Data modeling method based on prototype adapter

Abstract

The data modeling method based on the prototype adapter comprises the following steps: acquiring multi-source and multi-type data, and constructing a data set; evaluating the quality of sample points in the data set to obtain a sample point quality evaluation result, and splitting the data set into a training set and a test set; according to the training set data, utilizing the data fitting model to construct a preliminary adapter model; and correcting the obtained preliminary adapter model to obtain a final data fitting type universal adapter model. The method sequentially increases the sample data according to the maximum mean square error criterion, achieves the updating of the original agent model, improves the utilization rate of the training sample data and the modeling precision of the agent model, and solves a problem that a conventional model machine adapter modeling method needs a large amount of data for training the model.

Description

Technical Field

[0001] The invention relates to a data modeling method based on a prototype adapter, and belongs to the field of data modeling. Background Art

[0002] The development of equipment and subsequent stages of its life cycle are restricted by objective factors such as conditions and costs, and often only limited physical tests are allowed. However, numerical simulation has the advantages of repeatability, economy, and safety. With the development of computer technology, numerical simulation has been continuously optimized in terms of the fidelity of mechanism modeling and the accuracy of calculation results. By combining virtual and real, and using virtual to predict real, it has made up for the problem of limited physical test samples to a certain extent, and has occupied an increasingly important position in the development of equipment systems and even training, maintenance and other life cycle activities. In view of the characteristics of aerospace equipment with heavy tasks, short cycles, and high quality requirements, it is necessary to integrate models and data-driven new generation information technology, integrate various resources such as virtual simulation and physical tests, and strengthen virtual design and simulation verification applications to avoid major technical adjustments in the later stages of development, thereby shortening the development cycle of weapons and equipment.

[0003] In recent years, with the improvement of computer hardware capabilities, artificial intelligence and deep learning technologies based on neural networks and their applications have also shown an explosive development trend. In the fields of natural language processing, computer vision, intelligent decision-making, etc., a large number of revolutionary applications such as GPT-4, Stable Diffusion, autonomous driving, medical diagnosis, AlphaGo, climate prediction, and drug development have emerged, fully verifying its learning ability and application prospects for complex problems. In the field of engineering, neural networks have also been successfully applied as agent models in virtual experiments, optimization design, uncertainty analysis, control decision-making, state monitoring, fault diagnosis, life prediction, etc., effectively verifying its ability to fit nonlinear, time-delayed, and high-dimensional data compared to general statistical models and other machine learning models, as well as its flexible adaptability when facing multi-source heterogeneous data and modeling objects.

[0004] The integration solution based on model adapter in virtual-real integration relies on agent model, which can replace, reduce or automate the corresponding activities and processes in the real world, such as experiments, which require a lot of time and human resources, by integrating all available data in virtual space and conducting rapid simulation and analysis. Similarly, the construction process of agent model itself also needs to follow such principles. Although the data-driven agent modeling method is generally a basically automated pipeline operation, that is, after a set of training data is input into the computer program, the corresponding model will be obtained, and after a set of new data is input into this model, the corresponding response will be predicted. However, the fitting effect of agent modeling depends on the choice of hyperparameters when facing different data. However, the work of selecting algorithms, configuring hyperparameters, and organizing data sets that can determine the fitting effect of agent models is still mainly completed by experts in the fields of statistics and machine learning through subjective judgment and manual parameter adjustment, which not only wastes a lot of manpower, but also raises the application threshold of technology, which is not conducive to the staff in other fields such as aerospace to maximize the potential of algorithms. Summary of the invention

[0005] The technical problem solved by the present invention is: to overcome the shortcomings of the prior art, propose a data modeling method based on a prototype adapter, sequentially increase sample data according to the maximum mean square error criterion, realize the update of the original proxy model, improve the utilization rate of training sample data and the modeling accuracy of the proxy model, and solve the problem that the traditional prototype adapter modeling method requires a large amount of data for training the model.

[0006] The technical solution of the present invention is: a data modeling method based on a prototype adapter, comprising:

[0007] Acquire data from multiple sources and types to build data sets;

[0008] Evaluate the quality of sample points in the data set, obtain the sample point quality evaluation results, and split the data set into a training set and a test set;

[0009] Based on the training set data, a preliminary adapter model is constructed using the data fitting model;

[0010] The obtained preliminary adapter model was modified to obtain the final data-fitting general adapter model.

[0011] A comprehensive method of centroid mosaic method and gradient distribution method is adopted to evaluate the quality of sample points in the dataset. That is, the sample point coverage information, sample point radius information and gradient information of the test domain where the sample points are located are obtained respectively through the centroid mosaic method and gradient distribution method. Finally, the evaluation results of the centroid mosaic method and gradient distribution method are further analyzed by the Kullback-Leibler divergence method.

[0012] The comprehensive method of using the centroid mosaic method and the gradient distribution method to evaluate the quality of sample points in the data set includes:

[0013] If the sample points in the data set are evenly distributed, the centroid mosaic method is used to evaluate the quality of the sample points and calculate the sample point coverage and coverage radius;

[0014] If the data set presents a gradient distribution, the gradient distribution method is used to evaluate the quality of the sample points; the gradient modulus and KL divergence of the sample points are calculated;

[0015] Finally, the sample point quality assessment results are obtained.

[0016] The KL divergence is an asymmetric measure of the difference between two probability distributions, and simultaneously characterizes the probability distribution p and the approximate distribution q. The KL divergence is used to obtain the information lost when approximating one distribution to another. When the two distributions are exactly the same, the KL divergence is 0.

[0017] The data fitting model includes a Gaussian process regression model, a response surface model or a Kriging model.

[0018] The obtained preliminary adapter model is modified to obtain a final data fitting general adapter model, including:

[0019] 1) The fitted universal adapter model, i.e., the preliminary training model, is used as the update object, and a model update method based on sequential sampling is used to obtain a sequential sampling point set;

[0020] 2) The obtained point set is used as a supplement to the training set and added to the new round of training and modeling together with the previous training set;

[0021] 3) Building methods based on general adapters for data fitting, retraining and modeling;

[0022] 4) Determine whether the current model meets the stopping criteria. If so, execute step 5), otherwise return to step 1) and use the current model as the update object;

[0023] 5) The revised model is regarded as a qualified model.

[0024] The model updating method based on sequential sampling includes a maximum variance sequential method and a maximum gradient sequential method.

[0025] The maximum variance sequential method is as follows: by finding the maximum variance of the remaining unused training set sample points, arranging them from large to small, and taking the points with the largest variance as sequential sampling points.

[0026] The maximum gradient sequential method is as follows: find the point with the largest gradient modulus in the test set as the sequential sampling point.

[0027] The stopping criterion is the accuracy of the corrected fitting model. When the error between the calculation result of the corrected fitting model and the true value is less than a certain threshold, it is determined to meet the stopping criterion.

[0028] The advantages of the present invention compared with the prior art are:

[0029] The present invention realizes the updating of the original proxy model through the proxy model correction method based on sequential sampling, can apply sample data not used for training to the model updating, and improves the modeling accuracy of the proxy model.

[0030] The present invention realizes the construction of a higher-precision proxy model based on simulation data and limited experimental data. The data fitting general adapter supports the selection of three data fitting mathematical models, namely, the response surface model, the Gaussian process regression model, and the Kriging model, to construct a proxy model; at the same time, it supports the application of sample data not used for training to the model correction and update through the model correction and update method based on sequential sampling, thereby improving the credibility of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Build and update method routes for the data fitting class generic adapter;

[0032] Figure 2 The sample point quality assessment method route used in the prototype adapter construction process;

[0033] Figure 3 Correction of methodological routes for data fitting of generic adapter-like models. DETAILED DESCRIPTION

[0034] The present invention proposes a data modeling method based on a prototype adapter, such as Figure 1 As shown, the following steps are included:

[0035] ST101, aggregate data from multiple sources and multiple types to construct a data set. The data in the data set are data of the same object, multiple sources refer to experimental test data or simulation data, and multiple types refer to data of different parameter types.

[0036] ST102, evaluate the quality of sample points.

[0037] The sample point quality assessment method adopted by the present invention combines the centroid mosaic method and the gradient distribution method, and further analyzes the assessment results of the centroid mosaic method and the gradient distribution method by the Kullback-Leibler divergence method. The sample point coverage information, the sample point radius information and the gradient information of the test domain where the sample point is located are obtained by the centroid mosaic method and the gradient distribution method respectively;

[0038] ST103, split the dataset into training set and test set.

[0039] ST104, select a data fitting method, select one of the Gaussian process regression model, response surface model, and Kriging model as the data fitting model, and use its construction method as the data fitting method of the proxy model.

[0040] The present invention provides three data fitting mathematical models for selection, including response surface model, Gaussian random regression model and Kriging model. The most suitable model can be selected according to the fitting effect for fitting the same type of data in the later stage.

[0041] The response surface model is a proxy model that replaces the mechanism model by fitting the complex implicit relationship between structural characteristics and parameters with an explicit response surface function. When selecting the form of the response surface function, both accuracy and efficiency should be considered. This requires that the form of the response surface function can simulate the actual functional relationship with high accuracy, and at the same time, its form should be as simple as possible to improve the calculation efficiency. Commonly used response surface function forms include polynomial response surface, radial basis function response surface, and support vector machine response surface. Taking the quadratic polynomial response surface as an example, the calculation steps are:

[0042] (1) Assume that the initial iteration point is It is usually taken as the mean point;

[0043] (2) Select the experimental design method to carry out the experimental design, and set the parameters x = (x1, x2, ..., x n ) is divided into (k-1) groups according to different levels, and the variable of the mth group is

[0044] (3) Calculation and

[0045]

[0046] in, is the response value of the response surface model, x i (i=1,2,...,n) is the design parameter value, A=[a,b i ,c i ] is the unknown coefficient, n is the number of design parameters, ε is the error between the calculated value and the true value of the response surface, usually, the unknown coefficient A is calculated by the least squares method;

[0047] (4) Using the least squares method to solve the coefficient matrix A of the response surface;

[0048] (5) Among them, the order of magnitude of the components of the parameter to be corrected X may differ greatly, such as the difference between the unit of size and the Young's modulus of more than ten orders of magnitude. When solving the response surface coefficients using the least squares method, a singular matrix may be obtained. Therefore, when calculating the parameter to be corrected x i Perform normalization processing;

[0049]

[0050] In the formula, is the structural parameter x in the original design space i The corresponding normalization parameter is easy to know

[0051] Gaussian Process Regression (GPR) is a non-parametric model that uses Gaussian Process (GP) priors to perform regression analysis on data. The Gaussian Process Regression model can be fitted by following the steps below.

[0052] Let the random vector X = [x1, x2, ..., x n ] T It obeys the multivariate Gaussian distribution X~N(μ,Σ), where: X1=[x1,x2,...,x n ] T is the observed variable, X2=[x1,x2,...,x n ] T is an unobserved variable, then

[0053]

[0054] Thus there is So we can find the posterior distribution of X2 given X1 as

[0055] μ 2|1 =μ2+Σ 21 Σ 11 -1 (X1-μ1);

[0056] Σ 21 =Σ 22 -Σ 21 Σ -1 11 Σ 12

[0057] A sample set in The purpose is to find y i With x i Functional relationship between them. Model building: Assume that y i With x iThe functional relationship of is given by the function f and is interfered by Gaussian noise, that is, for any i,

[0058] y i =f(x i )+ξ i

[0059] Among them, ξ i is generated by a Gaussian random variable ξ.

[0060] X=[x1,...,x i ,...,x n ] T ,y=[y1,...,y i ,...,y n ] T

[0061] Use least squares regression to find the optimal solution for the following parameters

[0062]

[0063] Assume v is a Gaussian process with mean function 0 and variance function κ, and ξ is independent of f, with mean 0 and variance σ 2 A Gaussian random variable, that is

[0064]

[0065] The goal is to calculate the set of test points given the sample set D The function value at Based on the assumption that f is a Gaussian process, calculate y and f * The joint probability distribution of

[0066]

[0067] Where I is the identity matrix,

[0068] Therefore (y,f * ) has a mean of 0 and a covariance matrix of The joint Gaussian distribution of . The conditional probability density p = (f * |X * ,X,y)

[0069] p(f * |X * ,X,y)=N(f * |μ * ,Σ * ),

[0070] μ * =K* '(K+σ 2 I) -1 y,

[0071] Σ * =K ** -K * '(K+σ 2 I) -1 K *

[0072] Get f * The conditional probability density p=(f * |X * ,X,y), we can sample according to the density and get f * The sample value of is the predicted value.

[0073] The Kriging model is an interpolation model whose interpolation result is defined as the linear weighting of the known sample function response value, that is,

[0074]

[0075] Therefore, as long as the weighting coefficient ω=[ω (1) ,ω (2) ,...,ω (n) ] T The performance estimate of any solution in the design space can be obtained by using the expression of . In order to calculate the weighting coefficient, the Kriging model introduces a statistical assumption: the unknown function is regarded as a specific embodiment of a Gaussian static random process. The static random process is defined as

[0076] Y(x)=β0+Z(x)

[0077] Where β0 is an unknown constant, also called the global trend model, representing the mathematical expectation of Y(x); Z(·) has a mean of 0 and a variance of σ 2 A static random process. At different locations in the design space, these random variables have a certain correlation, and the covariance can be expressed as

[0078] Cov[Z(x),Z(x')]=σ 2 R(x,x')

[0079] Where R(x,x') is the correlation function, and satisfies R(x,x') = 1 when the distance is zero; R(x,x') = 0 when the distance is infinite; the correlation decreases as the distance increases.

[0080] Based on the above assumptions, the Kriging model seeks the optimal weighting coefficient ω so that the mean square error

[0081]

[0082] Minimum and meets the interpolation conditions

[0083]

[0084] Using the Lagrange multiplier method, it can be deduced that the optimal weighting coefficient ω is given by the following Kriging model equations:

[0085]

[0086] Where i = 1, 2, ..., n; μ is the Lagrange multiplier. Written in matrix form

[0087]

[0088] Solving equation (4.27), we can get the Kriging model estimate

[0089]

[0090] By inverting the block matrix, the model can be finally written as

[0091]

[0092] In the formula, β0=(F T R -1 F) -1 F T R -1 y s , column vector V = R -1 (y s -β0F), which is only related to known sample points. Predicting the x value at any position requires calculating the dot product between r(x) and V. In addition, the Kriging model gives an estimate of the mean square error of the estimate as

[0093]

[0094] The mean square error can be used to guide the addition of new sample points to improve the accuracy of the proxy model or approach the optimal solution to the optimization problem.

[0095] The training set was fitted by the response surface model, Gaussian random regression model and Kriging model respectively to obtain three proxy models. The model with the highest fitting accuracy was selected from the three proxy models as the general adapter for data fitting, which was used for subsequent testing and calculation of the same type of data. The general adapter module for data fitting of virtual-reality integrated test supports the selection, construction, modification and application of data fitting models, and provides efficient and practical data fitting test models for the design of virtual-reality integrated test schemes.

[0096] ST105, construct an adapter model based on the training set data through the selected data fitting method.

[0097] ST106, the adapter model obtained in ST105 is modified by a model updating method based on sequential sampling, including maximum gradient sequential and maximum mean square error sequential.

[0098] The adapters of the preliminary modeling need to be evaluated. The modeling effect can be judged by evaluating the maximum mean square error in the test set. For adapters whose maximum mean square error does not meet the stopping criteria, additional sample points need to be supplemented for remodeling. The process of screening and supplementing sample points is called sequential sampling, and the point selection criterion is called sequential sampling criterion. The study of sequential sampling based on maximum mean square error and gradient distribution obtains sample points with maximum variance and gradient close to 90° and 0° for further model correction.

[0099] ST107, obtain the data fitting class universal adapter model.

[0100] The sample point quality assessment method used to assess the sample point quality is as follows: Figure 2 As shown, the following steps are included:

[0101] ST201, obtain the original test or simulation data.

[0102] ST202: If the sample points are evenly distributed, the centroid mosaic method is used to evaluate the quality of the sample points.

[0103] ST203, calculate the sample point coverage and coverage radius.

[0104] ST204: If the sample points present a gradient distribution, use the gradient distribution method to evaluate the quality of the sample points.

[0105] ST205, calculate the gradient modulus of the sample point.

[0106] ST206, calculate the KL divergence of the sample points.

[0107] KL divergence is an asymmetric measure of the difference between two probability distributions, and it also characterizes the probability distribution p and the approximate distribution q.

[0108]

[0109] The KL divergence can be used to find out how much information is lost when approximating one distribution to another. When the two distributions are exactly the same, the KL divergence is 0.

[0110] For various data fitting general adapters, the sample point quality assessment results are obtained by studying the sample point quality assessment method before modeling. The assessment results are suitable for modeling sampling and subsequent sequential update adapters.

[0111] ST207, obtain the sample point quality assessment results.

[0112] The centroid mosaic method evaluates the quality of sample points by comparing the coverage of the test domain by the sample points. After obtaining the sample points for training, they are recorded as a set of n sample points in m dimensions.

[0113]

[0114] Find the distance R between each pair of sample points and arrange them in order from small to large. Half of the shortest distance is the required radius r.

[0115]

[0116] In the formula

[0117]

[0118] In the formula, i, j = 1, 2, ..., n, and the radius r is obtained after calculating the distance.

[0119] r=[r1,r2,...,r n ]

[0120] In the formula

[0121] r i =minR i·

[0122] The area of ​​the corresponding circle is calculated by the radius, and then the area ratio covered in the test domain is obtained, which is recorded as the coverage rate C

[0123]

[0124] The gradient distribution method evaluates the rate of change of sample response by comparing the gradient changes of sample points, so as to evaluate the quality of sample points. After obtaining the sample points for training, the distance information obtained in the centroid mosaic method is used to construct a local response surface for each sample point using the 5 points closest to it, and the gradient at the sample point is calculated.

[0125] For the gradient distribution method, the distance between each sample point is also used. The five points closest to each sample point are calculated using this distance. These five points are built into a second-order response surface, and the gradient information of the response surface at this point is calculated.

[0126]

[0127] mf=[mf1,mf2,...,mf n ]

[0128] Finally, the KL divergence result of the gradient modulus mf and the radius information r is calculated as the final sample point quality evaluation result.

[0129] KL divergence is an asymmetric measure of the difference between two probability distributions, and it also characterizes the probability distribution p and the approximate distribution q.

[0130]

[0131] The KL divergence can be used to find out how much information is lost when approximating one distribution to another. When the two distributions are exactly the same, the KL divergence is 0.

[0132] For various data fitting general adapters, the sample point quality assessment method is studied before modeling to obtain the sample point quality assessment results. The assessment results are suitable for modeling sampling and subsequent sequential update adapters. If the KL divergence between radius information and gradient information is too large, the sample point will lose a large amount of model information, and it is necessary to consider improving the sample point quality when modifying the model later.

[0133] Model updating methods based on sequential sampling are as follows Figure 3 As shown, the following steps are included:

[0134] ST301, the fitted universal adapter model, namely the preliminary training model, is used as an update object. Two optional sequential methods are provided, including the maximum variance sequential method and the maximum gradient sequential method.

[0135] ST302 is calculated by the maximum variance sequential method.

[0136] The maximum variance sequential sampling method finds the maximum variance of the remaining unused training set sample points, arranges them from large to small, and uses the points with the largest variance as sequential sampling points for subsequent model correction.

[0137] ST303, calculation is performed using the maximum gradient sequential method.

[0138] The maximum gradient sequential method is to find the point with the largest gradient modulus in the test set as the sequential sampling point, and add it to the training set for remodeling. The specific process is:

[0139] When determining the maximum gradient, the gradient modulus is converted into an angle. Its physical meaning is the steepness of the gradient change reflected on the response surface. When the gradient is close to 90 degrees, it means that the response of this segment changes dramatically; when the gradient is close to 0 degrees, it means that there may be an inflection point in this segment. The calculation formula is:

[0140] θ i =arctan(mf i )·180° / π

[0141] θ=[θ1,θ2,...,θ k ]

[0142] mθ=max(θ)

[0143] The sample point corresponding to mθ is the point with the largest gradient modulus.

[0144] ST304, after completing the calculation according to the selected sequential method, a sequential sampling point set is obtained. The newly obtained sequential sampling point set is used as a supplement to the training set and is added to a new round of training and modeling together with the previous training set.

[0145] ST305, constructs methods based on data fitting class general adapters, retraining and modeling.

[0146] ST306, determine whether the current model meets the stopping criterion. If so, execute ST307, otherwise return to ST301 and use the current model as the update object. The stopping criterion can be specified as the accuracy of the corrected fitting model. When the error between the calculation result of the corrected fitting model and the true value is less than a certain degree, it is determined to meet the stopping criterion.

[0147] ST307, the corrected model is regarded as a qualified model.

[0148] Although the present invention has been disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solution of the present invention by using the technical contents disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope of the technical solution of the present invention.

Claims

1. A data modeling method based on a prototype adapter, characterized in that: include: Acquire data from multiple sources and types to build data sets; Evaluate the quality of sample points in the data set, obtain the sample point quality evaluation results, and split the data set into a training set and a test set; Based on the training set data, a preliminary adapter model is constructed using the data fitting model; The obtained preliminary adapter model was modified to obtain the final data-fitting general adapter model.

2. A data modeling method based on a prototype adapter according to claim 1, characterized in that: A comprehensive method of centroid mosaic method and gradient distribution method is adopted to evaluate the quality of sample points in the dataset. That is, the sample point coverage information, sample point radius information and gradient information of the test domain where the sample points are located are obtained respectively through the centroid mosaic method and gradient distribution method. Finally, the evaluation results of the centroid mosaic method and gradient distribution method are further analyzed by the Kullback-Leibler divergence method.

3. A data modeling method based on a prototype adapter according to claim 2, characterized in that: The comprehensive method of using the centroid mosaic method and the gradient distribution method to evaluate the quality of sample points in the data set includes: If the sample points in the data set are evenly distributed, the centroid mosaic method is used to evaluate the quality of the sample points and calculate the sample point coverage and coverage radius; If the data set presents a gradient distribution, the gradient distribution method is used to evaluate the quality of the sample points; the gradient modulus and KL divergence of the sample points are calculated; Finally, the sample point quality assessment results are obtained.

4. The data modeling method based on the prototype adapter according to claim 3 is characterized in that: The KL divergence is an asymmetric measure of the difference between two probability distributions, and simultaneously characterizes the probability distribution p and the approximate distribution q. The KL divergence is used to obtain the information lost when approximating one distribution to another. When the two distributions are exactly the same, the KL divergence is 0.

5. The data modeling method based on the prototype adapter according to claim 1, characterized in that: The data fitting model includes a Gaussian process regression model, a response surface model or a Kriging model.

6. The data modeling method based on the prototype adapter according to claim 1, characterized in that: The obtained preliminary adapter model is modified to obtain a final data fitting general adapter model, including: 1) The fitted universal adapter model, i.e., the preliminary training model, is used as the update object, and a model update method based on sequential sampling is used to obtain a sequential sampling point set; 2) The obtained point set is used as a supplement to the training set and added to the new round of training and modeling together with the previous training set; 3) Building methods based on general adapters for data fitting, retraining and modeling; 4) Determine whether the current model meets the stopping criteria. If so, execute step 5), otherwise return to step 1) and use the current model as the update object; 5) The revised model is regarded as a qualified model.

7. The data modeling method based on the prototype adapter according to claim 6 is characterized in that: The model updating method based on sequential sampling includes a maximum variance sequential method and a maximum gradient sequential method.

8. The data modeling method based on the prototype adapter according to claim 7 is characterized in that: The maximum variance sequential method is as follows: by finding the maximum variance of the remaining unused training set sample points, arranging them from large to small, and taking the points with the largest variance as sequential sampling points.

9. The data modeling method based on prototype adapter according to claim 7, characterized in that: The maximum gradient sequential method is as follows: find the point with the largest gradient modulus in the test set as the sequential sampling point.

10. The data modeling method based on prototype adapter according to claim 6, characterized in that: The stopping criterion is the accuracy of the corrected fitting model. When the error between the calculation result of the corrected fitting model and the true value is less than a certain threshold, it is determined to meet the stopping criterion.

Images