A modeling method for complex multi-source data fusion

Abstract

The present invention belongs to the technical field of intelligent fluid mechanics, and discloses a modeling method for complex multi-source data fusion. The modeling method for complex multi-source data fusion of the present invention includes defining fidelity according to data quality; constructing a front-end model; constructing a sub-model; and constructing an aggregation model. The modeling method for complex multi-source data fusion of the present invention can solve the complex multi-source data fusion modeling problem with three or more types of data sources, further improving the data quality; moreover, it has flexible scalability, significantly reduces the difficulty of complex multi-source data fusion problems, shortens the modeling time, and improves the modeling efficiency. The modeling method for complex multi-source data fusion of the present invention can be applied to the multi-source aerodynamic data fusion problem in the aerospace field, and can also be expanded to other fields with multi-source data fusion requirements.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent fluid mechanics, and in particular relates to a modeling method for complex multi-source data fusion. Background Art

[0002] At present, the main means of obtaining aerodynamic data include engineering calculations, CFD numerical simulations, wind tunnel tests, and flight tests. The quality of the data obtained by them increases successively, but the cost also increases sharply. In order to solve this contradictory relationship between data quality and acquisition cost, more and more researchers have begun to pay attention to aerodynamic data fusion modeling methods to achieve efficient and low-cost acquisition of aerodynamic data. Aerodynamic data fusion modeling is to fuse aerodynamic data from multiple different sources to obtain richer and more comprehensive information by comprehensively utilizing the complementary and redundant information of multi-source data in time or space.

[0003] Most of the current aerodynamic data modeling methods focus on the fusion of two types of data, which are difficult to apply to the fusion of complex multi-source data (data sources reach 3 or more). In fact, the sources of aerodynamic data are more than two types. By fusing data from more sources, the data quality can be further improved. Moreover, some models for complex multi-source data fusion are quite complex and lack flexible scalability, resulting in long modeling time and low modeling efficiency.

[0004] Currently, there is an urgent need to develop a modeling method for complex multi-source data fusion. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a modeling method for complex multi-source data fusion.

[0006] The main idea of ​​the modeling method for complex multi-source data fusion of the present invention can be summarized as "decomposition-aggregation".

[0007] The purpose of decomposition is to decompose the original problem into multiple sub-problems of the same nature. The decomposition principle is shown in Figure 1 , the decomposition process is shown in Figure 2 , Figure 1 and Figure 2 The specific process of decomposition is as follows:

[0008] First, sort data class 1, data class 2, ..., data class n according to data quality, and obtain data class x, ..., data class b, data class a from low to high data quality; secondly, perform dimensionality reduction decomposition, and merge relatively high-quality data classes with relatively low-quality data classes in pairs to obtain sub-problems 1, sub-problems 2, ..., sub-problems w, and generate sub-problems; finally, deep learning methods are used to model the sub-problems and obtain A sub-problem.

[0009] in, Figure 1 in " refers to constructing w pairwise fusion sub-models of the same nature for n types of data fusion problems. The same nature means that each sub-problem contains a relatively low-quality data class and a relatively high-quality data class; the goal of each sub-problem is the same, that is, to achieve fusion modeling of high- and low-quality data. Since the sub-models have the same properties, the same method can be used to construct all sub-models. Usually, deep learning methods are used to model the sub-problems.

[0010] The purpose of aggregation is to combine the results of multiple sub-problems to obtain the solution to the original problem. Figure 3 .

[0011] Combination Figure 1~Figure 3 The modeling method for complex multi-source data fusion of the present invention comprises the following steps:

[0012] S10. Define fidelity based on data quality;

[0013] For given engineering calculation data, CFD numerical simulation data and wind tunnel test data, the data are evaluated according to the authenticity of the physical properties reflected by the data. The closer to the reality, the higher the data quality and the higher the data fidelity. Engineering calculation data is defined as low-fidelity data, CFD numerical simulation data is defined as medium-fidelity data, and wind tunnel test data is defined as high-fidelity data.

[0014] S20. Build the front-end model;

[0015] like Figure 4 As shown, for low-fidelity data, a front-end model Low net Model is constructed from the low-fidelity incoming flow condition Xlow to the corresponding low-fidelity data Ylow; the front-end model Low net Model reflects the mapping relationship between the low-fidelity incoming flow condition Xlow and the low-fidelity data Ylow, and is a proxy model for low-fidelity data; the model structure of the front-end model Low net Model is a fully connected network;

[0016] S30. Construct sub-model;

[0017] like Figure 5As shown, the low-fidelity incoming flow condition Xlow and the low-fidelity data Ylow construct the sub-model Modellow2high and the sub-model Modellow2mid, and the medium-fidelity incoming flow condition Xmid and the medium-fidelity data Ymid construct the sub-model Modelmid2high; each sub-model uses the same construction method, and each sub-model structure includes two parts: a nonlinear network with a nonlinear activation function and a nonlinear network without a nonlinear activation function; the sub-model Modellow2high outputs high-fidelity data Yhigh1, the sub-model Modellow2mid outputs medium-fidelity data Ymid, and the sub-model Modelmid2high outputs high-fidelity data Yhigh2;

[0018] S40. Figure 3 As shown, build an aggregation model;

[0019] S41. Build branch;

[0020] Use all sub-models to build several different branches; the construction of branches follows the following rules: each branch is a separate sub-model or multiple sub-models connected end to end; the starting point of each branch, i.e. the input, is low-fidelity data Ylow; the end point of each branch, i.e. the output, is high-fidelity data Yhigh; for branches built by multiple sub-models, the sub-models are connected in the order of increasing data fidelity;

[0021] S42. Allocate aggregation weights;

[0022] Assign a weight parameter to each branch, which represents the contribution of each branch to the result; the weight parameter value is between 0 and 1, and the sum of the weight parameters of all branches is 1;

[0023] S43. Weighted aggregation, feedback update;

[0024] According to the weight parameters of each branch, the results of all branches are weighted and aggregated to obtain the predicted high-fidelity data Yhigh; then, the predicted high-fidelity data Yhigh is compared with the true value, and then fed back to S42 to adjust the weight parameters of each branch until the difference between the predicted high-fidelity data Yhigh and the true value reaches a preset error range.

[0025] The modeling method for complex multi-source data fusion of the present invention can solve the complex multi-source data fusion modeling problem with three or more types of data sources, further improving the data quality; moreover, it has flexible scalability, significantly reduces the difficulty of the complex multi-source data fusion problem, shortens the modeling time, and improves the modeling efficiency.

[0026] The modeling method for complex multi-source data fusion of the present invention can be applied to the multi-source aerodynamic data fusion problem in the aerospace field, and can also be expanded to other fields with multi-source data fusion requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Decompose schematics into subproblems;

[0028] Figure 2 Decompose the flow chart into sub-problems;

[0029] Figure 3 Aggregate flow charts for sub-problems;

[0030] Figure 4 Build schematics for front-end models;

[0031] Figure 5 Construct schematics for submodels;

[0032] Figure 6 Construct schematics for aggregation models;

[0033] Figure 7a C of airfoil sg6043 L ~α data curve (Reynolds number Re=5×10 4 );

[0034] Figure 7b C of airfoil sg6043 L ~α data curve (Reynolds number Re=10×10 4 );

[0035] Figure 7c C of airfoil sg6043 L ~α data curve (Reynolds number Re=15×10 4 );

[0036] Figure 7d C of airfoil sg6043 L ~α data curve (Reynolds number Re=20×10 4 );

[0037] Figure 7e C of airfoil sg6043 L ~α data curve (Reynolds number Re=30×10 4 );

[0038] Figure 7f C of airfoil sg6043 L ~α data curve (Reynolds number Re=40×10 4 );

[0039] Figure 7gC of airfoil sg6043 L ~α data curve (Reynolds number Re=50×10 4 );

[0040] Figure 7h C of airfoil sg6043 L ~α data curve (Reynolds number Re=100×10 4 );

[0041] Figure 8a C of airfoil sg6043 D ~α data curve (Reynolds number Re=5×10 4 );

[0042] Figure 8b C of airfoil sg6043 D ~α data curve (Reynolds number Re=10×10 4 );

[0043] Figure 8c C of airfoil sg6043 D ~α data curve (Reynolds number Re=15×10 4 );

[0044] Figure 8d C of airfoil sg6043 D ~α data curve (Reynolds number Re=20×10 4 );

[0045] Figure 8e C of airfoil sg6043 D ~α data curve (Reynolds number Re=30×10 4 );

[0046] Figure 8f C of airfoil sg6043 D ~α data curve (Reynolds number Re=40×10 4 );

[0047] Figure 8g C of airfoil sg6043 D ~α data curve (Reynolds number Re=50×10 4 );

[0048] Figure 8h C of airfoil sg6043 D ~α data curve (Reynolds number Re=100×10 4 );

[0049] Figure 9a-1 C of airfoil sg6043 L~α prediction curve (Reynolds number Re=10×10 4 );

[0050] Figure 9a-2 C of airfoil sg6043 L Relative error bar graph corresponding to the prediction curve (Reynolds number Re=10×10 4 );

[0051] Figure 9b-1 C of airfoil sg6043 L ~α prediction curve (Reynolds number Re=15×10 4 );

[0052] Figure 9b-2 C of airfoil sg6043 L Relative error bar graph corresponding to the prediction curve (Reynolds number Re=15×10 4 );

[0053] Figure 9c-1 C of airfoil sg6043 L ~α prediction curve (Reynolds number Re=20×10 4 );

[0054] Figure 9c-2 C of airfoil sg6043 L Relative error bar graph corresponding to the prediction curve (Reynolds number Re=20×10 4 );

[0055] Figure 9d-1 C of airfoil sg6043 L ~α prediction curve (Reynolds number Re=30×10 4 );

[0056] Figure 9d-2 C of airfoil sg6043 L Relative error bar graph corresponding to the prediction curve (Reynolds number Re=30×10 4 );

[0057] Figure 9e-1 C of airfoil sg6043 L ~α prediction curve (Reynolds number Re=40×10 4 );

[0058] Figure 9e-2 C of airfoil sg6043 L Relative error bar graph corresponding to the prediction curve (Reynolds number Re=40×10 4 );

[0059] Figure 9f-1C of airfoil sg6043 L ~α prediction curve (Reynolds number Re=50×10 4 );

[0060] Figure 9f-2 C of airfoil sg6043 L Relative error bar graph corresponding to the prediction curve (Reynolds number Re=50×10 4 );

[0061] Figure 10a-1 C of airfoil sg6043 D ~α prediction curve (Reynolds number Re=10×10 4 );

[0062] Figure 10a-2 C of airfoil sg6043 D Relative error bar graph corresponding to the prediction curve (Reynolds number Re=10×10 4 );

[0063] Figure 10b-1 C of airfoil sg6043 D ~α prediction curve (Reynolds number Re=15×10 4 );

[0064] Figure 10b-2 C of airfoil sg6043 D Relative error bar graph corresponding to the prediction curve (Reynolds number Re=15×10 4 );

[0065] Figure 10c-1 C of airfoil sg6043 D ~α prediction curve (Reynolds number Re=20×10 4 );

[0066] Figure 10c-2 C of airfoil sg6043 D Relative error bar graph corresponding to the prediction curve (Reynolds number Re=20×10 4 );

[0067] Figure 10d-1 C of airfoil sg6043 D ~α prediction curve (Reynolds number Re=30×10 4 );

[0068] Figure 10d-2 C of airfoil sg6043 D Relative error bar graph corresponding to the prediction curve (Reynolds number Re=30×10 4 );

[0069] Figure 10e-1 C of airfoil sg6043 D ~α prediction curve (Reynolds number Re=40×10 4 );

[0070] Figure 10e-2 C of airfoil sg6043 D Relative error bar graph corresponding to the prediction curve (Reynolds number Re=40×10 4 );

[0071] Figure 10f-1 C of airfoil sg6043 D ~α prediction curve (Reynolds number Re=50×10 4 );

[0072] Figure 10f-2 C of airfoil sg6043 D Relative error bar graph corresponding to the prediction curve (Reynolds number Re=50×10 4 ). DETAILED DESCRIPTION

[0073] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0074] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0075] Embodiment: This embodiment uses the SG6043 airfoil to verify the reliability of the modeling method for complex multi-source data fusion of the present invention.

[0076] This embodiment achieves the prediction of high-fidelity lift and drag coefficients based on incoming flow conditions by integrating three types of aerodynamic data with different fidelity. Among them, the high-fidelity data comes from 110 wind tunnel test data of "UIUC Low-Speed ​​Airfoil Test" (Volume 3); the medium-fidelity data and low-fidelity data come from the Xfoil simulation results under different Ncrit values ​​(critical roughnessReynolds number) in "airfoiltools.com". Ncrit is often used to describe the roughness effect considered in the calculation. Ncrit=9 represents the simulation of general wind tunnel conditions, and Ncrit=5 represents the simulation of poor wind tunnel conditions. Specifically, this embodiment uses the simulation results of Ncrit=9 as medium-fidelity data, and the simulation results of Ncrit=5 as low-fidelity data.

[0077] In this embodiment, each data is composed of the angle of attack α, the Reynolds number Re, the lift coefficient C L And the drag coefficient C D Among them, the angle of attack α and the Reynolds number Re are used as input conditions, and the lift coefficient C L and the drag coefficient C D All 409 medium-fidelity data and 402 low-fidelity data were used for training. Among the 110 high-fidelity data, 70% of the data at each Reynolds number were randomly sampled as training data, and the remaining 30% were used as test data. The specific original data distribution and high-fidelity data training and test sampling points can be found in Figure 7a~Figure 7h and Figure 8a~Figure 8h .in, Figure 7a~Figure 7h The C of airfoil sg6043 at different Reynolds numbers are given respectively. L ~α data curve chart, Figure 8a~Figure 8h The CD~α data curves of airfoil sg6043 at different Reynolds numbers are given respectively.

[0078] The specific implementation process is as follows:

[0079] S10. Define fidelity based on data quality;

[0080] 110 wind tunnel test data from UIUC Low-Speed ​​Airfoil Test (Volume 3) are used as high-fidelity data; the simulation results of Ncrit=9 in Xfoil are used as medium-fidelity data, and the simulation results of Ncrit=5 are used as low-fidelity data;

[0081] S20. Build the front-end model;

[0082] Multilayer Perceptron (MLP) is used as the front-end model Low net Model, referred to as x2low model, and all low-fidelity data are used for training; the input low-fidelity flow condition Xlow is , the output low-fidelity data Ylow is :

[0083] ;

[0084] S30. Construct sub-model;

[0085] The low-fidelity incoming flow condition Xlow and the low-fidelity data Ylow construct the sub-model Modellow2highh and the sub-model Modellow2mid, and the medium-fidelity incoming flow condition Xmid and the medium-fidelity data Ymid construct the sub-model Modelmid2high; each sub-model uses the same construction method, and each sub-model structure includes two parts: a nonlinear network with a nonlinear activation function and a nonlinear network without a nonlinear activation function; the sub-model Modellow2high outputs high-fidelity data Yhigh1, the sub-model Modellow2mid outputs medium-fidelity data Ymid, and the sub-model Modelmid2high outputs high-fidelity data Yhigh2;

[0086] Inputs for all sub-models , Output The description is as follows:

[0087] ;

[0088] in, , represents the prediction result of the mapping model under the corresponding incoming flow condition;

[0089] Use all the mid-fidelity data to build the sub-model Modellow2mid; Input into the x2low model to get the corresponding low-fidelity , Then, and , Concatenate into a one-dimensional vector as the input of the sub-model Modellow2mid , and then iterate the training until convergence;

[0090] Use high-fidelity data to build the sub-model Modellow2high, and the process is similar to the construction process of the sub-model Modellow2mid;

[0091] Then, all high-fidelity data are used to construct the sub-model Modelmid2high. First, the x2low model is used to obtain the corresponding incoming flow condition , The prediction results; then, the future flow conditions and , Input the trained sub-model Modellow2mid and get the corresponding , ; Finally, the sub-model Modelmid2high is iteratively trained until convergence, just like the construction process of the sub-model Modellow2mid and the sub-model Modelmid2high;

[0092] S40. Constructing an aggregation model;

[0093] S41. Build branch;

[0094] like Figure 6 As shown, two branches are constructed using sub-models Modellow2high, Modellow2mid and Modelmid2high. The first branch only includes sub-model Modellow2high, and the second branch includes sub-models Modellow2mid and Modelmid2high; among them, the output of sub-model Modellow2mid , As part of the input of the submodel Modelmid2high. The input of all branches is low-fidelity data, and the output is high-fidelity data. The x2low model connects the starting points of the two branches and is used to provide low-fidelity lift and drag coefficients. , ;

[0095] S42. Allocate aggregation weights;

[0096] In this embodiment, the weight parameter of the first branch is ρ, and correspondingly, the weight of the second branch is 1-ρ; let ρ=0.5;

[0097] S43. Weighted aggregation, feedback update;

[0098] First, the incoming flow condition in the high-fidelity data Input to the x2low model, output , ;

[0099] Then, and , Concatenate into one-dimensional vector , input two branches at the same time, and get two high-fidelity prediction results and ;

[0100] Next, weighted aggregation of the results of all branches is performed to obtain the final prediction result. :

[0101] ;

[0102] Finally, feedback is used to update the aggregate model until convergence.

[0103] The training process of the aggregate model only adjusts the weight parameters of each branch, and the sub-models and x2low model do not participate in the update.

[0104] The results of this example are shown in Figure 9a-1~Figure 9f-1 , Figure 9a-2~Figure 9f-2 , Figure 10a-1~Figure 10f-1 and Figure 10a-2~Figure 10f-2 .in, Figure 9a-1 , Figure 9b-1 , Figure 9c-1 , Figure 9d-1 , Figure 9e-1 , Figure 9f-1 The C of airfoil sg6043 at different Reynolds numbers are given respectively. L ~α prediction curve chart, Figure 9a-2 , Figure 9b-2 , Figure 9c-2 , Figure 9d-2 , Figure 9e-2 , Figure 9f-2 The C of airfoil sg6043 at different Reynolds numbers are given respectively. L ~α Relative error bar graph corresponding to the prediction curve; Figure 10a-1 , Figure 10b-1 , Figure 10c-1 , Figure 10d-1 , Figure 10e-1 , Figure 10f-1 The C of airfoil sg6043 at different Reynolds numbers is given. D ~α prediction curve chart, Figure 10a-2 , Figure 10b-2 , Figure 10c-2 , Figure 10d-2 , Figure 10e-2 , Figure 10f-2 The C of airfoil sg6043 at different Reynolds numbers is given. D The relative error bar graph corresponding to the prediction curve of ~α. It can be seen that the modeling method for complex multi-source data fusion in this embodiment not only learns the characteristic trend of aerodynamic data based on a large amount of medium-fidelity data and low-fidelity data, but also captures key detail features from a small amount of high-fidelity data, and utilizes the favorable information provided by multi-source and multi-fidelity data to achieve the prediction of the high-fidelity lift and drag coefficient of airfoil sg6043 based on the incoming flow conditions.

Claims

1. A modeling method for complex multi-source data fusion, characterized in that: The following steps are involved: S10. Define fidelity based on data quality; For given engineering calculation data, CFD numerical simulation data and wind tunnel test data, the data are evaluated according to the authenticity of the physical characteristics reflected by the data. The closer to the reality, the higher the data quality and the higher the data fidelity; Define engineering calculation data as low-fidelity data, CFD numerical simulation data as medium-fidelity data, and wind tunnel test data as high-fidelity data; S20. Build the front-end model; For low-fidelity data, each data consists of the angle of attack α, Reynolds number Re, lift coefficient C L And the drag coefficient C D Composition; among them, the angle of attack α and the Reynolds number Re are used as input conditions, and the lift coefficient C L and the drag coefficient C D As the prediction object; construct the front-end model Low net Model from the low-fidelity incoming flow condition Xlow to the corresponding low-fidelity data Ylow; the front-end model Low net Model reflects the mapping relationship between the low-fidelity incoming flow condition Xlow and the low-fidelity data Ylow, and is a proxy model for the low-fidelity data; the model structure of the front-end model Low net Model is a fully connected network; The input low-fidelity incoming flow condition Xlow is , the output low-fidelity data Ylow is : ； S30. Construct sub-model; The low-fidelity incoming flow condition Xlow and the low-fidelity data Ylow construct the sub-model Modellow2high and the sub-model Modellow2mid, and the medium-fidelity incoming flow condition Xmid and the medium-fidelity data Ymid construct the sub-model Modelmid2high; each sub-model uses the same construction method, and each sub-model structure includes two parts: a nonlinear network with a nonlinear activation function and a nonlinear network without a nonlinear activation function; the sub-model Modellow2high outputs high-fidelity data Yhigh1, the sub-model Modellow2mid outputs medium-fidelity data Ymid, and the sub-model Modelmid2high outputs high-fidelity data Yhigh2; Inputs for all sub-models , Output The description is as follows: ； in, , represents the prediction result of the mapping model under the corresponding incoming flow condition; Use all the mid-fidelity data to build the sub-model Modellow2mid; Input into the x2low model to get the corresponding low-fidelity , Predict the results; then, and , Concatenate into a one-dimensional vector as the input of the sub-model Modellow2mid , and then iterate the training until convergence; Use high-fidelity data to build the sub-model Modellow2high, and the process is the same as that of the sub-model Modellow2mid; Then, all high-fidelity data are used to construct the sub-model Modelmid2high; first, the x2low model is used to obtain the corresponding incoming flow conditions , The prediction results; then, the future flow conditions and , Input the trained sub-model Modellow2mid and get the corresponding , ; Finally, the sub-model Modelmid2high is iteratively trained until convergence, just like the construction process of the sub-model Modellow2mid and the sub-model Modelmid2high; S40. Constructing an aggregation model; S41. Build branch; Use all sub-models to build several different branches; the construction of branches follows the following rules: each branch is a separate sub-model or multiple sub-models connected end to end; the starting point of each branch, i.e. the input, is low-fidelity data Ylow; the end point of each branch, i.e. the output, is high-fidelity data Yhigh; for branches built by multiple sub-models, the sub-models are connected in the order of increasing data fidelity; Use sub-model Modellow2high, sub-model Modellow2mid and sub-model Modelmid2high to construct two branches; the first branch only includes sub-model Modellow2high, and the second branch includes sub-model Modellow2mid and sub-model Modelmid2high; among them, the output of sub-model Modellow2mid , As part of the input of the sub-model Modelmid2high; the input end of all branches is the low-fidelity data end, and the output end is the high-fidelity data end; the x2low model connects the starting points of the two branches to provide low-fidelity lift and drag coefficients , ; S42. Allocate aggregation weights; Assign a weight parameter to each branch, which represents the contribution of each branch to the result; the weight parameter value is between 0 and 1, and the sum of the weight parameters of all branches is 1; S43. Weighted aggregation, feedback update; According to the weight parameters of each branch, the results of all branches are weighted and aggregated to obtain the predicted high-fidelity data Yhigh; then, the predicted high-fidelity data Yhigh is compared with the true value, and then fed back to S42 to adjust the weight parameters of each branch until the difference between the predicted high-fidelity data Yhigh and the true value reaches a preset error range.

Images