Data prediction method and device, equipment, storage medium and program product

Abstract

The invention provides a data prediction method and device, equipment, a storage medium and a program product, multiple pre-trained prediction models are obtained, the multiple prediction models are obtained through machine learning training based on a first sample data set, and at least two prediction models are different in structure; after a prediction instruction is received, weight coefficients of a plurality of prediction models are determined, and the weight coefficients of the plurality of prediction models are weight values obtained by carrying out weight optimization solution on the plurality of prediction models based on the second sample data set; target prediction is carried out on the to-be-predicted data through the multiple prediction models, and model prediction values output by the multiple prediction models are obtained; and on the basis of the weight coefficient of each prediction model, carrying out weighted summation calculation on the model prediction values of the plurality of prediction models, and outputting the obtained target prediction value as a prediction result corresponding to the to-be-predicted data, so that the prediction capability of the model combination on unknown data can be enhanced, and the prediction accuracy and the decision-making capability of the system are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular, to a data prediction method, apparatus, device, storage medium, and program product. Background Art

[0002] With the development of artificial intelligence and big data technologies, machine learning models have been widely used in many fields such as financial prediction, industrial control, medical diagnosis, and intelligent manufacturing. In order to improve the understanding and prediction ability of complex data, models such as deep machine learning are widely used to build prediction systems. However, in practical applications, the accuracy of machine learning models often fails to meet the actual prediction requirements, resulting in insufficient prediction accuracy of the models for data, thereby affecting the decision-making ability of the system based on the prediction results. Summary of the Invention

[0003] The present invention provides a data prediction method, apparatus, device, storage medium, and program product to solve the problem in the related art that the model accuracy is difficult to meet the actual prediction requirements, resulting in insufficient prediction accuracy of the model and affecting the decision-making ability of the system.

[0004] In a first aspect, an embodiment of the present application provides a data prediction method, including: Obtaining a plurality of pre-trained prediction models, the plurality of prediction models being obtained through machine learning training based on a first sample data set, and at least two of the plurality of prediction models having different structures; After receiving a prediction instruction for the data to be predicted, determining weight coefficients of the plurality of prediction models, the weight coefficients of the plurality of prediction models being weight values obtained by performing weight optimization solution on the plurality of prediction models based on a second sample data set; Using the plurality of prediction models to respectively perform target prediction on the data to be predicted, and obtaining model prediction values respectively output by the plurality of prediction models; Based on the weight coefficients of each prediction model, performing weighted summation calculation on the model prediction values respectively output by the plurality of prediction models to obtain a target prediction value; Outputting the target prediction value as the prediction result corresponding to the data to be predicted.

[0005] In a second aspect, an embodiment of the present application provides a data prediction apparatus, including: An obtaining module, configured to obtain a plurality of pre-trained prediction models, the plurality of prediction models being obtained through machine learning training based on a first sample data set, and at least two of the plurality of prediction models having different structures; A determination module, configured to determine the weight coefficients of multiple prediction models after receiving a prediction instruction for data to be predicted. The weight coefficients of the multiple prediction models are weight values obtained by performing weight optimization on the multiple prediction models based on a second sample data set; A prediction module, configured to perform target prediction on the data to be predicted by using the multiple prediction models respectively, and obtain model prediction values output by the multiple prediction models respectively; A calculation module, configured to perform weighted summation calculation on the model prediction values output by the multiple prediction models respectively based on the weight coefficients of each prediction model, and obtain a target prediction value; An output module, configured to output the target prediction value as a prediction result corresponding to the data to be predicted, so as to process the data to be predicted based on the target prediction value.

[0006] In a third aspect, an embodiment of the present application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above data prediction method are implemented.

[0007] In a fourth aspect, an embodiment of the present application provides a readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the above data prediction method are implemented.

[0008] In a fifth aspect, an embodiment of the present application provides a computer program product, which includes a computer program. When the computer program is run, the above data prediction method is executed.

[0009] In one solution provided by the above data prediction method, apparatus, device, storage medium, and program product, multiple pre-trained prediction models are obtained. The multiple prediction models are obtained through machine learning training based on a first sample data set, and at least two of the multiple prediction models have different structures. After receiving a prediction instruction for the data to be predicted, the weight coefficients of the multiple prediction models are determined. The weight coefficients of the multiple prediction models are weight values obtained by performing weight optimization on the multiple prediction models based on a second sample data set. The multiple prediction models are used to perform target prediction on the data to be predicted respectively, and model prediction values output by the multiple prediction models are obtained. Based on the weight coefficients of each prediction model, weighted summation calculation is performed on the model prediction values output by the multiple prediction models respectively to obtain a target prediction value, and the target prediction value is output as the prediction result corresponding to the data to be predicted. On the one hand, using multiple heterogeneous prediction models for prediction can complement the advantages and disadvantages of different models in prediction performance. By integrating the prediction results of multiple prediction models, the prediction bias and overfitting risk of a single model can be reduced, and the overall prediction performance and prediction accuracy can be improved. On the other hand, by optimizing the weights of the models on the second sample data set, the importance of each model in the weighted combination can be dynamically adjusted according to the actual situation. The prediction process is more flexible and adaptable, enabling the prediction result to better adapt to new data distributions and various different evaluation metrics, enhancing the prediction ability of the model combination for unknown data, and thus improving the prediction accuracy and the decision-making ability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0011] Figure 1 is a schematic structural diagram of a data prediction system according to an embodiment of the present invention; Figure 2 is a schematic flowchart of a data prediction method according to an embodiment of the present invention; Figure 3 is Figure 2 a schematic implementation flowchart of step S20 in Figure 4 is another schematic flowchart of a data prediction method according to an embodiment of the present invention; Figure 5 is a schematic diagram of the training process of multiple prediction models according to an embodiment of the present invention; Figure 6It is the mean square error curve graph of the optimal weight method of multiple prediction models trained in an embodiment of the present invention and other models; Figure 7 It is the coefficient of determination curve graph of the optimal weight method of multiple prediction models trained in an embodiment of the present invention and other models; Figure 8 It is a schematic diagram of the acquisition process of the weight values of multiple prediction models in an embodiment of the present invention; Figure 9 It is Figure 8 a schematic diagram of an implementation process of step S7 in Figure 10 It is Figure 1 a schematic diagram of a structure of the data prediction device in Figure 11 It is a schematic diagram of a structure of a computer device in an embodiment of the present invention. Specific Embodiments

[0012] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without making creative efforts belong to the scope of protection of the present invention.

[0013] It should be understood that when used in the specification of the present invention and the appended claims, the term "comprising" indicates the presence of the described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their combinations. It should also be understood that the term "and / or" refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations. Additionally, the terms "first", "second", "third", etc. are only used for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0014] The reference to "an embodiment" described in the specification of the present invention means that a specific feature, structure or characteristic described in conjunction with the embodiment is included in one or more embodiments of the present invention. The terms "comprising", "including", "having" and their variants all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0015] It should be understood that the magnitudes of the sequence numbers of the steps in the following embodiments do not mean the order of execution is prior or subsequent. The execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0016] To illustrate the technical solution of the present invention, specific embodiments are used for illustration below.

[0017] It should be understood that with the development of artificial intelligence and big data technologies, machine learning models based on deep learning have been widely applied in many fields such as financial prediction, industrial control, medical diagnosis, and intelligent manufacturing. To improve the understanding and prediction capabilities for complex data, models such as deep machine learning are widely used to build prediction systems. However, in practical applications, the accuracy of machine learning models often fails to meet the actual prediction requirements, resulting in insufficient prediction accuracy of the models for data, thereby affecting the decision-making ability of the system based on the prediction results.

[0018] Among them, the main reasons for the low accuracy of machine learning models mainly include the following aspects: 1. Single model structure: When a machine learning model with a single structure faces complex data with high nonlinearity, time dynamics, or cross-domain characteristics, its modeling ability has bottlenecks and it is difficult to capture all the effective features in the data, resulting in poor performance of the trained model and insufficient prediction result accuracy; 2. Limited training data: The model usually relies on a large amount of high-quality training data to learn the data distribution characteristics. However, in actual scenarios, especially in the optical field, it is usually difficult to obtain a sufficient amount of measured data as training samples. The training samples of the model often have noise, imbalance, or limited coverage, resulting in insufficient accuracy of the trained model and inability to accurately generalize to unseen data; 3. Overfitting and underfitting problems are widespread: In order to pursue higher training accuracy, the model is prone to overfitting, that is, it performs well on the training set but the prediction ability decreases on the test set or actual data; while a model with too simple a structure may result in underfitting due to insufficient learning ability; 4. Difficulties in model update and maintenance: In scenarios where the data distribution changes over time, if a statically trained model fails to be updated or adapted to new data in a timely manner, its prediction accuracy will also gradually decrease.

[0019] In view of the above problems, the embodiments of the present application provide a data prediction method, apparatus, device, storage medium, and program product. By obtaining multiple pre-trained prediction models, the multiple prediction models are obtained through machine learning training based on a first sample data set, and at least two of the multiple prediction models have different structures. After receiving a prediction instruction for the data to be predicted, determine the weight coefficients of the multiple prediction models. The weight coefficients of the multiple prediction models are weight values obtained by performing weight optimization on the multiple prediction models based on a second sample data set. Use the multiple prediction models to perform target prediction on the data to be predicted respectively, and obtain the model prediction values output by the multiple prediction models respectively. Based on the weight coefficients of each prediction model, perform weighted summation calculation on the model prediction values output by the multiple prediction models respectively to obtain the target prediction value, and output the target prediction value as the prediction result corresponding to the data to be predicted, so as to perform relevant processing on the data to be predicted and its corresponding target object based on the target prediction value. On the one hand, since models with different structures may have different prediction effects on the same data, this embodiment uses multiple heterogeneous prediction models for prediction, which can complement each other's advantages and disadvantages in prediction performance. By integrating the prediction results of multiple prediction models, the prediction deviation and overfitting risk of a single model can be reduced, and the overall prediction performance and prediction accuracy can be improved. On the other hand, by optimizing the weights of the models on the second sample data set, the importance of each model in the weighted combination can be dynamically adjusted according to the actual situation, making the prediction process more flexible and adaptable, enabling the prediction result to better adapt to the new data distribution and various different evaluation indicators, enhancing the prediction ability of the model combination for unknown data, and thus improving the prediction accuracy and the decision-making ability of the system. In addition, by using multiple heterogeneous prediction models for synchronous prediction and then obtaining the final prediction value through weighted summation of different weights, the dependence of the model on the training samples can be reduced. During the model training process, multiple prediction models can be trained with a small number of training samples, and accurate prediction results can also be obtained through multi-model prediction and weight weighting in the future, reducing the impact of factors such as noise, imbalance, or limited coverage of the training samples on the model accuracy, enabling the multi-model ensemble to accurately generalize to unseen data, and improving the prediction accuracy and the decision-making ability of the system.

[0020] The data prediction method provided by the embodiments of the present invention can be applied in a data prediction system as shown in Figure 1 The data prediction system includes a server and a terminal device. Among them, the terminal device communicates with the server through a network. Among them, the terminal device includes, but is not limited to, various personal computers, laptop computers, smart phones, tablet computers, and portable wearable devices. For example, the terminal device can be an optical data acquisition device. The server can be implemented by an independent server or a server cluster composed of multiple servers.

[0021] In the actual application process, the server can obtain multiple pre-trained prediction models. The multiple prediction models are obtained through machine learning training based on the first sample dataset, and the multiple prediction models are deployed locally for subsequent calls. At least two of the multiple prediction models have different structures. In other embodiments, multiple prediction models can also be pre-trained by other devices and sent to the server so that the server can deploy the multiple prediction models locally for subsequent calls. That is, the server can obtain multiple pre-trained prediction models.

[0022] In the process of model application, that is, when the server executes the prediction instruction, the terminal device obtains the data to be predicted and sends the prediction instruction for the data to be predicted to the server. After receiving the prediction instruction for the data to be predicted, the server determines the weight coefficients of the multiple pre-trained prediction models. Among them, the weight coefficients of the multiple prediction models are weight values obtained by performing weight optimization on the multiple prediction models based on the second sample dataset. Then, the server uses the multiple prediction models to perform target prediction on the data to be predicted respectively, and obtains the model prediction values output by the multiple prediction models respectively. Based on the weight coefficients of each prediction model, a weighted sum calculation is performed on the model prediction values output by the multiple prediction models respectively to obtain the target prediction value, and the target prediction value is output as the prediction result corresponding to the data to be predicted, so as to perform related processing on the data to be predicted and its corresponding target object based on the target prediction value.

[0023] In this embodiment, using multiple heterogeneous prediction models for prediction can complement the advantages and disadvantages of different models in prediction performance. By integrating the prediction results of multiple prediction models, the prediction bias and overfitting risk of a single model can be reduced, and the overall prediction performance and prediction accuracy can be improved. On the other hand, by optimizing the weights of the models on the second sample dataset, the importance of each model in the weighted combination can be dynamically adjusted according to the actual situation, making the prediction process more flexible and adaptable, enabling the prediction result to better adapt to the new data distribution and various different evaluation indicators, enhancing the prediction ability of the model combination for unknown data, and thus improving the prediction accuracy and the decision-making ability of the system. In addition, multiple prediction models can be trained with a small number of training samples, and accurate prediction results can also be obtained through multi-model prediction and weight weighting in the future, reducing the impact of the training samples on the model accuracy due to noise, imbalance, or limited coverage, etc., enabling the multi-model ensemble to accurately generalize to unseen data, and improving the prediction accuracy and the decision-making ability of the system.

[0024] In one embodiment, as Figure 2 shown, a data prediction method is provided. Taking the application of this method in a data prediction device as an example for illustration. Among them, the data prediction device can be Figure 1The terminal device in. The method specifically includes the following steps: S10: Obtain multiple pre-trained prediction models, which are obtained by performing machine learning training on a first sample data set.

[0025] In the actual application process, the server can obtain multiple pre-trained prediction models, and at least two of the multiple prediction models have different structures. Among them, the server can pre-train multiple machine learning models with different structures based on the first sample data set to obtain multiple prediction models, and deploy the multiple obtained prediction models locally for subsequent calls. In other embodiments, multiple prediction models can also be pre-trained by other devices and sent to the server so that the server can deploy the multiple prediction models locally for subsequent calls.

[0026] Among them, the multiple prediction models can include at least two types of models such as a support vector regression model, a random forest model, and an extreme gradient boosting (XGBoost) model or other machine learning models. In other embodiments, the multiple prediction models can also be multiple neural network models with different network structures; or, the multiple prediction models can also include at least one of a support vector regression model, a random forest model, and an XGBoost model, and at least one neural network model among multiple neural network models with different network structures. The multiple neural network models with different network structures can include at least two of models such as a feedforward neural network model, a convolutional neural network model, a recurrent neural network model, and a long short-term memory network model.

[0027] S20: After receiving a prediction instruction for the data to be predicted, determine the weight coefficients of the multiple prediction models, where the weight coefficients of the multiple prediction models are weight values obtained by performing weight optimization solution on the multiple prediction models based on a second sample data set.

[0028] After obtaining the multiple pre-trained prediction models, the server receives a prediction instruction for the data to be predicted and obtains the data to be predicted. Among them, the user can upload the data to be predicted through the terminal device and send a prediction instruction for the data to be measured to the server through the terminal device. The server receives the prediction instruction and obtains the data to be predicted carried by the prediction instruction.

[0029] In other embodiments, the terminal device can also have a data collection function. In the actual application process, the terminal device collects data on the target object in real time to obtain the data to be predicted, and then the terminal device sends the data to be predicted and a prediction instruction for the data to be predicted to the server. The server receives the prediction instruction and obtains the data to be predicted.

[0030] After receiving a prediction instruction for the data to be predicted, the server determines the weight coefficients of multiple prediction models. Among them, the weight coefficients of the multiple prediction models are weight values obtained by optimizing the weights of the multiple prediction models based on the second sample data set. Among them, the sample data in the second sample data set is different (or can be the same) from the sample data in the first sample data set and is derived from the same historical data sample set.

[0031] By actively solving the optimal weight combination of multiple heterogeneous prediction models on the second sample data set using mathematical programming methods, the prediction errors caused by a single weight or setting fixed weights are reduced, the accuracy of different prediction models is improved, and thus the prediction accuracy based on weighted integration of weights can be improved, realizing a double improvement in the model fitting ability and generalization performance in small-sample scenarios.

[0032] S30: Use multiple prediction models to respectively perform target prediction on the data to be predicted, and obtain the model prediction values output by the multiple prediction models.

[0033] After receiving a prediction instruction for the data to be predicted, the server uses multiple prediction models to respectively perform target prediction on the data to be predicted, and obtain the model prediction values output by the multiple prediction models. That is, the data to be predicted is respectively input into multiple prediction models, so that each prediction model respectively performs target prediction on the input data to be predicted and outputs the model prediction value for the data to be predicted, thereby obtaining the model prediction values output by the multiple prediction models.

[0034] S40: Based on the weight coefficients of each prediction model, perform weighted summation calculation on the model prediction values respectively output by the multiple prediction models to obtain the target prediction value.

[0035] After determining the weight coefficients of the multiple prediction models and the model prediction values output by each prediction model, the server performs weighted summation calculation on the model prediction values respectively output by the multiple prediction models based on the weight coefficients of each prediction model to obtain the target prediction value.

[0036] Among them, the target prediction value is calculated through the following formula: ; Among them, represents the target prediction value; represents the model prediction value output by the prediction model ; represents the weight coefficient of the prediction model ; represents the number of prediction models, .

[0037] S50: Output the target prediction value as the prediction result corresponding to the data to be predicted.

[0038] After obtaining the target prediction value, the server outputs the target prediction value as the prediction result corresponding to the data to be predicted, so that the user can timely learn the prediction result of the data to be predicted and send a processing instruction for the data to be predicted or the target corresponding to the data to be predicted, enabling the server to process the data to be predicted or the target corresponding to the data to be predicted in response to the processing instruction, and improving the system's decision-making ability based on the prediction result.

[0039] In this embodiment, multiple heterogeneous prediction models are used for prediction, which can complement the advantages and disadvantages of different models in prediction performance. By integrating the prediction results of multiple prediction models, the prediction bias and overfitting risk of a single model can be reduced, and the overall prediction performance and prediction accuracy can be improved. On the other hand, by optimizing the weights of the models on the second sample dataset, the importance of each model in the weighted combination can be dynamically adjusted according to the actual situation, making the prediction process more flexible and adaptable, enabling the prediction result to better adapt to the new data distribution and various different evaluation metrics, enhancing the prediction ability of the model combination for unknown data, and thus improving the prediction accuracy and the system's decision-making ability. In addition, multiple prediction models can be trained with a small number of training samples, and accurate prediction results can also be obtained through multi-model prediction and weight weighting in the follow-up, reducing the impact of the training samples on the model accuracy due to noise, imbalance, or limited coverage, enabling the multi-model ensemble to accurately generalize to unseen data, and improving the prediction accuracy and the system's decision-making ability.

[0040] In one embodiment, as Figure 3 shown, in step S20, that is, determining the weight coefficients of multiple prediction models, specifically includes the following steps: S201: Determine the type of prediction data for target prediction of the data to be predicted.

[0041] After receiving the prediction instruction for the data to be detected, the server can determine the type of prediction data for target prediction of the data to be predicted, that is, determine the data type of the prediction value for target prediction of the data to be detected. Among them, the type of prediction data can be determined by the received prediction instruction; the prediction instruction indicates the prediction target for predicting the data to be predicted (i.e., the expected prediction value), and the type of prediction data can be determined through the data type of the prediction target in the prediction instruction.

[0042] For example, the data to be predicted may be spectral data of a target object (such as a semiconductor device). Then, the prediction instruction for the data to be detected may be to indicate that based on the spectral data of the target object, predict the critical dimensions (such as length, width, and height) of the target object, or predict the thickness of the target object. That is, the prediction target for predicting the data to be predicted is the critical dimension or thickness, and the prediction data type is the dimension (size) prediction type, that is, the first type. Or, the prediction instruction for the data to be detected may also be to indicate that based on the spectral data of the target object, predict the probability of defects in the target object. That is, the prediction target for predicting the data to be predicted is the defect probability, and the prediction data type is the probability prediction type, that is, the second type. In other embodiments, the data to be predicted may be other data, and the prediction instruction and prediction target for the data to be detected may also be other, which will not be elaborated here.

[0043] S202: Obtain multiple weight data groups obtained by pre-solving. The weight data group includes weight values corresponding to multiple prediction models. The multiple weight values in each weight data group are obtained by performing weight optimization on the multiple prediction models based on the second sample data set.

[0044] Meanwhile, the server can read the weight data for multiple prediction models pre-stored in the local database. The weight data includes one or more weight data groups. Each weight data group includes weight values corresponding to multiple prediction models. Among them, the multiple weight values in each weight data group are weight values obtained by performing weight optimization on the multiple prediction models based on the second sample data set. When there are multiple weight data groups obtained by pre-solving stored in the server local database, each weight data group corresponds to a prediction data type. That is, the server performs weight optimization on the multiple prediction models based on the second sample data set, using different weight solving algorithms for different prediction data types, to obtain weight data groups corresponding to different prediction data types.

[0045] When there is one weight data group obtained by pre-solving stored in the server local database, the server directly assigns the weight values corresponding to the prediction models in the weight data group to the corresponding prediction models to obtain the weight coefficients of the prediction models.

[0046] S203: Determine the target weight data group corresponding to the prediction data type among the multiple weight data groups, and determine the corresponding weight values in the target weight data group as the weight coefficients of each prediction model.

[0047] When there are multiple weight data groups obtained by pre-solving stored in the server local database, the server determines the target weight data group corresponding to the prediction data type among the multiple weight data groups, and determines the corresponding weight values in the target weight data group as the weight coefficients of each prediction model.

[0048] In this embodiment, the type of prediction data for target prediction of the data to be predicted is determined, and then the target weight data group corresponding to the prediction data type is determined from multiple weight data groups, and the corresponding weight values in the target weight data group are determined as the weight coefficients of each prediction model. Each set of weights is the optimal solution optimized based on a specific type of prediction data, enabling each prediction task to use the most suitable model weight combination method, enhancing the adaptability and generalization ability of model integration; moreover, by determining the type of data to be predicted and then selecting the weight data group optimized in advance for this type, the model integration can be more in line with the data characteristics, thereby effectively improving the prediction accuracy. In addition, since the multiple weight data groups are obtained by offline pre-solving, only the corresponding weight group needs to be selected according to the data type during actual prediction, without the need to solve the optimization problem in real time, significantly enhancing the efficiency and real-time performance of the system's online prediction.

[0049] In a specific embodiment, the data to be predicted can be the spectral data of a target object (such as a semiconductor device), and the prediction instruction for the data to be predicted can be a prediction instruction for key dimension prediction of the target object based on the spectral data of the target object. Taking the prediction instruction of key dimension prediction of the target object based on the spectral data of the target object as an example, as Figure 4 shown, the data prediction method provided in this embodiment specifically includes the following steps: S101: Obtain multiple pre-trained prediction models, where the multiple prediction models are obtained through machine learning training based on a first sample data set.

[0050] Among them, the first sample data set includes multiple spectrogram samples. That is, the server uses multiple spectrogram samples to train two models among the support vector regression model, the random forest model, or multiple heterogeneous neural network models, to obtain multiple prediction models including at least two heterogeneous prediction models. The prediction model is used to predict the key dimensions of an object based on the spectral data (i.e., spectrogram) of the object.

[0051] S102: After receiving a prediction instruction for key dimension prediction of the target object, determine the weight coefficients of the multiple prediction models, where the weight coefficients of the multiple prediction models are weight values obtained by performing weight optimization solution on the multiple prediction models based on a second sample data set.

[0052] The server receives a prediction instruction for the data to be predicted, that is, receives a prediction instruction for key dimension prediction of the target object based on the spectral data of the target object, and obtains the spectral data of the target object. Among them, the user can upload the spectral data of the target object through a terminal device and send the prediction instruction for this to the server through the terminal device. The server receives the prediction instruction and obtains the spectral data of the target object carried by the prediction instruction.

[0053] Among them, the second sample data set includes multiple spectrogram samples. After obtaining multiple spectrogram samples of the same type of item collected historically, the multiple spectrogram samples are divided into a first sample data set and a second sample data set, and each data set contains at least two different spectrogram samples. After training multiple prediction models based on the multiple spectrogram samples in the first sample data set, the server can optimize the weights of the multiple prediction models based on the multiple spectrogram samples in the second sample data set to obtain at least one weight data group, and the weight data group includes weight values corresponding to the multiple prediction models.

[0054] After receiving a prediction instruction for the data to be predicted, that is, after receiving a prediction instruction for performing critical dimension prediction on the target object based on the spectral data of the target object, the server determines the weight coefficients of the multiple prediction models. For example, when multiple weight data groups are stored in the server, the prediction data type for target prediction of the data to be predicted is determined according to the prediction instruction, the target weight data group corresponding to the prediction data type is determined among the multiple weight data groups, and the corresponding weight values in the target weight data group are determined as the weight coefficients of each prediction model.

[0055] S103: Use multiple prediction models to perform critical dimension prediction on the spectral data of the target object respectively, and obtain model prediction values respectively output by the multiple prediction models.

[0056] After receiving a prediction instruction for the data to be predicted, the server uses multiple prediction models to perform critical dimension prediction on the spectral data of the target object respectively, and obtains model prediction values respectively output by the multiple prediction models. That is, the spectral data of the target object is input into the multiple prediction models respectively, so that each prediction model performs critical dimension prediction on the spectral data of the target object and outputs the size prediction value of the spectral data of the target object, thereby obtaining the model prediction values output by the multiple prediction models.

[0057] S104: Based on the weight coefficients of each prediction model, perform weighted summation calculation on the model prediction values respectively output by the multiple prediction models to obtain the target prediction value.

[0058] After determining the weight coefficients of the multiple prediction models and the model prediction values output by each prediction model, the server performs weighted summation calculation on the model prediction values respectively output by the multiple prediction models based on the weight coefficients of each prediction model to obtain the target prediction value, that is, to obtain the target prediction value of the critical dimension of the target object.

[0059] S105: Output the target prediction value as the prediction result of the critical dimension of the target object.

[0060] After obtaining the target prediction value, the server outputs the target prediction value as the prediction result of the critical dimension of the target object, so that the user can timely know the prediction result of the critical dimension of the target object, and send a processing instruction for the target object, so that the server responds to the processing instruction to process the target object, such as adjusting the parameter model of the target object to improve the decision-making ability of the system based on the prediction result.

[0061] In this embodiment, multiple heterogeneous prediction models are used for critical dimension prediction, which can complement the advantages and disadvantages of different models in prediction performance. By integrating the critical dimension prediction results of multiple prediction models, the prediction deviation and overfitting risk of a single model can be reduced, and the overall prediction performance and the prediction accuracy of the critical dimension of the object can be improved. On the other hand, by optimizing the weights of the models on the second sample data set, the importance of each model in the weighted combination can be dynamically adjusted according to the actual situation, and the prediction process is more flexible and adaptable, so that the prediction result of the critical dimension can better adapt to the new data distribution, which can enhance the prediction ability of the model combination to predict the critical dimension based on spectral data, thereby improving the prediction accuracy, enabling the server to adaptively adjust the parameters of the target object based on the critical dimension of the object, and improving the decision-making ability of the system.

[0062] In addition, multiple prediction models can be trained with a small number of spectral map samples, and accurate prediction results can also be obtained through multi-model prediction and weight weighting in the follow-up, reducing the impact of training samples on the model accuracy due to noise, imbalance, limited coverage, etc., enabling the multi-model ensemble to accurately generalize to subsequent spectral data, improving the prediction accuracy of the critical dimension of the object and the decision-making ability of the system. At the same time, the embodiment of the present application also solves the problems of poor model accuracy and poor critical dimension prediction effect caused by insufficient samples in the optical field, and improves the accuracy of the model to predict the critical dimension of the object based on spectral data.

[0063] In one embodiment, the server can train multiple initial models with different structures using the first sample data set to obtain multiple prediction models for subsequent use of the multiple prediction models to perform target prediction on the data to be predicted. Among them, as Figure 5 shown, the multiple prediction models are trained through the following steps: S1: Obtain multiple historical data samples, and the historical data samples and the data to be predicted are of the same type of data.

[0064] The server obtains multiple historical data samples. Among them, the historical data samples and the data to be predicted are of the same type of data. For example, if the data to be predicted is the spectral data of an object, the historical data samples are also the spectral data (i.e., spectral maps) of the same type of object (or the same target object), such as the historical spectral data obtained by collecting spectral maps of semiconductor devices.

[0065] S2: Perform feature augmentation processing on multiple historical data samples to obtain a first sample dataset including multiple training samples.

[0066] After obtaining multiple historical data samples, perform dataset partitioning on the multiple historical data samples to obtain an initial training set, a generalization set (i.e., the second sample dataset, used for solving the weights of the model), and a test set. Among them, the number of historical data samples in the training set and the prediction set is less than the number of historical data samples in the test set to simulate the scenario of small-sample machine learning.

[0067] For example, the number of historical data samples obtained this time is 1,989. If 1,989 real historical spectrograms of a certain target object are obtained, the server divides the first dozens of historical data samples among the 1,989 historical data samples into the initial training set, divides the last dozens of historical data samples among the 1,989 historical data samples into the generalization set, and divides the remaining other historical data samples into the test set to simulate the scenario of small-sample machine learning. Use the initial training set to train the model and use the test set to perform performance testing on the trained model until multiple prediction models with converged performance are obtained. Then, use the generalization set to perform weight optimization solution on the multiple prediction models to obtain at least one weight data group to obtain the weight coefficients of the multiple prediction models. Using data of the same type in the same dataset for model training and weight solution of the model can improve the accuracy of the weights and thus improve the accuracy of the target prediction value obtained by subsequent weighted summation.

[0068] Specifically, after obtaining multiple historical data samples and partitioning the dataset, the server performs feature augmentation processing on the multiple historical data samples in the initial training set to obtain multiple training samples to form a first sample dataset including multiple training samples.

[0069] S3: Perform dataset construction on the multiple training samples in the first sample dataset to obtain at least one target sample dataset.

[0070] After obtaining the first sample dataset, the server performs dataset construction on the multiple training samples in the first sample dataset to obtain at least one target sample dataset. Among them, the server can use the sampling algorithm and / or the bootstrap sampling algorithm (such as the Bagging bootstrap sampling algorithm) in the k-fold cross-validation method to perform data sampling on the multiple training samples in the first sample dataset to construct multiple datasets, thereby obtaining multiple target sample datasets.

[0071] S4: Use the first sample dataset and at least one target sample dataset to perform model training on multiple heterogeneous initial models to obtain multiple prediction models.

[0072] After obtaining the first sample data set and the target sample data set, the server uses the first sample data set and at least one target sample data set to train multiple heterogeneous initial models to obtain multiple prediction models. Among them, the multiple initial models are machine learning models with different structures.

[0073] For example, use the first sample data set to train one or more initial models, and use the test set to test the trained initial models until the performance of the initial models meets the requirements, obtaining one or more prediction models with model convergence; at the same time, use the target sample data set to train one or more initial models, and use the test set to test the trained initial models until the performance of the initial models meets the requirements, obtaining one or more prediction models with model convergence; output the prediction models obtained by training with the two data sets to obtain multiple prediction models.

[0074] In this embodiment, by performing feature augmentation processing on historical data samples, potential data features can be mined, a higher-dimensional input space can be constructed, the model's learning ability for complex patterns can be enhanced, and the target sample data set can be constructed using the first sample data set after feature augmentation, making the training data more comprehensive to cover data distributions in different situations or states, enhancing the robustness and generalization ability of the model, and improving the model prediction performance. In this solution, through systematic feature augmentation and sample construction of historical data samples, multiple structurally heterogeneous prediction models are trained, thereby significantly improving the model's learning ability for complex data features and the accuracy, robustness, and scalability of the overall prediction system.

[0075] In one embodiment, in step S2, that is, using the first sample data set and at least one target sample data set to train multiple initial models to obtain multiple prediction models, specifically includes the following steps: S21: Extract features from multiple historical data samples to obtain feature vectors of the multiple historical data samples, and fill in missing values for the vector matrix formed by the multiple feature vectors to obtain the first vector matrix.

[0076] After obtaining multiple historical data samples and dividing the dataset to get an initial training set, the server performs feature extraction on each of the multiple historical data samples in the initial training set to obtain the feature vectors of each historical data sample. Then, the server converts the feature vectors of each historical data sample into a vector matrix, obtaining a vector matrix formed by the feature vectors of multiple historical data samples. Among them, each row (i.e., the feature row) in this vector matrix represents the feature vector of a historical data sample, and each feature vector includes multiple feature values. The feature values at the same position in multiple feature vectors form a feature column. That is, each column (i.e., the feature column) in this vector matrix includes the feature values at the same position in multiple feature vectors, that is, the feature vector of a historical data sample consists of multiple feature values. The number of rows in this vector matrix is the number of historical data samples, and the number of columns in this vector matrix is the number of feature values in the feature vector.

[0077] For example, the historical data sample can be a spectrogram sample. The server performs feature extraction on each of the multiple spectrogram samples in the initial training set at wavelength points to obtain the feature vectors of each spectrogram sample. Among them, the feature vector of the spectrogram sample includes the feature values of multiple wavelength points in the spectrogram sample. The server converts the feature vectors of multiple spectrogram samples into a vector matrix. Each row in the vector matrix represents the feature vector of a spectrogram sample, and each column (i.e., the feature column) in the vector matrix includes the values of the wavelength points at the same position in multiple feature vectors. That is, the number of multiple spectrogram samples in the initial training set is N, and the number of wavelength points in a spectrogram sample is d. Then the number of feature values of the wavelength points in the feature vector of a spectrogram sample is d, and the feature vectors of multiple spectrogram samples form an N*d vector matrix.

[0078] After obtaining the vector matrix formed by the feature vectors of multiple historical data samples, the server fills in the missing values in this vector matrix to obtain a first vector matrix. This first vector matrix includes the feature vectors of multiple historical data samples after missing value filling. Due to data acquisition errors, one or more data points in the collected historical data samples may be missing, resulting in some feature values missing in the extracted feature vectors. Therefore, it is necessary to fill in the missing values in the vector matrix formed by multiple feature vectors to obtain a first vector matrix for subsequent data calculations, and then obtain accurate training samples.

[0079] For example, the historical data sample can be a spectrogram sample. In the spectrogram sample, there may be cases where the actual values of some wavelength points are not collected, that is, the measured values of some wavelength points in the spectrogram sample may be 0 or missing. After the server extracts the feature vectors of each spectrogram sample and converts them into a vector matrix, it can perform an operation to fill in the missing values for the eigenvalues in the vector matrix. That is, at the positions where the eigenvalues are 0 or missing in each row and each column of the vector matrix, a preset eigenvalue is filled. The preset eigenvalue can be a calibrated fixed value, or the mean or median of the eigenvalues in the column where the position is located, so as to obtain the first vector matrix after filling in the missing values.

[0080] By filling in the missing values in the feature vector matrix, the common missing problems in the actual collected data are solved, ensuring that the dimensions of all training samples are complete and effectively preventing training errors or model instability caused by missing values.

[0081] S22: Perform feature augmentation processing on the first vector matrix to obtain a second vector matrix, and perform normalization processing on the second vector matrix to obtain a normalized vector matrix.

[0082] After obtaining the first vector matrix, the server performs feature augmentation processing on the first vector matrix to obtain a second vector matrix. Through feature augmentation processing (such as constructing combined features, time window features, etc.), more potential correlation patterns can be discovered, improving the expression ability of the data dimension and helping the model learn more complex rules.

[0083] Among them, for the feature vector of each historical data sample in the first vector matrix, the server can calculate the target statistical vector of the feature vector of the historical data sample (the target statistical vector can be at least one of skewness, kurtosis, mean, and standard deviation of the feature vector), that is, calculate the target statistical vector (such as mean and standard deviation) of all eigenvalues (such as eigenvalues of all wavelength points) in each feature vector; then, the server takes the mean and standard deviation of the feature vector of the historical data sample as eigenvalues respectively and splices them to any position (such as the end or a random position) in the original feature vector. That is, the server calculates the mean and standard deviation of each row of data in the first vector matrix, and then inserts the mean and standard deviation of this row of data into any position of this row of data to form a new row of data (that is, form a new feature vector of the historical data sample), so as to obtain the second vector matrix after feature augmentation. The second vector matrix includes the feature vectors of multiple historical data samples, and the feature vector of each historical data sample includes the original eigenvalues (and the preset eigenvalues after filling in the missing values) extracted from the spectrogram sample, and the calculated target statistical vector.

[0084] After obtaining the second vector matrix, the server normalizes the second vector matrix to obtain a normalized vector matrix. Specifically, after obtaining the second vector matrix, the server calculates the mean and standard deviation of each column of data in the second vector matrix; then, for each column in the second vector matrix, the server subtracts the mean of the column data from each eigenvalue in the column data and divides it by the standard deviation of the column data to obtain the normalized column data, thereby obtaining a normalized vector matrix in which each column of data is normalized.

[0085] S23: Perform dimensionality reduction and data augmentation on the normalized vector matrix to obtain a target vector matrix, and output each eigenvector in the target vector matrix as a training sample to obtain a first sample dataset.

[0086] After obtaining the normalized vector matrix, the server performs dimensionality reduction and data augmentation on the normalized vector matrix to obtain a target vector matrix, and outputs each eigenvector in the target vector matrix as a training sample to obtain a first sample dataset. By performing dimensionality reduction to remove redundant and weakly correlated features, the representational compactness of the data is improved, the risk of model overfitting is reduced, and the training speed is accelerated; through data augmentation processing (such as adding noise, perturbation, time translation, etc.), more diverse training samples can be generated. Among them, after obtaining the normalized vector matrix, the server can adopt a preset dimensionality reduction algorithm, which can be a Recursive Feature Elimination (RFE) algorithm, a Principal Component Analysis (PCA) algorithm, etc., to perform dimensionality reduction on the normalized vector matrix to obtain a dimensionality-reduced normalized vector matrix. Then, the server adopts a preset data augmentation algorithm (such as an image data augmentation algorithm, a text data augmentation algorithm) to perform data augmentation on each row of data in the dimensionality-reduced normalized vector matrix, and outputs the augmented data of each row (that is, the eigenvector of each historical data sample) as a training sample to obtain a first sample dataset including multiple training samples.

[0087] In this embodiment, by performing steps such as feature extraction → missing value filling → feature expansion → normalization → dimensionality reduction → data augmentation on historical data samples, multiple training samples are obtained, the diversity and accuracy of the training samples are improved, more accurate and robust input data is provided for subsequent model training, the problems such as underfitting that are prone to occur due to the small number of samples in small-sample machine learning are solved, and the generalization ability and stability of the model are significantly improved in scenarios where the number of samples is insufficient.

[0088] In one embodiment, in step S4, that is, using the first sample data set and at least one target sample data set to train multiple initial models to obtain multiple prediction models, specifically including the following steps: S41: Use the first sample data set to train multiple initial models respectively to obtain multiple first prediction models with converged training.

[0089] The server obtains multiple initial models with different structures, uses the first sample data set to train multiple initial models respectively to obtain multiple initial models with updated parameters, then uses the test set to perform performance tests on the initial models with updated parameters, and after determining that their performance meets the requirements, outputs the initial model with updated parameters as the first prediction model with converged training to obtain heterogeneous first prediction models.

[0090] S42: Use the target sample data set to train multiple initial models respectively to obtain multiple second prediction models with converged training.

[0091] Meanwhile, after obtaining at least one target sample data set, the server uses one target sample data set to train multiple initial models respectively to obtain multiple initial models with updated parameters, and then uses the test set to perform performance tests on the initial models with updated parameters respectively. After determining that their performance meets the requirements, outputs the initial model with updated parameters as the second prediction model with converged training to obtain multiple heterogeneous second prediction models. Wherein, when the number of target sample data sets is K (K is an integer greater than 1), for each target sample data set, the same initial model is trained to obtain K isomorphic second prediction models. S43: Output multiple first prediction models and second prediction models to obtain multiple prediction models.

[0092] After training to obtain multiple first prediction models and multiple second prediction models, the server jointly outputs multiple first prediction models and multiple second prediction models as multiple prediction models. The multiple prediction models include multiple heterogeneous prediction models and multiple isomorphic prediction models.

[0093] For example, the server obtains N initial models with different structures, where the N initial models include at least two models among a support vector regression model and a random forest model, and multiple neural network models with different network structures; constructs K target sample data sets from multiple training samples in the first sample data set. The server trains the N initial models respectively using the first sample data set to obtain N first prediction models with converged training; after constructing K target sample data sets from multiple training samples in the first sample data set, the server trains the N initial models respectively using the target sample data sets to obtain N*K second prediction models with converged training; then, the N first prediction models with converged training and the N*K second prediction models are jointly output as prediction models to obtain N*(1 + K) prediction models, so as to subsequently perform target prediction on the data to be predicted using these N*(1 + K) prediction models.

[0094] According to the experimental results, it can be seen that using the data prediction method provided in this embodiment for data prediction, its performance is also significantly better than that of a single prediction model obtained by separate training in the traditional method. As Figure 6 and Figure 7 shown, it shows the prediction performance curves of using multiple prediction models provided in this embodiment and using traditional prediction models for data prediction when the training sample sizes are 20, 30, 40, 50, and 60 respectively. Among them, Figure 6 the curves 1, 2, and 3 in Figure 6 respectively represent the mean squared error curves of the prediction results obtained by using the random forest model and the extreme gradient boosting model (i.e., the XGBoost model) trained in the traditional way for prediction, and the mean squared error curve of the prediction result obtained by predicting with multiple prediction models trained in this embodiment and weighting the predicted values. Figure 6 The abscissa in Figure 7 represents the training sample size, Figure 6 and the ordinate in Figure 6 represents the mean squared error of the prediction result. Among them,

[0095] It should be understood that the smaller the mean squared error of the model on multiple training samples, the better the model performance; the larger the coefficient of determination of the model on multiple training samples, the better the model performance. From Figure 6 and Figure 7It can be seen that the performance of the prediction model obtained by training in this embodiment is significantly better than that of the model obtained by separately training with the traditional solution; moreover, as the number of training samples increases, the prediction model obtained by training in this embodiment exhibits better performance.

[0096] In this embodiment, a first sample data set is used to train multiple initial models respectively to obtain multiple first prediction models with converged training, and a target sample data set is used to train multiple initial models respectively to obtain multiple second prediction models with converged training. Finally, the multiple first prediction models and the second prediction models are output as multiple prediction models. By separately training multiple initial models with different structures on the first sample data set and the target sample data set, the models can learn the rules at different levels or dimensions respectively during the separate training, which can reduce the possible training bias or overfitting when training all models based on the same training data, and can generate multiple prediction models that are different in terms of data perspective and model structure, thereby enhancing the performance coverage of different prediction models, and further enhancing the generalization ability and robustness of multiple prediction models under different data distributions and improving the prediction accuracy.

[0097] In one embodiment, the server can perform weight optimization solution on the multiple prediction models obtained by training by using a second sample data set to obtain at least one weight data group, and the weight data group includes the weight values of multiple prediction models, so that after using the multiple prediction models to perform target prediction on the data to be predicted, based on the weight data group obtained by pre-solving, the current weight coefficients of the multiple prediction models are determined, and then the model prediction values output by each prediction model are weighted and summed based on the corresponding weight coefficients to obtain the target prediction value and output it as the prediction result.

[0098] Among them, as Figure 8 shown, after step S4, that is, after training multiple prediction models, the weight values corresponding to the multiple prediction models in the weight data group are obtained through the following steps: S5: Obtain a second sample data set, and the second sample data set includes multiple historical data samples.

[0099] After the server trains multiple prediction models, it obtains a second sample data set, and the second sample data set includes multiple historical data samples. Among them, the second sample data set is obtained by dividing the obtained multiple historical data samples.

[0100] S6: For each historical data sample, use multiple prediction models to perform target prediction on the historical data sample respectively, and determine the prediction vector corresponding to the historical data sample based on the sample prediction values output by the multiple prediction models respectively.

[0101] After obtaining a second sample data set including multiple historical data samples, for each historical data sample in the second sample data set, the server uses multiple prediction models to respectively perform target prediction on the historical data sample, and obtains sample prediction values respectively output by the multiple prediction models for the historical data sample.

[0102] That is, the historical data sample is respectively input into multiple prediction models for target prediction to obtain sample prediction values respectively output by the multiple prediction models; then the server determines a prediction vector corresponding to the historical data sample based on the sample prediction values respectively output by the multiple prediction models for the historical data sample, so as to obtain prediction vectors corresponding to multiple historical data samples. For example, the server can splice the sample prediction values respectively output by the multiple prediction models for the historical data sample into a vector to obtain the prediction vector corresponding to the historical data sample, and the prediction vector is composed of the sample prediction values of the multiple prediction models for the historical data sample.

[0103] S7: Optimally solve the weight values of multiple prediction models based on the prediction vectors corresponding to multiple historical data samples to obtain the weight values corresponding to each prediction model, so as to obtain a weight data set.

[0104] After obtaining the prediction vector corresponding to the historical data sample, the server optimally solves the weight values of multiple prediction models based on the prediction vectors corresponding to multiple historical data samples, with the goal of optimizing (such as maximizing or minimizing) a preset index and with the constraint that the sum of the weight values is 1, and uses a mathematical programming method to obtain the weight values corresponding to each prediction model, so as to obtain a weight data set. Each historical data sample corresponds to a true measured value.

[0105] Among them, the server can convert the prediction vectors corresponding to multiple historical data samples into a prediction vector matrix. Each row of the prediction vector matrix is the prediction vector corresponding to a historical data sample, and the prediction vector includes the sample prediction values of multiple prediction models for the historical data sample. The server can also set an initial weight value for each prediction model and splice the weight values of multiple prediction models into a weight matrix, where each row of the weight matrix is a weight value. Then, the server multiplies the prediction vector matrix by the weight matrix to obtain a target matrix, where each row of the target matrix is the target prediction value of a historical data sample, that is, the target matrix includes the target prediction values of multiple historical data samples. Among them, the target prediction value is obtained by weighted summing the sample prediction values of each prediction model in the prediction vector of the historical data sample using multiple weight values in the weight matrix. After obtaining the target matrix, that is, obtaining the target prediction values of multiple historical data samples, the server determines a preset metric based on the target prediction values and the true measurement values of multiple historical data samples. The preset metric is represented by an optimization objective function composed of the target prediction values and the true measurement values of multiple historical data samples. The server takes optimizing the preset metric as the goal and the sum of the weight values being 1 as the constraint, and iteratively updates the weight values of each prediction model to perform multi-objective optimization on the weight values of multiple prediction models to obtain the weight values of multiple prediction models, so as to form a weight data group.

[0106] Among them, the prediction vector matrix obtained by converting the prediction vectors corresponding to multiple historical data samples can be represented by the following matrix: = ; Among them, 、 … respectively represent the prediction vector of the first historical data sample, the prediction vector of the second historical data sample... the prediction vector of the th historical data sample, is the number of historical data samples in the second sample dataset; represents the th model's sample prediction value for the th historical data sample, , ; , ; is the number of prediction models.

[0107] Among them, the weight matrix obtained by splicing the weight values of multiple prediction models can be represented by the following matrix: ; Among them, 、 … respectively represent the weight values of the first prediction model, the weight values of the second prediction model... the weight values of the th prediction model,

[0108] Then, the target matrix can be expressed by the following formula: ; where 、 … respectively represent the target prediction values of the first historical data sample in the target matrix, the target prediction values of the second historical data sample... the target prediction values of the 、 … respectively represent the prediction vectors of the first historical data sample, the prediction vectors of the second historical data sample... the prediction vectors of the 、 … respectively represent the weight values of the first prediction model, the weight values of the second prediction model... the weight values of the th prediction model, is the number of prediction models;

[0109] For example, the preset index can be the coefficient of determination. After determining the optimization objective function of the coefficient of determination based on the target prediction values of multiple historical data samples in the target matrix, the server takes maximizing the coefficient of determination as the goal and the sum of the weight values being 1 as the constraint, and uses the mathematical programming method to optimize and solve the weight values of multiple prediction models to obtain the weight values corresponding to each prediction model.

[0110] Among them, the objective function of the coefficient of determination can be expressed by the following formula: ; where represents the coefficient of determination; represents the th target prediction value of the historical data sample; represents the th true measured value of the historical data sample, , …; is the number of historical data samples in the second sample dataset; represents the average value of the true measured values of all historical data samples in the second sample dataset.

[0111] In this embodiment, by obtaining the prediction outputs of each prediction model on the second sample data set and constructing a prediction vector, the performance differences of each prediction model on the actual data can be comprehensively evaluated, so that the weights of each prediction model can be optimally allocated based on the true effect, the accuracy of the weights of each prediction model can be improved, and thus the prediction accuracy after weight fusion can be significantly improved. In addition, the weight optimization process considers the overall prediction errors of multiple samples, which helps to suppress the influence of a certain model overfitting to specific samples, thereby improving the performance stability of the overall model combination on different input data and the accuracy of the model weights.

[0112] In one embodiment, a weight data group corresponds to a type of sample prediction value, that is, a weight data group corresponds to a prediction data type. The prediction targets for the server to perform target prediction on the historical data samples can include multiple types. The prediction target can be a prediction target of the dimension prediction type, such as the key dimension and thickness of the target object; the prediction target can also be a prediction target of the probability prediction type, such as the probability that the target object has a defect. The server can, according to requirements, perform predictions on the historical data samples with different prediction targets to obtain the sample prediction values of different types of prediction targets, so as to obtain the prediction vectors of different types of prediction targets corresponding to the historical data samples, so as to subsequently optimize and solve the weight values of multiple prediction models according to the prediction vectors of different types of prediction targets corresponding to multiple historical data samples, that is, obtain multiple weight data groups corresponding to different prediction data types, and one type of prediction target (that is, one prediction data type) corresponds to one weight data group. The acquisition process of each weight data group is as described above and will not be elaborated here.

[0113] In one embodiment, as Figure 9 shown, in step S7, that is, based on the prediction vectors corresponding to multiple historical data samples, the weight values of multiple prediction models are optimally solved to obtain the weight values corresponding to each prediction model, which specifically includes the following steps: S71: Determine the optimization index according to the prediction data type to which the sample prediction value output by the prediction model belongs and the prediction vectors corresponding to multiple historical data samples.

[0114] The server determines the prediction data type to which the sample prediction value output by the prediction model belongs. Among them, if the sample prediction value output by the prediction model is a numerical value of the dimension type, such as the sample prediction value becomes a critical dimension, thickness, etc., it is determined that the prediction data type to which the sample prediction value output by the prediction model belongs is the dimension prediction type, that is, the first type; if the sample prediction value output by the prediction model is a numerical value of the probability type, such as the sample prediction value becomes the probability of the current defect of the object, etc., it is determined that the prediction data type to which the sample prediction value output by the prediction model belongs is the probability prediction type, that is, the second type; in addition, if the sample prediction value output by the prediction model is a numerical value other than the dimension type or the probability type, it is determined that the prediction data type to which the sample prediction value output by the prediction model belongs is the other prediction type, that is, the third type.

[0115] After determining the prediction data type to which the sample prediction value output by the prediction model belongs, the server determines the optimization index according to the prediction data type to which the sample prediction value output by the prediction model belongs, and the prediction vectors corresponding to multiple historical data samples. Among them, the optimization indexes corresponding to different prediction data types are different. The optimization index can be the determination coefficient, mean square error, cross entropy loss, etc.

[0116] S72: Taking the maximized or minimized optimization index as the goal, based on the prediction vectors corresponding to multiple historical data samples, perform an optimization solution on the weight values of multiple prediction models to obtain the weight values corresponding to each prediction model.

[0117] After determining the optimization index, the server takes the maximized or minimized optimization index as the goal, and based on the prediction vectors corresponding to multiple historical data samples, performs an optimization solution on the weight values of multiple prediction models to obtain the weight values corresponding to each prediction model. Among them, the determination process of the weight values corresponding to each prediction model can be referred to above and will not be elaborated here.

[0118] In this embodiment, the weight is solved with the goal of maximizing or minimizing a certain type of optimization index to ensure that the overall performance of the fusion result on historical samples is optimal, and to avoid the local sub-optimal or non-optimal problems caused by artificially setting weights. Moreover, since different types of prediction tasks apply different evaluation criteria, if a certain index is uniformly used for optimization, it may lead to incompatibility with some tasks. In this embodiment, by selecting a suitable optimization index according to the prediction data type to which the sample prediction value belongs, the evaluation criteria of different task scenarios can be dynamically adapted, and the prediction effect of the fusion model under the target task can be improved.

[0119] In one embodiment, in step S71, that is, according to the prediction data type to which the sample prediction value output by the prediction model belongs, and the prediction vectors corresponding to multiple historical data samples, determining the optimization index specifically includes the following steps: S711: When the prediction data type to which the sample prediction value output by the prediction model belongs is the first type, determine that the optimization metric is the coefficient of determination of the target prediction values of multiple historical data samples.

[0120] After determining the prediction data type to which the sample prediction value output by the prediction model belongs, when the prediction data type to which the sample prediction value output by the prediction model belongs is the first type (such as the dimension prediction type), the server determines that the optimization metric is the coefficient of determination of the target prediction values of multiple historical data samples.

[0121] Among them, the target prediction value of each historical data sample is determined according to the prediction vector of the corresponding historical data sample and the weight value of the model. The determination process of the target prediction value of each historical data sample can be referred to the above description and will not be elaborated here.

[0122] S712: When the prediction data type to which the sample prediction value output by the prediction model belongs is the second type, determine that the optimization metric is the mean square error of the target prediction values of multiple historical data samples.

[0123] After determining the prediction data type to which the sample prediction value output by the prediction model belongs, when the prediction data type to which the sample prediction value output by the prediction model belongs is the second type (such as the probability prediction type), the server determines that the optimization metric is the mean square error of the target prediction values of multiple historical data samples. In other embodiments, when the prediction data type to which the sample prediction value output by the prediction model belongs is the third type (a type other than the probability prediction type and the dimension prediction type), the server may determine the cross-entropy loss of the target prediction values of multiple historical data samples as the optimization metric.

[0124] In this embodiment, the corresponding optimization metric is selected according to the prediction data type to which the sample prediction value belongs. The coefficient of determination is used for the first type of prediction data, which can measure the explanatory ability of the prediction model for data variation. The mean square error is used for the second type of prediction data, which can measure the overall deviation between the predicted value and the true value. The targeted optimization of the multi-model fusion weights is realized, thereby significantly improving the accuracy of the model weights and the adaptability to different types of prediction targets.

[0125] In one embodiment, in step S72, that is, with the maximized or minimized optimization metric as the target, based on the prediction vectors corresponding to multiple historical data samples, the weight values of multiple prediction models are optimized to obtain the weight values corresponding to each prediction model, which specifically includes the following steps: S721: When the optimization metric is the coefficient of determination of the target prediction values of multiple historical data samples, with the maximization of the coefficient of determination as the target, based on the prediction vectors corresponding to multiple historical data samples, the weight values of multiple prediction models are optimized to obtain the weight values corresponding to each prediction model.

[0126] Among them, when the prediction data type to which the sample prediction value output by the prediction model belongs is the first type (such as the dimension prediction type), that is, when the optimization metric is the coefficient of determination of the target prediction values of multiple historical data samples, the server aims to maximize the coefficient of determination and is constrained by the sum of the weight values being equal to 1. Based on the prediction vectors corresponding to multiple historical data samples, the server performs an optimization solution for the weight values of multiple prediction models to obtain the weight values corresponding to each prediction model, so as to form a weight data group. Among them, this weight data group is the weight data group corresponding to the first type (such as the probability prediction type). The solution process of the weight values of each prediction model can be referred to in the previous text and will not be elaborated here.

[0127] S722: When the optimization metric is the mean square error of the target prediction values of multiple historical data samples, aiming to minimize the mean square error, based on the prediction vectors corresponding to multiple historical data samples, the server performs an optimization solution for the weight values of multiple prediction models to obtain the weight values corresponding to each prediction model.

[0128] When the prediction data type to which the sample prediction value output by the prediction model belongs is the second type (such as the probability prediction type), that is, when the optimization metric is the mean square error of the target prediction values of multiple historical data samples, the server can determine the target matrix according to the above text, and this target matrix includes the target prediction values of multiple historical data samples. After obtaining the target matrix, that is, after obtaining the target prediction values of multiple historical data samples, the server calculates the mean square error between the target prediction values of multiple historical data samples and the true measured values, and uses this mean square error as the optimization metric. The server aims to minimize the mean square error and is constrained by the sum of the weight values being equal to 1, and performs a multi-objective optimization solution for the weight values of multiple prediction models to obtain the weight values of multiple prediction models, so as to form a weight data group. Among them, this weight data group is the weight data group corresponding to the second type (such as the probability prediction type).

[0129] In this embodiment, when the prediction data type to which the sample prediction value output by the prediction model belongs is the first type (such as the size prediction type), the coefficient of determination is used as the optimization index. The goal is to maximize the explanatory ability of the prediction model for data variation, improve the adaptability of the model weights to the size prediction target (i.e., the prediction target of the first type), and thus the prediction ability of the prediction model for the size prediction target. When the prediction data type to which the sample prediction value output by the prediction model belongs is the second type (such as the probability prediction type), the mean square error is used to measure the error fluctuation. The goal is to minimize the prediction error of the prediction model, improve the adaptability of the model weights to the probability prediction target (i.e., the prediction target of the first type), and thus the prediction ability of the prediction model for the probability prediction target. This solution selects the optimization goal as needed, and then determines the weight data group corresponding to different prediction data types, which can improve the adaptability of the system to different types of prediction tasks, improve the scientificity and accuracy of the model weights. Subsequently, different weight data groups can be selected according to actual prediction needs to optimize the prediction results, improve the accuracy of the prediction results, and make the prediction results more in line with the requirements of the target scenario.

[0130] It should be understood that the magnitudes of the sequence numbers of the steps in the above embodiments do not imply the order of execution. The execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0131] In one embodiment, a data prediction device is provided, which corresponds to the data prediction method in the above embodiment one by one. As Figure 10 shown, the data prediction device includes an acquisition module 101, a determination module 102, a prediction module 103, a calculation module 104, and an output module 105. The detailed description of each functional module is as follows: The acquisition module 101 is configured to acquire a plurality of pre-trained prediction models, and the plurality of prediction models are obtained through machine learning training based on a first sample data set; The determination module 102 is configured to determine the weight coefficients of the plurality of prediction models after receiving a prediction instruction for the data to be predicted, and the weight coefficients of the plurality of prediction models are weight values obtained by performing weight optimization on the plurality of prediction models based on a second sample data set; The prediction module 103 is configured to respectively perform target prediction on the data to be predicted by using the plurality of prediction models, and obtain model prediction values respectively output by the plurality of prediction models; The calculation module 104 is configured to perform weighted summation calculation on the model prediction values respectively output by the plurality of prediction models based on the weight coefficients of each prediction model to obtain a target prediction value; The output module 105 is configured to output the target prediction value as the prediction result corresponding to the data to be predicted, so as to process the data to be predicted based on the target prediction value.

[0132] In one embodiment, the determining module 102 is specifically configured to: determine the type of prediction data for performing target prediction on the data to be predicted; obtain a plurality of pre-solved weight data groups, where each weight data group includes weight values corresponding to a plurality of prediction models, and the plurality of weight values in each weight data group are obtained by performing weight optimization on the plurality of prediction models based on the second sample data set; determine the target weight data group corresponding to the type of prediction data from the plurality of weight data groups, and determine the corresponding weight values in the target weight data group as the weight coefficients of each prediction model.

[0133] In one embodiment, the determining module 102 is further configured to: obtain a second sample data set, where the second sample data set includes a plurality of historical data samples; for each historical data sample, use the plurality of prediction models to perform target prediction on the historical data sample respectively, and determine the prediction vector corresponding to the historical data sample based on the sample prediction values respectively output by the plurality of prediction models; perform optimization on the weight values of the plurality of prediction models based on the prediction vectors corresponding to the plurality of historical data samples to obtain the weight values corresponding to each prediction model, so as to obtain a weight data group, and one weight data group corresponds to a type of sample prediction value.

[0134] In one embodiment, the determining module 102 is further configured to: determine an optimization index according to the type of prediction data to which the sample prediction value output by the prediction model belongs, and the prediction vectors corresponding to the plurality of historical data samples; take the maximized or minimized optimization index as the target, and perform optimization on the weight values of the plurality of prediction models based on the prediction vectors corresponding to the plurality of historical data samples to obtain the weight values corresponding to each prediction model.

[0135] In one embodiment, the data prediction device further includes a training module, and the training module is configured to: obtain a plurality of historical data samples, where the historical data samples are of the same type as the data to be predicted; perform feature augmentation processing on the plurality of historical data samples to obtain a first sample data set including a plurality of training samples; perform data set construction on the plurality of training samples in the first sample data set to obtain at least one target sample data set; use the first sample data set and the at least one target sample data set to perform model training on the plurality of initial models to obtain a plurality of prediction models, and the plurality of initial models are machine learning models with different structures.

[0136] In one embodiment, the training module is specifically configured to: extract features from multiple historical data samples to obtain feature vectors of the multiple historical data samples, and fill missing values in the vector matrix formed by the multiple feature vectors to obtain a first vector matrix; perform feature expansion processing on the first vector matrix to obtain a second vector matrix, and perform normalization processing on the second vector matrix to obtain a normalized vector matrix; perform dimensionality reduction and data augmentation processing on the normalized vector matrix to obtain a target vector matrix, and output each feature vector in the target vector matrix as a training sample to obtain a first sample dataset.

[0137] In one embodiment, the training module is further specifically configured to: respectively train multiple initial models using the first sample dataset to obtain multiple first prediction models that converge in training; respectively train multiple initial models using the target sample dataset to obtain multiple second prediction models that converge in training; output the multiple first prediction models and the second prediction models to obtain multiple prediction models.

[0138] It should be noted that for the information interaction, execution process, etc. between the above-mentioned devices / units, since they are based on the same concept as the method embodiments of the present application, their specific functions and the technical effects brought can be specifically referred to in the method embodiment part, and will not be elaborated here.

[0139] Those skilled in the art can clearly understand that for the convenience and brevity of description, only the above-mentioned division of each functional unit and module is used as an example. In actual applications, the above-mentioned functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit. In addition, the specific names of the functional units and modules are for the convenience of mutual distinction and do not limit the protection scope of the present application. The specific working process of the units and modules in the above system can refer to the corresponding process in the foregoing method embodiments and will not be elaborated here.

[0140] The embodiment of the present application further provides a computer device, which can be a terminal device or a server, such as Figure 11 shown, the computer device includes: at least one processor, a memory, and a computer program stored in the memory and executable on at least one processor. When the processor executes the computer program, it implements the steps in any of the above method embodiments, or when the processor executes the computer program, it implements the functions of each module / unit in the above device embodiments.

[0141] Exemplarily, a computer program may be divided into one or more modules / units. One or more modules / units are stored in a memory and executed by a processor to complete the present application. One or more modules / units may be a series of computer program instruction segments capable of performing specific functions, and these instruction segments are used to describe the execution process of the computer program in a computer device. Those skilled in the art can understand that Figure 11 merely examples of computer devices, which do not constitute limitations on computer devices, may include more or fewer components than shown in the figure, or combine certain components, or different components. For example, a computer device may also include input / output devices, network access devices, buses, etc.

[0142] The above-mentioned processor may be a central processing unit, or may also be other general-purpose processors, digital signal processors, application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The memory may be an internal storage unit of the computer device, such as the hard disk or memory of the computer device. The memory may also be an external storage device of the computer device, such as a plug-in hard disk, smart memory card, secure digital card, flash card, etc. The memory may also include both the internal storage unit and the external storage device of the computer device.

[0143] The embodiments of the present application also provide a readable storage medium storing a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned various method embodiments can be implemented.

[0144] The embodiments of the present application provide a computer program product, and when the computer program product runs on an electronic device, the electronic device can implement the steps in the above-mentioned various method embodiments when executed.

[0145] When an integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, to implement all or part of the processes in the above-described embodiment methods of this application, a computer program can be used to instruct relevant hardware to complete. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-described method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can at least include: any entity or device that can carry the computer program code to the photographing device / terminal device, recording medium, computer memory, read-only memory, random access memory, electrical carrier signal, telecommunication signal, and software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk, or an optical disc, etc. In some jurisdictions, according to legislation and patent practice, the computer-readable medium cannot be an electrical carrier signal and a telecommunication signal.

[0146] In the above embodiments, the descriptions of each embodiment have their own focuses. For the parts not detailed or recorded in a certain embodiment, reference can be made to the relevant descriptions of other embodiments. Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed in this application can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0147] In the embodiments provided in this application, it should be understood that the disclosed device / equipment and method can be implemented in other ways. For example, the device / equipment embodiments described above are only illustrative. For example, the division of modules or units is only a logical function division. In actual implementation, there can be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection to each other can be through some interfaces. The indirect coupling or communication connection of the device or unit can be in an electrical, mechanical, or other form. The unit described as a separated component can be or can not be physically separated, and the component displayed as a unit can be or can not be a physical unit, that is, it can be located in one place, or can be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0148] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A data prediction method, characterized in that, Including: Obtain multiple pre-trained prediction models, where the multiple prediction models are obtained through machine learning training based on a first sample data set, and at least two of the multiple prediction models have different structures; After receiving a prediction instruction for the data to be predicted, determine the weight coefficients of the multiple prediction models, where the weight coefficients of the multiple prediction models are weight values obtained by performing weight optimization on the multiple prediction models based on a second sample data set; Use the multiple prediction models to respectively perform target prediction on the data to be predicted, and obtain model prediction values respectively output by the multiple prediction models; Based on the weight coefficients of each prediction model, perform weighted summation calculation on the model prediction values respectively output by the multiple prediction models to obtain a target prediction value; Output the target prediction value as the prediction result corresponding to the data to be predicted.

2. The data prediction method according to claim 1, wherein The multiple prediction models are trained through the following method: Obtain multiple historical data samples, where the historical data samples are of the same type as the data to be predicted; Perform feature augmentation processing on the multiple historical data samples to obtain the first sample data set including multiple training samples; Perform data set construction on the multiple training samples in the first sample data set to obtain at least one target sample data set; Use the first sample data set and at least one of the target sample data sets to perform model training on multiple initial models to obtain the multiple prediction models, where the multiple initial models are machine learning models with different structures.

3. The data prediction method according to claim 2, wherein The step of using the first sample data set and at least one of the target sample data sets to perform model training on multiple initial models to obtain the multiple prediction models includes: Use the first sample data set to respectively train the multiple initial models to obtain multiple first prediction models with training convergence; Use the target sample data set to respectively train the multiple initial models to obtain multiple second prediction models with training convergence; Output the multiple first prediction models and the second prediction models to obtain the multiple prediction models.

4. The data prediction method according to claim 2, wherein The step of performing feature augmentation processing on the multiple historical data samples to obtain the first sample data set including multiple training samples includes: Extract features from the multiple historical data samples to obtain feature vectors of the multiple historical data samples, and fill in missing values for the vector matrix formed by the multiple feature vectors to obtain a first vector matrix; Perform feature augmentation processing on the first vector matrix to obtain a second vector matrix, and perform normalization processing on the second vector matrix to obtain a normalized vector matrix; Perform dimensionality reduction and data augmentation processing on the normalized vector matrix to obtain a target vector matrix, and output each feature vector in the target vector matrix as a training sample to obtain the first sample data set.

5. The data prediction method according to any one of claims 1-4, characterized in that The step of determining the weight coefficients of the multiple prediction models includes: Determine the prediction data type for performing target prediction on the data to be predicted; Obtain multiple groups of weight data obtained by pre-solving. Each group of weight data includes the weight values corresponding to multiple of the prediction models. The multiple weight values in each group of weight data are obtained by performing weight optimization solution on multiple of the prediction models based on the second sample data set; Determine the target group of weight data corresponding to the prediction data type among the multiple groups of weight data, and determine the corresponding weight values in the target group of weight data as the weight coefficients of each of the prediction models.

6. The data prediction method according to claim 5, wherein The weight values corresponding to multiple of the prediction models in the group of weight data are determined by the following method: Obtain the second sample data set, which includes multiple historical data samples; For each of the historical data samples, use multiple of the prediction models to respectively perform target prediction on the historical data sample, and determine the prediction vector corresponding to the historical data sample based on the sample prediction values respectively output by multiple of the prediction models; Based on the prediction vectors corresponding to multiple of the historical data samples, perform optimization solution on the weight values of multiple of the prediction models to obtain the weight values corresponding to each of the prediction models, so as to obtain a group of weight data.

7. A data prediction device, characterized in that, Including: An acquisition module, configured to acquire multiple pre-trained prediction models. The multiple prediction models are obtained by performing machine learning training based on a first sample data set. At least two of the multiple prediction models have different structures; A determination module, configured to determine the weight coefficients of multiple of the prediction models after receiving a prediction instruction for data to be predicted. The weight coefficients of multiple of the prediction models are weight values obtained by performing weight optimization solution on multiple of the prediction models based on the second sample data set; A prediction module, configured to use multiple of the prediction models to respectively perform target prediction on the data to be predicted to obtain the model prediction values respectively output by multiple of the prediction models; A calculation module, configured to perform weighted summation calculation on the model prediction values respectively output by multiple of the prediction models based on the weight coefficients of each of the prediction models to obtain a target prediction value; An output module, configured to output the target prediction value as the prediction result corresponding to the data to be predicted.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, the steps of the data prediction method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, the steps of the data prediction method according to any one of claims 1 to 6 are implemented.

10. A computer program product, characterized in that, When the computer program is run, the data prediction method according to any one of claims 1 to 6 is executed.

Images