the remaining product performance parameters in the final second data set to predict product performance.

[0195] In one embodiment, the device also includes a model prediction module, which is used to input the samples to be predicted into the product performance prediction models constructed for the entire data set and the data subset respectively, to obtain multiple prediction results, and take the optimal value or average value among the multiple prediction results as the product performance prediction result.

[0196] Each module in the aforementioned product performance prediction model construction device can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0197] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 10As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store product performance parameters. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for constructing a product performance prediction model is implemented.

[0198] Those skilled in the art will understand that Figure 10 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0199] In one embodiment, a computer device is further provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0200] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0201] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0202] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions.

[0203] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processor involved in the various embodiments provided herein may be, but are not limited to, a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, and the like.

[0204] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0205] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A method for constructing a product performance prediction model, characterized in that: The method comprises: Obtaining a complete data set, where the complete data set is a training set in a collection of data, and the collection of data includes product performance parameters corresponding to multiple versions of product control systems; the collection of data includes a training set and a test set, and the version numbers of the product control systems corresponding to the training set and the test set are not completely consistent; Based on the full data set, at least one data subset is obtained, and the full data set and each data subset are respectively merged with a test set in the collected data set to obtain a first data set to be processed; For each first set of data to be processed, selecting data from the first set of data to be processed as model processing data, and dividing the model processing data into training data and verification data; Perform model training based on the training data in all model processing data to obtain a first model, and determine the importance ranking of each product performance parameter based on the training results. The importance is used to represent the degree of influence of the product performance parameter on the model prediction result; Validating the first model based on validation data in all model-processed data to obtain a first model performance score, where the first model performance score is used to characterize model prediction accuracy; If the first model performance score is greater than a preset threshold, then based on the ranking result from high to low importance, a preset number of product performance parameters before sorting are removed from the model-processed data, the model-processed data after removal is used as the model-processed data again, and the model training based on the training data in all the model-processed data is returned to obtain the first model step and the execution is continued until the first model performance score is less than the preset threshold; The remaining product performance parameters of the final model-processed data are used as identically distributed variables, and a product performance prediction model is constructed based on the identically distributed variables to perform product performance prediction.

2. The method according to claim 1, characterized in that The constructing of a product performance prediction model based on the identically distributed variables to perform product performance prediction includes: Selecting data from the first set of data to be processed as model testing data; Determining product performance parameters corresponding to the identically distributed variables from the model test data; Testing the first model based on product performance parameters corresponding to the identically distributed variables to obtain a second model performance score; determining spatial distribution information of product performance parameters corresponding to the identically distributed variables by a principal component analysis method, and determining a model performance score of the first model based on the spatial distribution information; If the second model performance score satisfies the corresponding conditions and / or the model performance score determined based on the spatial distribution information satisfies the corresponding conditions, a product performance prediction model is constructed based on the identically distributed variables to perform product performance prediction.

3. The method according to claim 1, characterized in that Before performing model training based on the training data in all model processing data to obtain the first model, the method further includes: Delete product performance parameters whose variable correlation coefficients in the first data set to be processed are lower than a threshold value, where the variable correlation coefficients are used to characterize the correlation between each product performance parameter and the product performance prediction result.

4. The method according to claim 1, wherein The acquiring of at least one data subset based on the entire data set includes: Based on the entire data set, obtaining multiple candidate subsets; A target candidate subset that meets the screening conditions is determined from multiple candidate subsets as the data subset.

5. The method according to claim 4, characterized in that The screening conditions include at least one of the following: The proportion of the number of product control system versions corresponding to the target candidate subset in the number of product control system versions corresponding to the full data set exceeds a corresponding preset proportion; The number of samples included in the target candidate subset accounts for a proportion of the number of samples included in the full data set that exceeds a corresponding preset proportion; the full data set includes samples corresponding to each version of the product control system, different samples corresponding to the same version of the product control system are distinguished based on the time of collection, and each sample includes product performance parameters corresponding to the corresponding version of the product control system; The difference between the maximum value of the variable correlation coefficient of the target candidate subset and the maximum value of the variable correlation coefficient of the entire data set is greater than the corresponding preset difference, and the variable correlation coefficient is used to represent the correlation between the product performance parameter and the product performance prediction result; The difference between the number of versions of the product control systems corresponding to the target candidate subset and the number of versions of the product control systems corresponding to other candidate subsets is greater than the corresponding preset difference.

6. The method according to claim 1, characterized in that The constructing of a product performance prediction model based on the identically distributed variables to perform product performance prediction includes: For each first data set to be processed, determining identically distributed variables from the first data set to be processed to obtain a second data set to be processed, determining a training set from the second data set to be processed, and dividing the training set into training data and validation data; Performing model training based on the training data in the second set of data to be processed to obtain a second model, and determining an importance ranking of each product performance parameter based on the training results of the second model; Validating the second model based on the validation data in the second set of data to be processed to obtain a third model performance score; If the change in the performance score of the third model exceeds a preset range of change, then based on the ranking result of the product performance parameters in the second model from high to low importance, a preset number of product performance parameters with the lowest ranking are removed from the second set of data to be processed, and the removed second set of data to be processed is used as the second set of data to be processed again. The step of training the model based on the training data in the second set of data to be processed to obtain the second model is returned to and continued until the change in the performance score of the third model does not exceed the preset range of change; A product performance prediction model is constructed based on the remaining product performance parameters in the final second data set to be processed to perform product performance prediction.

7. The method according to claim 1, characterized in that The method further comprises: The samples to be predicted are respectively input into the product performance prediction models constructed for the entire data set and the data subset to obtain multiple prediction results, and the optimal value or average value of the multiple prediction results is used as the product performance prediction result.

8. A device for constructing a product performance prediction model, characterized in that: The device comprises: A full set acquisition module is used to acquire a full set of data, wherein the full set of data is a training set in a collection of data, and the collection of data includes product performance parameters corresponding to multiple versions of product control systems. The collection of data includes a training set and a test set, and the version numbers of the product control systems corresponding to the training set and the test set are not completely consistent. a subset acquisition module, configured to acquire at least one data subset based on the full data set, and merge the full data set and each data subset with a test set in the collected data set to obtain a first data set to be processed; A selection module is used for selecting each first set of data to be processed. Selecting data from the first set of data to be processed as model processing data, and dividing the model processing data into training data and verification data; The first module is a construction module, which is used to perform model training based on the training data in all model processing data to obtain a first model, and determine the importance ranking of each product performance parameter based on the training results. The importance is used to represent the degree of influence of the product performance parameter on the model prediction result; a model validation module, configured to validate the first model based on validation data in all model-processed data to obtain a first model performance score, wherein the first model performance score is used to characterize the prediction accuracy of the model; a loop module configured to, if the first model performance score is greater than a preset threshold, remove a preset number of product performance parameters from the model-processed data based on the sorting results from high to low importance, use the removed model-processed data as the new model-processed data, return to the step of performing model training based on the training data in all the model-processed data, and continue executing the steps of obtaining the first model until the first model performance score is less than a preset threshold; The second module is a construction module, which is used to use the remaining product performance parameters of the final model processing data as identically distributed variables, and to construct a product performance prediction model based on the identically distributed variables to perform product performance prediction.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images