Aircraft distributed load multi-fidelity fusion method and system combined with space coordinates

Abstract

The invention discloses an aircraft distributed load multi-fidelity fusion method and system combined with space coordinates, and the method comprises the steps: (1) calculating a distributed load data set of an aircraft under different working conditions through numerical simulation, and defining the data set as low-fidelity aerodynamic data; obtaining a distributed load data set of the aircraft under different working conditions through an actual test, and defining the distributed load data set as high-fidelity pneumatic data; (2) performing preliminary fusion on the low-fidelity aerodynamic data by using high-fidelity data in the pre-trained original data fusion layer to improve the quality of the low-fidelity aerodynamic data; (3) space coordinate features of the high-fidelity data and the preliminarily fused low-fidelity data are optimized in the pre-trained self-defined space coordinate data layer, and self-defined high-fidelity data and self-defined low-fidelity data are generated; and (4) further fusing the self-defined high-fidelity data and the self-defined low-fidelity data in the pre-trained combined space coordinate data fusion layer. According to the invention, the high-precision distributed load data of the aircraft can be obtained at low cost.

Description

Technical Field

[0001] The present invention belongs to the field of aircraft aerodynamics, and in particular relates to a multi-fidelity fusion method and system for aircraft distributed loads combined with space coordinates. Background Art

[0002] In the field of aerodynamics, wind tunnel experiments, numerical calculations, and simulated flight tests are usually used to obtain aerodynamic data of aircraft. However, each method has its own advantages and disadvantages. At present, it is very difficult to improve the accuracy of any of the above aerodynamic evaluation methods in a short period of time, and it requires high costs.

[0003] For example, a Chinese patent document with publication number CN112182771A discloses a data processing method based on numerical simulation, including: determining a flow field numerical calculation type in a fluid mechanics numerical simulation, and according to the flow field numerical calculation type, correcting flow field values ​​solved by an NS equation solution type, flow field values ​​solved by a boundary condition solution type, and flow field values ​​solved by a turbulence equation solution type; and performing a numerical simulation of aerodynamic or thermal characteristics of an aircraft based on the correction result.

[0004] A Chinese patent document with publication number CN114168796A discloses a method for establishing an aircraft high-altitude aerodynamic database, which relates to the field of aircraft research, including: obtaining aircraft shape parameters; setting state parameters, and generating a number of virtual flight states based on the state parameters; performing numerical simulation calculations on each of the virtual flight states based on the aircraft shape parameters, and obtaining corresponding aerodynamic data for each of the virtual flight states; and obtaining an aircraft high-altitude aerodynamic database based on the aerodynamic data corresponding to all the virtual flight states.

[0005] In order to shorten the development cycle and reduce the design cost, the establishment of an aerodynamic proxy model as a means of efficiently obtaining aerodynamic data has attracted much attention. The establishment of such a model usually involves the physical characteristics of the aircraft, mathematical statistical laws, or fitting and interpolating existing data to predict the response of any flight state. Although the establishment of an aerodynamic proxy model is an efficient way to obtain aerodynamic data, the premise for establishing a good model is sufficient and high-precision aerodynamic data. It is difficult to significantly improve the accuracy of the proxy model's aerodynamic data through a single data.

[0006] In order to solve the above problems, it is urgent to design a new data fusion technology to integrate aerodynamic data from different sources to improve the quality and scope of aerodynamic research, and to improve the accuracy of data through data management, data fusion and data reuse technology. Summary of the invention

[0007] The present invention provides a multi-fidelity fusion method and system for aircraft distributed loads combined with spatial coordinates, which can obtain high-precision distributed load data of aircraft at low cost by using limited test data and numerical simulation data.

[0008] A multi-fidelity fusion method for aircraft distributed load combined with space coordinates includes the following steps:

[0009] (1) The distributed load data set of the aircraft under different working conditions is obtained through numerical simulation calculation, which is defined as low-fidelity aerodynamic data; the distributed load data set of the aircraft under different working conditions is obtained through actual tests, which is defined as high-fidelity aerodynamic data;

[0010] (2) In the raw data fusion layer after pre-training, the low-fidelity aerodynamic data is initially fused with the high-fidelity data to obtain high-quality low-fidelity aerodynamic data, while retaining the original high-fidelity aerodynamic data;

[0011] (3) In the pre-trained custom spatial coordinate data layer, the spatial coordinate features of the high-fidelity aerodynamic data and the initially fused low-fidelity aerodynamic data are optimized to generate a set of custom high-fidelity data and custom low-fidelity data samples;

[0012] (4) In the pre-trained combined spatial coordinate data fusion layer, the spatial coordinate feature learning of the sample points is introduced to enrich the relationship information between the data sample points, and the custom high-fidelity and custom low-fidelity data sets in the previous layer are further integrated to obtain the final fused aircraft target distribution load data.

[0013] In step (2), the network structure of the original data fusion layer is a fully connected layer. During the pre-training process, the losses of the two high-fidelity original data sets and the low-fidelity original data sets are calculated respectively. The formula is:

[0014]

[0015] in, Forward calculation prediction value of high-fidelity raw data in the raw data fusion layer, The forward calculation prediction value of the low-fidelity aerodynamic data in the original data fusion layer, For high-fidelity original data, is the low-fidelity original data, l is the number of high-fidelity original data, and q is the number of low-fidelity original data; the total loss is calculated by weighted method Calculate the total loss by weighting The formula is as follows:

[0016]

[0017] Among them, ρ is an adjustable parameter, and the optimal solution is found through the grid optimization method.

[0018] In step (3), the custom spatial coordinate data layer includes a generator and a discriminator. The generator is used to generate samples with specified spatial coordinate features. During the pre-training process, the loss function of the generator is

[0019]

[0020] The discriminator selects a custom data set that satisfies the real data. The loss function of the discriminator is

[0021]

[0022] Where z is a random noise vector from a custom space, x is a real sample, G(z) is a data sample with random noise as input and generated as output, D(x) represents the probability that sample x is identified as a real sample, and D(G(z)) represents the probability that a generated sample is identified as a real sample. represents the average of all possible inputs in the latent space, p data (x) represents the distribution of real data, p z (z) represents the distribution in the latent space.

[0023] In step (4), the encoder and decoder are used in combination with the spatial coordinate data fusion layer to generate the spatial position sequence relationship between samples, and the self-attention mechanism is used to focus on the spatial coordinate sequence between samples, capture the relationship between different samples, and extract the detailed features of the data;

[0024] In the pre-training process of the spatial coordinate data fusion layer, the total loss J(θ1) is calculated in a weighted manner MSE* , the formula is

[0025]

[0026] In the formula, and To combine the spatial coordinate data fusion layer to forward calculate the predicted value, For high-fidelity original data, is the low-fidelity original data, q is the number of low-fidelity original data, w is the number of high-fidelity original data, and λ is an adjustable parameter used to control the proportion of the loss of the two data sets in the total loss. This layer is combined with the sample space coordinate sequence to fuse the customized high- and low-fidelity data again.

[0027] In the HyperBand multi-armed bandit framework, Bayesian optimization is used to select the configuration of the adjustable parameter λ. The specific process is as follows:

[0028] First, initialize the maximum resource R, reduce the factor η, and calculate the maximum number of steps s max=log η (R); then execute the following loop: randomly select hyperparameter configuration according to the current step number s, configure training cycles for each hyperparameter and evaluate its performance, and use the evaluation results to update the Bayesian optimization model; check whether the termination condition is met, if not, continue the optimization process; if the termination condition is met, output the hyperparameter λ with the best performance in all evaluations.

[0029] Furthermore, the termination condition is the total loss When the maximum number of evaluations is reached or the pre-training phase is combined with the spatial coordinate data fusion layer, the performance no longer improves significantly.

[0030] A multi-fidelity fusion system for aircraft distributed loads combined with spatial coordinates includes a memory and one or more processors. The memory stores executable code. When the one or more processors execute the executable code, they are used to implement the above-mentioned multi-fidelity fusion method for aircraft distributed loads combined with spatial coordinates.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The present invention improves the quality of low-fidelity aerodynamic data by preliminarily fusing high-fidelity data in the original data fusion layer; optimizes the spatial coordinate features of high-fidelity data and preliminarily fused low-fidelity data in the spatial coordinate data customization layer to generate customized high-fidelity data and customized low-fidelity data sample sets; introduces spatial coordinate feature learning of sample points in the combined spatial coordinate data fusion layer to enrich the relationship information between data sample points, further fuses the customized high-fidelity and customized low-fidelity data sets in the previous layer, and uses Bayesian optimization to select hyperparameter configuration under the multi-armed bandit framework of HyperBand to achieve fusion of aircraft distributed load data obtained by different methods. Combined with spatial coordinates, the data is fused twice, and the model efficiently realizes data fusion of two aerodynamic data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of a multi-fidelity fusion method of aircraft distributed load combined with spatial coordinates according to an embodiment of the present invention;

[0034] Figure 2 This is a numerical calculation simulation diagram of the ONERAM6 standard wing model according to an embodiment of the present invention;

[0035] Figure 3 This is a physical picture of the ONERAM6 standard wing wind tunnel test of an embodiment of the present invention;

[0036] Figure 4 The comparison between the results of the embodiment of the present invention and the numerical calculation results at the horizontal coordinate X=0.2 is shown in the figure;

[0037] Figure 5 The comparison between the results of the embodiment of the present invention and the numerical calculation results at the horizontal coordinate X=0.44 is shown in the figure;

[0038] Figure 6 The comparison between the results of the embodiment of the present invention and the numerical calculation results at the horizontal coordinate X=0.65 is shown in the figure;

[0039] Figure 7 The comparison between the results of the embodiment of the present invention and the numerical calculation results at the horizontal coordinate X=0.8 is shown in the figure;

[0040] Figure 8 The comparison between the results of the embodiment of the present invention and the numerical calculation results at the horizontal coordinate X=0.9 is shown in the figure;

[0041] Fig. 9 The comparison between the results of the embodiment of the present invention and the numerical calculation results at the horizontal coordinate X=0.96 is shown in the figure. DETAILED DESCRIPTION

[0042] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0043] The present invention needs to fuse two kinds of data: the first kind of test data is high-fidelity pneumatic data; the second kind of CFD numerical calculation data is low-fidelity pneumatic data. The data fusion of the two pneumatic data sets is achieved through the three data processing layers of the model.

[0044] like Figure 1 As shown, a multi-fidelity fusion method of aircraft distributed load combined with space coordinates includes the following steps:

[0045] S01, obtain the distributed load data set of the aircraft under different working conditions through numerical simulation calculation, which is defined as low-fidelity aerodynamic data; obtain the distributed load data set of the aircraft under different working conditions through actual tests, which is defined as high-fidelity aerodynamic data;

[0046] S02, in the raw data fusion layer after pre-training, the low-fidelity aerodynamic data is initially fused with the high-fidelity data to improve the quality of the low-fidelity aerodynamic data;

[0047] S03, in the pre-trained custom spatial coordinate data layer, optimizing the spatial coordinate features of the high-fidelity data and the initially fused low-fidelity data, and generating a set of custom high-fidelity data and custom low-fidelity data samples;

[0048] S04, in the combined spatial coordinate data fusion layer after pre-training, introduce the spatial coordinate feature learning of the sample points, enrich the relationship information between the data sample points, and further integrate the custom high-fidelity and custom low-fidelity data sets in the previous layer to obtain the final fused aircraft target distribution load data.

[0049] The following is a detailed description of the pre-training process of the original data fusion layer, the custom spatial coordinate data layer, and the combined spatial coordinate data fusion layer in the model.

[0050] First, the distributed pressure data set 1 of the aircraft under different working conditions is obtained by numerical calculation, and the working condition parameters (Mach number, angle of attack and position coordinates of the force point) in data set 1 are used as input features x l , the distribution pressure at different positions (C p ) as the output label y com The distributed pressure data set 2 of the aircraft under different working conditions is obtained through experiments, and the working condition parameters (Mach number, angle of attack and position coordinates of the force point) in data set 2 are used as input features x h , the distribution pressure at different positions (C p ) as the output label y exp Two working conditions in data set 1 and data set 2 are taken as test data sets respectively, and 80% of the remaining data are combined as training data and 20% of the data are combined as verification data.

[0051] Step S02: During the pre-training of the original data fusion layer, the model is trained with x l and x h Get the predicted value Calculated by custom loss function, the formula is

[0052]

[0053] in, It is the forward calculation prediction value of high-fidelity aerodynamic data in the original data fusion layer, The forward calculation prediction value of the low-fidelity aerodynamic data in the original data fusion layer, For high-fidelity original data, is the low-fidelity original data, q is the number of low-fidelity original data, l is the number of high-fidelity original data, ρ is an adjustable parameter, and the optimal solution is found through the grid optimization method.

[0054] High-fidelity data 2 and low-fidelity data 1 are forward-calculated to obtain the predicted values ​​of the labels and Through the weighted loss function Calculate the total loss, and in each training cycle, update the model parameters through backpropagation and optimizer to minimize the total loss To learn high-fidelity and low-fidelity data information. This layer initially fuses high-fidelity and low-fidelity data to generate high-quality low-fidelity aerodynamic data. The loss value of the forward calculation is obtained. ρ is set to 0.01. After calculation through this layer, the data quality of dataset 1 is improved to obtain a high-quality low-fidelity aerodynamic dataset 3, and the original high-fidelity aerodynamic dataset 2 is retained.

[0055] Step S03, in the process of pre-training the custom spatial coordinate data layer, based on the spatial coordinate features of the ONERA M6 standard wing model, samples with abscissa x of 0.2, 0.44, 0.65, 0.8, 0.9, and 0.96 are taken as groups, and the data set is divided based on this sequence. This layer consists of two parts: the generator and the discriminator. The loss function of the generator is

[0056]

[0057] The discriminator selects a custom data set that satisfies the real data. The loss function of the discriminator is

[0058]

[0059] Where z is a random noise vector from a custom space, x is a real sample, G(z) is a data sample with random noise as input and generated as output, D(x) represents the probability that sample x is identified as a real sample, and D(G(z)) represents the probability that a generated sample is identified as a real sample. represents the average of all possible inputs in the latent space, p data (x) represents the distribution of real data, p z (z) represents the distribution in the latent space.

[0060] In each training cycle, the model parameters are updated through back propagation and the optimizer, and the generator minimizes the loss L according to the conditions that the sample horizontal coordinate x is 0.2, 0.44, 0.65, 0.8, 0.9, and 0.96. G , generate samples that meet the custom spatial coordinate characteristics, pass the generated samples through the discriminator to calculate, and minimize L D The target sample data is filtered out by using loss. The high-fidelity pneumatic dataset 1 and the high-quality low-fidelity pneumatic dataset 3 in step 1 are processed by this layer to generate sample data with specific spatial coordinate features and corresponding sample labels, thereby obtaining a custom high-fidelity pneumatic dataset 4 and a custom low-fidelity pneumatic dataset 5.

[0061] Step S03, in the process of pre-training combined with the spatial coordinate data fusion layer, the encoder and decoder are used to process the sample groups when the horizontal coordinate x is 0.2, 0.44, 0.65, 0.8, 0.9, and 0.96, generate the spatial position sequence relationship between the samples, and use the self-attention mechanism to focus on the spatial coordinate sequence between the samples, capture the relationship between different samples, and extract the detailed features of the data. The total loss of the custom high-fidelity pneumatic data and the custom low-fidelity pneumatic data set is calculated in a weighted manner The formula is

[0062]

[0063] In the formula, and They are the forward calculation prediction values ​​of high-fidelity aerodynamic data and low-fidelity aerodynamic data in the combined spatial coordinate data fusion layer, For high-fidelity original data, is the low-fidelity original data, q is the number of low-fidelity original data, w is the number of high-fidelity original data, where λ is an adjustable parameter used to control the proportion of the loss of the two datasets in the total loss.

[0064] In each training cycle, the custom high-fidelity pneumatic data 4 and the custom low-fidelity pneumatic data 5 are combined with the sample space coordinate sequence in the encoder and decoder during the calculation process, and the forward calculation is performed to obtain the predicted values ​​of the labels. and Through the weighted loss function Calculate the total loss, and in each training cycle, update the model parameters through backpropagation and optimizer to minimize the total loss To learn to combine the high-fidelity and low-fidelity data information of the spatial coordinate sequence. Through this layer, the high-fidelity and low-fidelity aerodynamic data of the spatial coordinate sequence are fused to generate the target aerodynamic data.

[0065] In the pre-training process of the spatial coordinate data fusion layer, the hyperparameter with different initial values ​​λ is used to construct the proxy model (Gaussian process). Train the proxy model for the previously evaluated λ and corresponding performance value of the objective function. Allocate resources according to HyperBand's multi-arm gambling mechanism, and gradually increase the amount of resources for each hyperparameter configuration. The details are as follows: First, initialize the maximum resource R, reduce the factor η, and calculate the maximum number of steps s max =log η (R). Then execute the following loop: randomly select hyperparameter configuration according to the current step number s, configure the training cycle for each hyperparameter and evaluate its performance, and use the evaluation results to update the Bayesian optimization model. Check whether the termination condition (total loss When the maximum number of evaluations is reached or the performance of the pre-training phase combined with the spatial coordinate data fusion layer is no longer significantly improved), the best hyperparameter λ in all evaluations is output.

[0066] In a specific embodiment of the present invention, the object is the ONERA M6 standard wing. Fluent is a general fluid flow simulation software that can quickly evaluate aerodynamic performance and is used to provide a simulation calculation data set for the M6 ​​standard wing. The ONERA M6 wing test data is described in the AGARD report. For details, see Figure 2 and Figure 3 . The operating parameters Mach number, angle of attack and position coordinates of the force point are taken as input features, and the distributed pressure at different positions is taken as the output label. The Mach number range is 0.6998 to 0.9262, and the corresponding angle of attack range is 0.04 to 4.08. The test data set contains 679 samples, and the calculation simulation data contains 7906 samples. The test data set results of this embodiment show that the average absolute error of the multi-fidelity fusion method model of the aircraft distributed load combined with spatial coordinates of the present invention is better than the numerical calculation results. The example results of the present invention are compared with the numerical calculation results at the horizontal coordinates X=0.2, 0.44, 0.65, 0.8, 0.9, and 0.96, respectively as shown in the following figure. Figures 4 to 9 As shown in the figure, the horizontal axis is the spatial horizontal coordinate of the sample point, the vertical axis is the pressure value at the sample point, the curve CFD is the numerical calculation pressure result, the curve EXP is the actual pressure value of the aircraft, and the curve DNN is the pressure calculation result of the example of the present invention.

[0067] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-fidelity fusion method for aircraft distributed loads combined with spatial coordinates, characterized in that: The following steps are involved: (1) The distributed load data set of the aircraft under different working conditions is obtained through numerical simulation calculation, which is defined as low-fidelity aerodynamic data; the distributed load data set of the aircraft under different working conditions is obtained through actual tests, which is defined as high-fidelity aerodynamic data; (2) In the raw data fusion layer after pre-training, the low-fidelity aerodynamic data is initially fused with the high-fidelity data to obtain high-quality low-fidelity aerodynamic data, while retaining the original high-fidelity aerodynamic data; (3) In the pre-trained custom spatial coordinate data layer, the spatial coordinate features of the high-fidelity aerodynamic data and the initially fused low-fidelity aerodynamic data are optimized to generate a set of custom high-fidelity data and custom low-fidelity data samples; (4) In the pre-trained combined spatial coordinate data fusion layer, the spatial coordinate feature learning of the sample points is introduced to enrich the relationship information between the data sample points, and the custom high-fidelity and custom low-fidelity data sets in the previous layer are further integrated to obtain the final fused aircraft target distribution load data.

2. The multi-fidelity fusion method of aircraft distributed load combined with spatial coordinates according to claim 1 is characterized in that: In step (2), the network structure of the original data fusion layer is a fully connected layer. During the pre-training process, the losses of the two high-fidelity original data sets and the low-fidelity original data sets are calculated respectively. The formula is: in, It is the forward calculation prediction value of high-fidelity aerodynamic data in the original data fusion layer, The forward calculation prediction value of the low-fidelity aerodynamic data in the original data fusion layer, For high-fidelity original data, is the low-fidelity original data, l is the number of high-fidelity original data, and q is the number of low-fidelity original data; the total loss is calculated by weighted method By minimizing the total loss Come learn about high-fidelity and low-fidelity data information.

3. The multi-fidelity fusion method of aircraft distributed load combined with spatial coordinates according to claim 2 is characterized in that: Calculate the total loss by weighting The formula is as follows: Among them, ρ is an adjustable parameter, and the optimal solution is found through the grid optimization method.

4. The multi-fidelity fusion method of aircraft distributed load combined with spatial coordinates according to claim 1 is characterized in that: In step (3), the custom spatial coordinate data layer includes a generator and a discriminator. The generator is used to generate samples with specified spatial coordinate features. During the pre-training process, the loss function of the generator is The discriminator selects a custom data set that satisfies the real data. The loss function of the discriminator is Where z is a random noise vector from a custom space, x is a real sample, G(z) is a data sample with random noise as input and generated as output, D(x) represents the probability that sample x is identified as a real sample, and D(G(z)) represents the probability that a generated sample is identified as a real sample. represents the average of all possible inputs in the latent space, p data (x) represents the distribution of real data, p z (z) represents the distribution in the latent space.

5. The multi-fidelity fusion method of aircraft distributed load combined with spatial coordinates according to claim 1 is characterized in that: In step (4), the encoder and decoder are used in combination with the spatial coordinate data fusion layer to generate the spatial position sequence relationship between samples, and the self-attention mechanism is used to focus on the spatial coordinate sequence between samples, capture the relationship between different samples, and extract the detailed features of the data; Combined with the pre-training process of the spatial coordinate data fusion layer, the total loss is calculated in a weighted manner The formula is In the formula, and They are the forward calculation prediction values ​​of high-fidelity aerodynamic data and low-fidelity aerodynamic data in the combined spatial coordinate data fusion layer, For high-fidelity original data, is the low-fidelity original data, q is the number of low-fidelity original data, w is the number of high-fidelity original data, and λ is an adjustable parameter used to control the proportion of the loss of the two data sets in the total loss. This layer is combined with the sample space coordinate sequence to fuse the customized high- and low-fidelity data again.

6. The multi-fidelity fusion method of aircraft distributed load combined with spatial coordinates according to claim 5 is characterized in that: In the HyperBand multi-armed bandit framework, Bayesian optimization is used to select the configuration of the adjustable parameter λ. The specific process is as follows: First, initialize the maximum resource R, reduce the factor η, and calculate the maximum number of steps s max =log η (R); then execute the following loop: randomly select hyperparameter configuration according to the current step number s, configure training cycles for each hyperparameter and evaluate its performance, and use the evaluation results to update the Bayesian optimization model; check whether the termination condition is met, if not, continue the optimization process; If the termination condition is met, output the hyperparameter λ with the best performance in all evaluations.

7. The multi-fidelity fusion method of aircraft distributed load combined with spatial coordinates according to claim 6 is characterized in that: Termination condition is total loss When the maximum number of evaluations is reached or the pre-training phase is combined with the spatial coordinate data fusion layer, the performance no longer improves significantly.

8. A multi-fidelity fusion system for aircraft distributed loads combined with spatial coordinates, characterized in that: It comprises a memory and one or more processors, wherein the memory stores executable codes, and when the one or more processors execute the executable codes, they are used to implement the multi-fidelity fusion method of aircraft distributed load combined with spatial coordinates as described in any one of claims 1-7.

Images