Adaptive multi-fidelity data fusion method based on multi-credibility model

By introducing an adaptive spatial mapping method of Pearson's correlation coefficient in the multi-confidence model, the spatial mapping matrix is ​​optimized, and the problem of ignoring phase deviation in the existing technology is solved, and the precise fusion and accuracy of high and low confidence data is achieved. predict.

CN118965271APending Publication Date: 2024-11-15XIDIAN UNIV
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411174721.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

When the existing multi-credibility model is fused with high and low-credibility data, it only considers the amplitude deviation and ignores the phase deviation, which leads to the limitations in practical applications and cannot be widely used in multi-credibility model settings.

Method used

Adaptive spatial mapping method based on Pearson's correlation coefficient is adopted to calculate the correlation between high and low confidence samples, optimize the spatial mapping matrix, dynamically adjust the weights, and improve the robustness and accuracy of the model.

Benefits of technology

Achieve adaptive data fusion that takes into account amplitude and phase information in a multi-confidence model, improves the accuracy and robustness of the model, and can better predict the corresponding values ​​of the high-confidence model, suitable for multi-confidence Degree model settings.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118965271A_ABST
    Figure CN118965271A_ABST
Patent Text Reader

Abstract

The invention discloses a self-adaptive multi-fidelity data fusion method based on a multi-credibility model. The method comprises the following steps of: (1) inputting high and low credibility analysis sample quantities into an algorithm model, and optimizing high and low credibility space mapping parameters and training parameters; (2) calculating high and low credibility analysis model response samples, obtaining corresponding model response values, and obtaining high and low credibility sample data sets; (3) calculating the correlation between the samples with high and low credibility; (4) defining a space mapping matrix, and generating a high-reliability sample space mapping loss function and a low-reliability sample space mapping loss function; (5) performing high and low credibility spatial mapping optimization to obtain a spatial mapping matrix; (6) projecting the high credibility to the low credibility space through the space mapping matrix; step (7), establishing a multi-credibility neural network; and step (8), outputting a multi-credibility prediction result. According to the invention, self-adaptive multi-credibility data fusion is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of data fusion, and in particular relates to an adaptive multi-fidelity data fusion method based on a multi-credibility model. Background Art

[0002] For some large electronic equipment, such as array antennas and reflector antennas, their functional surfaces should be supported and designed in detail to maintain the required surface shape during the detailed design stage. However, the analysis process with the overall antenna is extremely time-consuming, so the model needs to be simplified so that optimization can be used. However, the simplified model often deviates from the refined model, resulting in the accuracy of the results being affected. Therefore, in the process of simplifying the model, combining the simplified model with the refined model is the key to optimizing the design process and reducing the computational cost.

[0003] In order to fuse multi-fidelity data, Meng et al. proposed a method in the paper "A composite neural network that learns from multi-fidelity data: Application to function approximation and inverse PDE problems (Journal of Computational Physics, 2020, 401: 109020)" to divide the input of the final prediction model into the input value information of the current input point and the corresponding low-confidence corresponding value, and perform linear and nonlinear transformations on them. The above method only considers the deviation of the corresponding value and requires a high correlation between the high and low multi-confidence analysis models, so it shows limitations in the actual multi-confidence model.

[0004] Meng et al. pointed out in the paper "A fast multi-fidelity method with uncertainty quantification for complex data correlations: Application to vortex-induced vibrations of marine risers" (Computer Methods in Applied Mechanics and Engineering, 2021, 386: 114212) that the deviation between high and low credibility models comes not only from amplitude error, but also from phase deviation. In order to solve this problem, it is proposed to correct it in the modal space. Since correcting the phase deviation in the modal space usually requires a large number of samples, this method has shown limitations in practical applications and cannot be widely used in the multi-credibility model setting. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide an adaptive multi-fidelity data fusion method based on a multi-credibility model, which converts the spatial mapping objective function from the regularization function set by the original hyperparameter to a constraint using the Pearson correlation coefficient, thereby obtaining a spatial mapping matrix that is adaptively adjusted according to the sample and is conducive to multi-credibility model modeling, which can be used for subsequent optimization procedures. By introducing the Pearson correlation coefficient, the model can better predict the corresponding value of the high-credibility model and dynamically adjust the weight during the training process to improve the robustness and accuracy of the model.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] An adaptive multi-fidelity data fusion method based on a multi-credibility model comprises the following steps;

[0008] Step (1): input the number of high and low confidence analysis samples into the algorithm model, optimize the high and low confidence space mapping parameters, and the training parameters;

[0009] Step (2): Calculate high and low credibility analysis model response samples, initialize the analysis model, divide different credibility models according to the degree of conformity with the actual situation to obtain the corresponding high credibility analysis model and low credibility analysis model, and conduct experimental design to obtain the corresponding model response values ​​to obtain high and low credibility sample data sets;

[0010] Step (3): Calculate the correlation between high and low confidence samples;

[0011] Step (4): define the spatial mapping matrix M and generate the high and low confidence sample spatial mapping loss function;

[0012] Step (5): Optimize the high and low confidence spatial mapping to obtain the spatial mapping matrix M opt ;

[0013] Step (6): Through the spatial mapping matrix M opt , projecting high credibility to low credibility space;

[0014] Step (7): Use the transformed high and low confidence samples to establish a multi-confidence neural network;

[0015] Step (8): Output multi-credibility prediction results.

[0016] In the step (1), the number of samples of high and low credibility analysis models, the optimization rounds of the spatial mapping function, the number of optimized particles, the particle swarm optimization parameters, the neural network structure, the training rounds, the learning step length and the multi-credibility model training parameters are input; the high and low credibility analysis models are simplified models divided according to the actual situation, and the model that is close to the actual situation is considered to be a high credibility model, and the model with a greater degree of simplification is regarded as a low credibility analysis model;

[0017] The samples are divided into high confidence model input samples x h , output sample y h ; Low confidence model input sample x l , output sample y l .

[0018] The step (2) is specifically:

[0019] In the design space D, different credibility models F h , F l Sampling is performed, where F h For a high-confidence analysis model, F l For low confidence analysis model, F h With F l The number of samples is N h and N l , and satisfies N l >>N h , F h The sampling method is the optimal Latin hypercube design, F l Using optimal Latin hypercube design or full factorial design, x h is a high confidence input sample, x l is a low-confidence input sample, and F h With F l Inputting a sample will get the corresponding model value y h With y l The high-confidence sample set D is formed by the high-confidence and low-confidence samples obtained above. h =(x h ,F h ) and low confidence sample set D l =(x l ,F l );

[0020] x h is a high confidence input sample, x l is a low confidence input sample;

[0021] F h With F l By inputting the sample, we can get the corresponding model value y hWith y l .

[0022] The step (3) is specifically:

[0023] Calculate the Pearson correlation coefficient λ for the obtained high and low confidence samples:

[0024]

[0025] The subscript h represents a high-confidence sample, l represents a low-confidence sample, and the superscript i represents the sample number. and Represents the corresponding mean of high-confidence samples and the corresponding mean of low-confidence samples, y l When calculating the correlation coefficient, the corresponding value of the low-confidence sample closest to the high-confidence sample input value is found;

[0026] in The value of is determined by finding a value closest to x H The point is solved by optimization:

[0027]

[0028] in represents the i-th sample with low confidence, represents the jth sample with low confidence, is the maximum allowable error, is the square of the 2-norm.

[0029] The step (4) is specifically:

[0030] The spatial mapping is divided into the spatial mapping matrix M and the spatial deviation vector B. The spatial mapping is expressed as:

[0031] x HSM =Mx H +B

[0032] where x H is the high-confidence sample point sampled in step (2), x HSM is the sample point after spatial mapping;

[0033] Set x H If is n-dimensional, the spatial mapping matrix is ​​specifically expressed as:

[0034]

[0035] Among them, M i,j Represents the spatial transformation coefficient of the i-th sample in the j-th dimension;

[0036] Then, the high-low confidence sample space mapping loss function is defined according to the Pearson correlation coefficient in step (3):

[0037]

[0038] where y H (x) represents the high-confidence sample response value corresponding to the current high-confidence sample input value, y L,HSM (Mx+B) represents the projection of the high-confidence input sample to the response value of the low-confidence sample closest to the low-confidence sample space, M0 and B0 are the regularization matrix and regularization vector of the space mapping, ||·||2 represents the 2-norm of the matrix, and ||·||2 represents the 1-norm of the vector;

[0039] The function input is actually a particle optimized in step (5), and its optimization is expressed as:

[0040] find M,B

[0041] min Cost

[0042] sM BL ≤M≤M BH

[0043] B BL ≤B≤B BH

[0044] In the current particle, we first need to perform spatial transformation x on the high confidence samples HSM =Mx H +B, and then find the value corresponding to the new coordinate x by running the following optimization: l,near Closest point:

[0045] find x l

[0046]

[0047] stx l ∈D.

[0048] The step (5) is specifically as follows:

[0049] The optimal spatial mapping matrix M obtained by optimizing in step (5) opt , the optimal space mapping vector B opt Perform spatial coordinate transformation x HSM =M opt x H +B opt , specifically expressed as:

[0050]

[0051] Construct a high-confidence dataset D based on the obtained high-confidence new coordinates h =(x HSM ,F h ).

[0052] The step (6) is specifically as follows:

[0053] By using the optimized spatial mapping matrix M and vector B, the high-confidence sample is spatially transformed to obtain the transformed (x HSM ,f H ) High confidence sample dataset.

[0054] The step (7) is specifically:

[0055] The training set D is constructed by using the transformed high-confidence sample data and the original low-confidence sample data. train =(x l ,y l ,x h ,y h ), and then use the transfer learning neural network to build a multi-credibility model, first by using low-credibility samples (x l ,y l ) for training, where the loss function is expressed as follows:

[0056]

[0057] Among them, θ is the parameter that the neural network needs to learn. represents the i-th low-confidence sample input value, represents the output value of the i-th low-confidence sample, The neural network prediction output is specifically expressed as:

[0058]

[0059] Among them, W is the neural network parameter, ψ is the neural network activation function, is the operator, which updates the parameter W by gradient descent;

[0060] The model obtained by training only with low-confidence samples is then trained with high-confidence samples:

[0061]

[0062] A multi-credibility model is obtained through two learning steps.

[0063] The step (8) is specifically as follows:

[0064] When changing the input to predict the design point, first pass the space mapping matrix M optand vector B opt It is mapped and then predicted by a multi-credibility model, and its predicted value is output.

[0065] Beneficial effects of the present invention:

[0066] 1. When designing a multi-credibility model, the present invention not only considers the amplitude information and phase information between the high-credibility and low-credibility models, but also considers the correlation between the high-credibility and low-credibility models in the high-credibility and low-credibility samples. First, the correlation between the high-credibility samples and the low-credibility samples is calculated as the adaptive regularization term of the adaptive spatial mapping, and then the adaptive regularization term is introduced into the spatial mapping loss function, and the optimal spatial mapping matrix is ​​obtained by optimizing and solving this loss function. The optimal spatial mapping matrix is ​​used to map the high-credibility samples to the low-credibility sample space, and the multi-credibility model is used for modeling, thereby realizing adaptive multi-credibility data fusion.

[0067] 2. Traditional multi-credibility models have high requirements on the correlation between high and low credibility models, and the resulting multi-credibility model has high errors. Although the multi-credibility model combined with spatial mapping is capable of alleviating problems such as poor accuracy when the correlation between high and low credibility models is low, it cannot meet most multi-credibility cases due to the existence of hyperparameters, and the hyperparameters need to be obtained a posteriori. The present invention uses the Pearson correlation coefficient between high and low credibility samples to adaptively establish the regularization term in the loss function of spatial mapping. And use the mapped high-credibility samples to model the multi-credibility model. The resulting multi-credibility model results have high accuracy and can adaptively adjust the hyperparameters of the regularization term. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 A training flow chart of a multi-credibility model method based on adaptive multi-fidelity data fusion provided in an embodiment of the present invention.

[0069] Figure 2 A prediction flow chart of the multi-credibility model method based on adaptive multi-fidelity data fusion provided by the present invention.

[0070] Figure 3 The structure of the multi-credibility model method based on adaptive multi-fidelity data fusion provided by the present invention on the test function. DETAILED DESCRIPTION

[0071] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0072] See also Figure 1 , Figure 1 The training flow chart of the multi-credibility model method based on adaptive multi-fidelity data fusion provided for the implementation of the present invention includes:

[0073] Step 1, input the user-provided high and low confidence models (, sample number; spatial mapping optimization parameters such as the number of spatial mapping optimization particles and the number of spatial mapping optimization iterations; multi-credibility neural network parameters such as architecture, activation function, learning rate, optimizer, and learning iterations; taking the mesh antenna as an example, the high-credibility analysis model considers the truss deformation, and the low-credibility analysis model does not consider the truss deformation, where the design parameter is the cable length variable. Specific input parameter examples are: the high-credibility model considers the truss deformation, the low-credibility model does not consider the truss deformation, and the sample number is 1000; the number of spatial mapping optimization particles is 50, and the number of spatial mapping optimization iterations is 200; the multi-credibility neural network parameters include 3 layers of fully connected neural networks (64 neurons per layer), ReLU activation function, 0.001 learning rate, Adam optimizer, and 5000 learning iterations. Through these parameter settings, a multi-credibility neural network model can be constructed;

[0074] Step 2: Generate high and low confidence training sets D = (x l ,f l ,x h ,f h );

[0075] In the design space D, different credibility models F h , F l Sampling is performed, where F h For a high-confidence analysis model, F l For low confidence analysis model, F h With F l The number of samples is N h and N l , and satisfies N l >>N h , F h The sampling method is the optimal Latin hypercube design, F l Using optimal Latin hypercube design or full factorial design, x h is a high confidence input sample, x l is a low-confidence input sample, and F h With F l Inputting a sample will get the corresponding model value y h With y l The high-confidence sample set D is formed by the high-confidence and low-confidence samples obtained above. h =(x h ,F h ) and low confidence sample set D l =(x l ,F l );

[0076] x h is a high confidence input sample, x l is a low confidence input sample;

[0077] F h With F l By inputting the sample, we can get the corresponding model value y h With y l .

[0078] Step 3, in order to be able to adaptively adjust the regularization term in its spatial mapping loss function according to D = (x l ,f l ,x h ,f h ) Calculate the Pearson correlation coefficient between high and low confidence, and use the Pearson correlation coefficient as a hyperparameter of the regularization term in the spatial mapping loss function. The Pearson correlation coefficient between high and low confidence samples is calculated as;

[0079]

[0080] The subscript h represents a high-confidence sample, l represents a low-confidence sample, and the superscript i represents the sample number. and Represents the corresponding mean of high-confidence samples and the corresponding mean of low-confidence samples, y l When calculating the correlation coefficient, the corresponding value of the low-confidence sample closest to the high-confidence sample input value is found;

[0081] in The value of is determined by finding a value closest to x H The point is solved by optimization:

[0082]

[0083] in represents the i-th sample with low confidence, represents the jth sample with low confidence, is the maximum allowable error, is the square of the 2-norm.

[0084] in, is the high confidence sample mean, is the mean of low confidence samples. The regularization term in the spatial mapping loss function can be expressed as:

[0085] λ(||M-M0||2+||B-B0||1)

[0086] Step 4: In order to reduce the spatial deviation between high-confidence samples and low-confidence samples, the spatial mapping loss function is generated according to the regularization term obtained by the Pearson correlation coefficient as follows:

[0087]

[0088] Among them, y H (x) represents the high-confidence sample response value corresponding to the current high-confidence sample input value, y L,HSM (Mx+B) represents the projection of the high-confidence input sample to the response value of the low-confidence sample closest to the low-confidence sample space. λ(||M-M0||2+||B-B0||1) is the regularization term, λ is the regularization coefficient taken as the Pearson correlation coefficient; M0 and B0 are the regularization matrix and regularization vector of the space mapping, ||·||2 represents the 2-norm of the matrix, and ||·||1 represents the 1-norm of the vector.

[0089] Step 5: In order to reduce the impact of the spatial deviation between high and low confidence on the modeling of the multi-confidence model, it is necessary to improve the sample correlation between high and low confidence. First, use spatial mapping to project high-confidence samples into the low-confidence sample space.

[0090] The projection expression is:

[0091] x HSM =Mx H +B

[0092] Among them, x HSM Represents high-confidence samples after spatial projection transformation.

[0093] According to the obtained spatial mapping loss function, particle swarm is used for optimization, and the optimization is shown as follows:

[0094] find M,B

[0095] min Cost

[0096] sM BL ≤M≤M BH

[0097] B BL ≤B≤B BH

[0098] The current particle in the particle swarm optimization first needs to perform a spatial transformation x on the high-confidence sample HSM =Mx H +B, and then find the closest point to the new coordinates by running the following optimization:

[0099] find x l

[0100]

[0101] stx l ∈D

[0102] Get the optimal space mapping matrix M opt and the optimal offset vector B opt ;

[0103] Step 6: According to the optimal spatial mapping matrix and offset vector, the high-confidence samples are spatially mapped to obtain a new data set D = (x l ,f l ,x HSM ,f h )

[0104] Step 7: Initialize the multi-credibility neural network and use the dataset (x l ,f l ) Train the neural network;

[0105] The training set D is constructed by using the transformed high-confidence sample data and the original low-confidence sample data. train =(x l ,y l ,x h ,y h ). Then, a multi-credibility model is constructed using a transfer learning neural network. First, by using low-credibility samples (x l ,y l )1 for training, where the loss function is expressed as follows:

[0106]

[0107] Among them, θ is the parameter that the neural network needs to learn. represents the i-th low-confidence sample input value, Represents the output value of the i-th low-confidence sample;

[0108] The model obtained by training only with low-confidence samples is then trained with high-confidence samples:

[0109]

[0110] A multi-credibility model is obtained through two learning steps.

[0111] Step 8: Based on the obtained multi-credibility neural network model, use the data set (x HSM ,f h ) Train the neural network;

[0112] Step 9, outputting multi-credibility neural network parameters;

[0113] This method first inputs the number of high and low credibility samples, neural network structure parameters, and neural network training parameters; secondly, divides the high and low credibility models according to the computational cost, calculates the corresponding values ​​of high and low credibility samples, and calculates the Pearson correlation coefficient between the high and low credibility samples; then defines the spatial mapping loss function, and optimizes and solves the loss function to obtain the optimal spatial mapping matrix; then, the low credibility samples are learned by using the multi-credibility neural network model based on transfer learning, and then the high credibility samples are converted using the optimal spatial mapping matrix, and the multi-credibility model is continued to be learned using the data set; finally, the multi-credibility model is output. When predicting, the sample to be predicted is first converted by the above-mentioned optimal spatial mapping matrix, and then used as the input value of the multi-credibility model to obtain its prediction result.

[0114] See also Figure 2 , Figure 2 The prediction flow chart of the multi-credibility model method based on adaptive multi-fidelity data fusion provided for the implementation of the present invention includes:

[0115] Step 1, input high and low confidence sample data sets, neural network parameters, optimal space mapping matrix and optimal space offset vector;

[0116] Step 2, using the optimal spatial mapping matrix and the optimal spatial offset vector to perform spatial mapping conversion on the predicted samples;

[0117] Step 3, load the neural network parameters and input the new sample set obtained in step 2;

[0118] Step 4: Output the prediction results.

[0119] The advantages of the present invention can be further illustrated by the following simulation experiments:

[0120] Test function simulation conditions:

[0121] The test function is shown in Table 1, where 5 points are sampled for high confidence and 50 points for low confidence. This method is used to perform multi-confidence data fusion on this test.

[0122] Table 1 Test

[0123]

[0124] Simulation results:

[0125] The multi-credibility model method based on adaptive multi-fidelity data fusion of the present invention is adopted. Figure 3 ,from Figure 3It can be seen that although there is an obvious phase deviation between high credibility and low credibility, a more accurate high credibility can be predicted by combining adaptive spatial mapping with multi-credibility neural network.

[0126] The multi-credibility model in this implementation scheme can be used in the development and optimization process of electronic equipment such as radomes and can be used as a basic analysis model.

[0127] The parts not described in detail in this embodiment belong to the commonly used means known in the industry and are not described here one by one. The above examples are only for illustration of the present invention and do not constitute a limitation on the protection scope of the present invention.

[0128] Or similar designs all belong to the protection scope of the present invention.

Claims

1. An adaptive multi-fidelity data fusion method based on a multi-credibility model, characterized in that: The steps include: Step (1): input the number of high and low confidence analysis samples, optimize the high and low confidence space mapping parameters, and the training parameters; Step (2): Calculate high and low credibility analysis model response samples, initialize the analysis model, divide different credibility models according to the degree of conformity with the actual situation to obtain the corresponding high credibility analysis model and low credibility analysis model, and conduct experimental design to obtain the corresponding model response values ​​to obtain high and low credibility sample data sets; Step (3): Calculate the correlation between high and low confidence samples; Step (4): define the spatial mapping matrix M and generate high and low confidence sample spatial mapping loss functions; Step (5): Optimize the high and low confidence spatial mapping to obtain the optimized spatial mapping matrix M opt ; Step (6): Through the spatial mapping matrix M opt , projecting high credibility to low credibility space; Step (7): Use the transformed high and low confidence samples to establish a multi-confidence neural network; Step (8): Output multi-credibility prediction results.

2. The adaptive multi-fidelity data fusion method based on a multi-credibility model according to claim 1, characterized in that: In the step (1), the number of samples of high and low credibility analysis models, the optimization rounds of the spatial mapping function, the number of optimized particles, the particle swarm optimization parameters, the neural network structure, the training rounds, and the learning step length multi-credibility model training parameters are input; the high and low credibility analysis models are simplified models divided according to the actual situation, and the one that is close to the actual situation is considered to be a high credibility model, and the model with a greater degree of simplification is regarded as a low credibility analysis model; The samples are divided into high confidence model input samples x h , output sample y h ; Low confidence model input sample x l , output sample y l .

3. The adaptive multi-fidelity data fusion method based on a multi-credibility model according to claim 1, characterized in that: The step (2) is specifically: In the design space D, different credibility models F h , F l Sampling is performed, where F h For a high-confidence analysis model, F l For low confidence analysis model, F h With F l The number of samples is N h and N l , and satisfies N l >>N h , F h The sampling method is the optimal Latin hypercube design, F l Use an optimal Latin hypercube design or a full factorial design. h is a high confidence input sample, x l is a low-confidence input sample, and F h With F l Inputting a sample will get the corresponding model value y h With y l The high-confidence sample set D is formed by the high-confidence and low-confidence samples obtained above. h =(x h ,y h ) and low confidence sample set D l =(x l ,y l ); x h is a high confidence input sample, x l is a low confidence input sample; F h With F l By inputting the sample, we can get the corresponding model value y h With y l .

4. The adaptive multi-fidelity data fusion method based on a multi-credibility model according to claim 3 is characterized in that: The step (3) is specifically: Calculate the Pearson correlation coefficient λ for the obtained high and low confidence samples: The subscript h represents a high-confidence sample, l represents a low-confidence sample, and the superscript i represents the sample number. and Represents the corresponding mean of high-confidence samples and the corresponding mean of low-confidence samples, y l When calculating the correlation coefficient, the corresponding value of the low-confidence sample closest to the high-confidence sample input value is found; where y l The value of i is to find the one closest to x H The point is solved by optimization: in represents the i-th sample with low confidence, represents the jth sample with low confidence, is the maximum allowable error, is the square of the 2-norm.

5. The adaptive multi-fidelity data fusion method based on a multi-credibility model according to claim 4, characterized in that: The step (4) is specifically: The spatial mapping is divided into the spatial mapping matrix M and the spatial deviation vector B. The spatial mapping is expressed as: x HSM =Mx H +B where x H is the high-confidence sample point sampled in step (2), x HSM is the sample point after spatial mapping; Set x H If is d-dimensional, the spatial mapping matrix is ​​specifically expressed as: Among them, M i,j Represents the spatial transformation coefficient of the i-th sample in the j-th dimension; Then, the high-low confidence sample space mapping loss function is defined according to the Pearson correlation coefficient in step (3): where y H (x) represents the high-confidence sample response value corresponding to the current high-confidence sample input value, y L,HSM (Mx+B) represents the projection of the high-confidence input sample to the response value of the low-confidence sample closest to the low-confidence sample space, M0 and B0 are the regularization matrix and regularization vector of the space mapping, ||·||2 represents the 2-norm of the matrix, and ||·||1 represents the 1-norm of the vector; The function input is actually a particle optimized in step (5), and its optimization is expressed as: find M,B min Cost s.t.M BL ≤M≤M BH B BL ≤B≤B BH In the current particle, we first need to perform spatial transformation x on the high confidence samples HSM =Mx H +B, and then find the value corresponding to the new coordinate x by running the following optimization: l,near Closest point:

6. The adaptive multi-fidelity data fusion method based on a multi-credibility model according to claim 5, characterized in that: The step (5) is specifically as follows: The optimal spatial mapping matrix M obtained by optimizing in step (5) opt , the optimal space mapping vector B opt Perform spatial coordinate transformation x HSM =M opt x H +B opt , specifically expressed as: Construct a high-confidence dataset D based on the obtained high-confidence new coordinates h =(x HSM ,F h ).

7. The adaptive multi-fidelity data fusion method based on a multi-credibility model according to claim 6, characterized in that: The step (6) is specifically as follows: By using the optimized spatial mapping matrix M opt and vector B opt , perform spatial transformation on high-confidence samples and obtain the transformed (x HSM ,f H ) High confidence sample dataset.

8. The adaptive multi-fidelity data fusion method based on a multi-credibility model according to claim 7, characterized in that: The step (7) is specifically: The training set D is constructed by using the transformed high-confidence sample data and the original low-confidence sample data. train =(x l ,y l ,x h ,y h ), and then use the transfer learning neural network to build a multi-credibility model, first by using low-credibility samples (x l ,y l ) for training, where the loss function is expressed as follows: Among them, θ is the parameter that the neural network needs to learn. represents the i-th low-confidence sample input value, represents the output value of the i-th low-confidence sample, The neural network prediction output is specifically expressed as: Where W is the neural network parameter, ψ is the neural network activation function, and 'ο' is the operator, which updates the parameter W by gradient descent; The model obtained by training only with low-confidence samples is then trained with high-confidence samples: A multi-credibility model is obtained through two learning steps.

9. The adaptive multi-fidelity data fusion method based on a multi-credibility model according to claim 8, characterized in that: The step (8) is specifically as follows: When changing the input to predict the design point, first pass the space mapping matrix M opt and vector B opt It is mapped and then predicted by a multi-credibility model, and its predicted value is output.

Citation Information

Cited By

  • Multi-fidelity data fusion method, device, equipment, medium and product

    CN119849336A

  • A multi-fidelity data fusion method, device, equipment, medium and product

    CN119849336B