A method and system for equivalent airflow shear force based on PCA and neural networks
By combining PCA with neural networks, a surrogate model was constructed and optimized using a genetic algorithm. This solved the accuracy and efficiency problems of equivalent shear force on the surface of an aircraft, and enabled efficient and accurate simulation of wind tunnel tests.
Patent Information
- Application Number
- CN202611079087.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies cannot accurately obtain the shear force on the surface of an aircraft, wind tunnel tests cannot reproduce the actual flight conditions, and traditional methods have low equivalent accuracy and poor operating condition matching efficiency, which cannot meet the needs of rapid engineering design.
By combining PCA with neural networks, a surrogate model is constructed, and a genetic algorithm is used for optimization to achieve equivalence between wind tunnel conditions and real flight airflow shear forces. This includes principal component analysis, neural network training, and iterative optimization.
It achieves efficient processing of high-dimensional data, improves equivalent accuracy and operating condition matching efficiency, shortens the design cycle, and outputs reliable wind tunnel test operating condition parameters.
Smart Images

Figure CN123116902A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of aerodynamics technology, and more specifically, relates to an equivalent method and system for airflow shear force based on PCA and neural networks. Background Technology
[0002] In the field of laser irradiation effect research, the shear force on the surface of high-speed aircraft directly determines the irradiation response characteristics, and the shear force causes significant surface erosion. Real-world laser irradiation experiments are costly, time-consuming, and difficult to collect data. The industry often uses wind tunnel simulations of shear force to conduct ground-based equivalent tests, but practical applications face two major challenges.
[0003] Accurately obtaining the shear force on the surface of an aircraft is difficult. Existing methods include flight testing, wind tunnel testing, and CFD simulation. Flight test data is accurate, but it is high-risk, has poor repeatability, and cannot support multiple iterative verifications. CFD simulation can simulate the shear force distribution in the flow field, but its accuracy is limited by the mesh and turbulence model. The aircraft mesh forms high-dimensional data, which is computationally time-consuming and cannot meet the needs of rapid engineering design.
[0004] Wind tunnels are constrained by physical conditions and cannot replicate real flight conditions. Wind tunnel testing is highly controllable, repeatable, and lower in cost, making it the mainstream ground testing method. However, it can only construct high-density, low-Mach-number flow fields and cannot replicate the high-altitude, low-density, high-Mach-number flight environment. The shear force distribution on the object surface deviates significantly from the actual state.
[0005] Current methods for equivalent flight conditions rely heavily on empirical formulas or simplified machine learning models, which have significant drawbacks. Empirical formulas depend on human experience, lack versatility, and require re-deriving when the aircraft configuration changes. Simplified machine learning struggles with high-dimensional grid data, is prone to overfitting leading to insufficient prediction accuracy, and fails to fully integrate multi-condition characteristics, resulting in inefficient condition matching and optimization, and difficulty in quickly aligning real flight conditions with wind tunnel conditions. Therefore, the industry urgently needs an equivalent design method for surface shear force that can efficiently process high-dimensional data, improve equivalence accuracy, and shorten the design cycle. Summary of the Invention
[0006] This invention aims to overcome the shortcomings of existing methods in handling high-dimensional shear force data, low equivalent accuracy, and poor working condition matching efficiency. By constructing a surrogate model through PCA combined with neural networks and optimizing it with genetic algorithms, the wind tunnel working conditions are equivalent to the shear force of real flight airflow, which can efficiently support the equivalent test of laser irradiation on the ground.
[0007] To address the aforementioned deficiencies or improvement needs of existing technologies, as a first aspect of this invention, the present invention provides an equivalent method for airflow shear force based on PCA and neural networks, comprising: S1. Obtain the first operating condition parameters and first airflow shear force data of multiple real flight operating condition samples, as well as the second operating condition parameters and second airflow shear force data of multiple wind tunnel test operating condition samples, and construct the original dataset; S2. Perform principal component analysis on the first airflow shear force data and the second airflow shear force data in the original dataset, extract the principal component feature vectors, and calculate the corresponding principal component coefficient data; S3. Construct a neural network model, using the first working condition parameters and the second working condition parameters as inputs and the corresponding principal component coefficient data as outputs for training, to obtain a trained neural network proxy model; S4. Using the airflow shear force under actual flight conditions as the target value, the wind tunnel test conditions parameters are input as decision variables into the trained neural network surrogate model. The model is iteratively optimized using a genetic algorithm. When the error between the reconstructed wind tunnel test airflow shear force and the actual flight conditions airflow shear force meets the preset error threshold, the optimal wind tunnel test conditions parameters are output to complete the airflow shear force equivalence.
[0008] Furthermore, in S1, the actual flight condition is a low-density, high-Mach number condition; the wind tunnel test condition is a high-density, low-Mach number condition; and the condition parameters include Mach number, density, temperature, and pressure.
[0009] Furthermore, in step S2, the principal component feature vectors are extracted, and the corresponding principal component coefficient data are calculated, including an adaptive processing mechanism, specifically: The airflow shear force data in the original dataset are subjected to eigenvalue decomposition to obtain multiple eigenvalues arranged in descending order of numerical value; The decay rate is calculated based on two adjacent feature values in the feature value sequence. The feature value index that the decay rate is first less than a preset decay threshold is selected and used as the initial principal component dimension number. For the airflow shear force data in the original dataset, calculate the shear force gradient value at each grid node on the aircraft surface, and count the proportion of the number of grid nodes with shear force gradient values greater than a preset gradient threshold to the total number of grid nodes. Use this proportion as an indicator of flow field complexity. The initial principal component dimension number is corrected based on the flow field complexity index to obtain the target principal component dimension number; wherein, when the flow field complexity index is greater than a preset complexity threshold, a preset number of dimension compensation values are added to the initial principal component dimension number. The number of feature vectors of the target principal component dimension that are first extracted and sorted are used as the principal component feature vectors. The airflow shear force data in the original dataset is projected onto the space formed by the principal component feature vectors to calculate the principal component coefficient data.
[0010] Furthermore, the attenuation rate is calculated as follows: In the formula, Indicates the first The attenuation rate at each location; This indicates the number of items sorted in descending order of numerical value. One eigenvalue; Indicates the first One eigenvalue; This represents the total number of the feature values; Indicates the summation index variable; Represents the sum of all eigenvalues; It is a positive integer, and .
[0011] Furthermore, the process of correcting the initial principal component dimension number based on the flow field complexity index to obtain the target principal component dimension number is as follows: Among them, the target cumulative variance contribution rate The calculation formula is: In the formula, This indicates the number of dimensions of the target principal components; Indicates the number of candidate dimensions, which is a positive integer; Represents the set of positive integers; This indicates the number of items sorted in descending order of numerical value. One eigenvalue; This represents the total number of the feature values; Indicates the summation index variable; The complexity index represents the flow field complexity. The dynamically changing cumulative variance contribution rate of the target; This represents the preset baseline cumulative variance contribution rate; This represents the preset maximum cumulative variance contribution rate; This represents the maximum reference value of the flow field complexity index; This means that the cumulative variance contribution rate is greater than or equal to Under the given conditions, take the smallest positive integer. As .
[0012] Furthermore, the neural network model in S3 includes an input layer, at least two hidden layers, and an output layer; The training process, which uses the first operating condition parameters and the second operating condition parameters as inputs and the corresponding principal component coefficient data as outputs, includes: The first operating condition parameter and the second operating condition parameter are input into the neural network model, and nonlinear mapping is performed through the hidden layer. The predicted principal component coefficient data are output by the output layer. Calculate the mean squared error between the predicted principal component coefficient data and the actual principal component coefficient data, and construct a loss function based on the mean squared error; The backpropagation algorithm and adaptive moment estimation optimizer are used to update the weight parameters of the neural network model according to the loss function until the loss function converges, thus obtaining the trained neural network surrogate model.
[0013] Furthermore, in step S4, the process of obtaining the reconstructed wind tunnel test airflow shear force includes: The predicted principal component coefficient data output by the trained neural network surrogate model is multiplied with the principal component feature vector to obtain the reconstructed one-dimensional vector of airflow shear force. The reconstructed one-dimensional vector of airflow shear force is mapped back to the aircraft surface grid according to the preset grid node order to obtain the reconstructed wind tunnel test airflow shear force.
[0014] Furthermore, S4 also includes: Obtain surface shear force data measured through wind tunnel testing; compare the surface shear force data with the equivalent airflow shear force output by the neural network model, and calculate the error between the two; If the error is greater than or equal to a preset error threshold, then return to step S2 to supplement CFD simulation data, and repeat steps S3 and S4 to update and train the neural network model until the error is less than the preset error threshold.
[0015] As a second aspect of the present invention, an equivalent system of airflow shear force based on PCA and neural networks is also provided, comprising: The original dataset construction unit is used to obtain the first operating condition parameters and first airflow shear force data of multiple real flight operating condition samples, as well as the second operating condition parameters and second airflow shear force data of multiple wind tunnel test operating condition samples, to construct the original dataset; The principal component feature extraction unit is used to perform principal component analysis on the first airflow shear force data and the second airflow shear force data in the original dataset, extract the principal component feature vectors, and calculate the corresponding principal component coefficient data. The neural network model training unit is used to construct a neural network model, taking the first working condition parameters and the second working condition parameters as inputs and the corresponding principal component coefficient data as outputs for training, to obtain a trained neural network surrogate model. The operating condition parameter iterative optimization unit is used to take the airflow shear force of the real flight condition as the target value, input the wind tunnel test operating condition parameters as decision variables into the trained neural network surrogate model, and perform iterative optimization through a genetic algorithm. When the error between the reconstructed wind tunnel test airflow shear force and the airflow shear force of the real flight condition meets the preset error threshold, the optimal wind tunnel test operating condition parameters are output to complete the airflow shear force equivalence.
[0016] As a third aspect of the invention, a computer-readable storage medium is also provided, on which a computer program is stored, which is executed by a processor as described in any one of the claims, a method for equivalent airflow shear force based on PCA and neural networks.
[0017] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. The airflow shear force equivalent method based on PCA and neural networks of this invention achieves dimensionality reduction of high-dimensional data by performing adaptive principal component analysis on the shear force data of the aircraft surface. First, the one-dimensional dimensionless vector of shear force is decomposed into eigenvalues. The initial number of principal components is determined by combining the eigenvalue decay rate. Then, the cumulative variance contribution rate is dynamically adjusted according to the flow field complexity index to obtain the target principal component dimension that adapts to the flow field characteristics. After the original shear force data is projected onto the principal component space, it is transformed into low-dimensional coefficients, fully preserving the key features of the flow field and effectively reducing the computational load brought by massive grid data, thus reducing the computational cost for subsequent model training and operational condition optimization.
[0018] 2. The airflow shear force equivalent method based on PCA and neural networks of this invention constructs a multi-layer neural network and completes surrogate model training using standardized operating conditions and principal component coefficient samples. The network takes flight and wind tunnel operating condition parameters as inputs and outputs the corresponding principal component coefficients. The weights are updated using a mean squared error loss function and an adaptive moment estimation optimizer until convergence. The trained model can quickly map operating conditions to shear force features, avoiding the overfitting problem that occurs when simple machine learning models process high-dimensional data, eliminating the need for repeated CFD simulations, and stably outputting reliable principal component coefficient prediction results.
[0019] 3. The airflow shear force equivalent method based on PCA and neural networks of this invention uses a genetic algorithm with wind tunnel operating parameters as decision variables for iterative optimization. Using the shear force under actual flight conditions as the benchmark, the complete surface shear force is reconstructed using model prediction coefficients combined with principal component eigenvectors. The error between the reconstructed data and the target data is compared to determine the superiority of the operating conditions. If the measured equivalent error does not meet the threshold, simulation data can be added to re-execute feature extraction and model training. By relying on closed-loop iteration to continuously optimize parameters, the method improves the low efficiency of traditional equivalent scheme operating condition search and outputs the optimal wind tunnel test operating conditions that meet the error requirements. Attached Figure Description
[0020] Figure 1 This is a flowchart of an equivalent method for airflow shear force based on PCA and neural networks according to an embodiment of the present invention; Figure 2 This is an overall flowchart of the equivalent design method for surface shear force according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the equivalent shear force at the wing surface of a high-speed aircraft according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the system units in an embodiment of the present invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0022] Example 1 Please refer to Figure 1 This embodiment 1 provides an equivalent method for airflow shear force based on PCA and neural networks, including: S1. Obtain the first operating condition parameters and first airflow shear force data of multiple real flight operating condition samples, as well as the second operating condition parameters and second airflow shear force data of multiple wind tunnel test operating condition samples, and construct the original dataset; S2. Perform principal component analysis on the first airflow shear force data and the second airflow shear force data in the original dataset, extract the principal component feature vectors, and calculate the corresponding principal component coefficient data; S3. Construct a neural network model, using the first working condition parameters and the second working condition parameters as inputs and the corresponding principal component coefficient data as outputs for training, to obtain a trained neural network proxy model; S4. Using the airflow shear force under actual flight conditions as the target value, the wind tunnel test conditions parameters are input as decision variables into the trained neural network surrogate model. The model is iteratively optimized using a genetic algorithm. When the error between the reconstructed wind tunnel test airflow shear force and the actual flight conditions airflow shear force meets the preset error threshold, the optimal wind tunnel test conditions parameters are output to complete the airflow shear force equivalence.
[0023] Please refer to Figure 2 This embodiment 1 further elaborates on the above steps.
[0024] (1) Construction of the original dataset In this embodiment, constructing the original dataset requires acquiring sample data under two different operating conditions: real flight condition samples and wind tunnel test condition samples. Real flight conditions typically exhibit low density and high Mach number physical characteristics, and multiple flight state points in the high-altitude atmospheric environment are selected as sample sources. For each real flight condition sample, its corresponding Mach number, atmospheric density, atmospheric temperature, and atmospheric pressure, etc., are acquired as first operating condition parameters, and first airflow shear force data on the aircraft surface are obtained through computational fluid dynamics numerical simulation. In contrast, wind tunnel test conditions are limited by the capabilities of ground equipment and exhibit high density and low Mach number physical characteristics, and multiple test state points in the wind tunnel test section with gas density higher than the high-altitude environment and Mach number lower than the real flight condition are selected as sample sources. For each wind tunnel test condition sample, its Mach number, airflow density, airflow temperature, and airflow pressure, etc., are acquired as second operating condition parameters, and second airflow shear force data on the aircraft surface are obtained through numerical simulation or historical wind tunnel test.
[0025] After obtaining the two types of airflow shear force data mentioned above, data preprocessing is required. After meshing the aircraft surface, each mesh node corresponds to an airflow shear force value. The three-dimensional airflow shear force distribution data of the aircraft surface is flattened into a one-dimensional vector according to a preset mesh node order. The preset mesh node order is arranged sequentially along the axial direction from the aircraft's nose to its tail, and also along the spanwise direction of the aircraft's circumference. Following this preset order, the airflow shear force values at each mesh node are extracted sequentially and combined into a one-dimensional vector. This processing method transforms the complex space physics field into a mathematical vector that the algorithm can process.
[0026] To eliminate the gradient impact of differences in the absolute values of airflow shear force under different operating conditions on subsequent model training, the flattened one-dimensional vector is dimensionless. The Min-Max normalization method is used, calculating the minimum and maximum airflow shear force values in the current sample one-dimensional vector to map each original airflow shear force value in the vector to a standard interval. After dimensionless processing, the values in all airflow shear force one-dimensional vectors are at the same scale, ensuring that large differences between different physical quantities do not cause convergence difficulties in subsequent neural network training.
[0027] After data preprocessing, the parameters of the actual flight conditions are correlated with the preprocessed one-dimensional vector of airflow shear force, and the parameters of the wind tunnel test conditions are also correlated with the preprocessed one-dimensional vector of airflow shear force to jointly construct the original dataset. The constructed original dataset is divided into a training set, a validation set, and a test set according to a preset ratio. The training set is used for subsequent parameter updates of the neural network model, the validation set is used for hyperparameter adjustment and overfitting monitoring during model training, and the test set is used to evaluate the generalization ability of the final model.
[0028] (2) Principal component feature extraction In this embodiment, principal component analysis is performed on the first and second airflow shear force data in the original dataset. An adaptive processing mechanism is used to extract the principal component eigenvectors and calculate the corresponding principal component coefficients. Since the first airflow shear force data corresponds to actual flight conditions and the second airflow shear force data corresponds to wind tunnel test conditions, the two types of data differ in physical characteristics and numerical distribution. Therefore, the two types of data need to be processed uniformly during the principal component analysis to ensure the effectiveness of the subsequent equivalent transformation.
[0029] Before eigenvalue decomposition, the first and second airflow shear force data in the original dataset are merged to construct a unified airflow shear force data matrix. Each row of the data matrix represents a one-dimensional vector of airflow shear force for a sample, and each column represents the shear force value at a certain grid node on the aircraft surface. Since the airflow shear force data for the two operating conditions have already undergone dimensionless processing in the previous steps, they can be directly merged to form a unified data matrix. The covariance matrix is then calculated on the merged data matrix. The covariance matrix reflects the correlation between the airflow shear force values at different grid nodes, providing a foundation for subsequent eigenvalue decomposition.
[0030] Eigenvalue decomposition is performed on the covariance matrix of the unified data matrix, yielding multiple eigenvalues arranged in descending order of numerical value. Eigenvalue decomposition is a core step in principal component analysis. By solving for the eigenvalues and eigenvectors of the covariance matrix, the variance distribution of the data in different directions can be obtained. Each eigenvalue corresponds to an eigenvector, and the magnitude of the eigenvalue reflects the degree of variance contribution of the data in that eigenvector direction. Arranging the eigenvalues in descending order ensures that the eigenvectors at the beginning capture more information about the changes in the data. Since the first and second airflow shear force data are decomposed in the same feature space, the two types of data share the same eigenvector basis, providing a unified mathematical framework for subsequent equivalent transformations.
[0031] The decay rate is calculated based on two adjacent eigenvalues in the eigenvalue sequence. The method for calculating the decay rate is as follows: ,in Indicates the first The attenuation rate at each location; This indicates the number of items sorted in descending order of numerical value. One eigenvalue; Indicates the first One eigenvalue; This represents the total number of the feature values; Indicates the summation index variable; Represents the sum of all eigenvalues; It is a positive integer, and .
[0032] The decay rate reflects the degree of difference between adjacent feature values. A large decay rate indicates a significant difference between the current feature value and the next feature value, and the corresponding feature vector contains important information. Feature values whose decay rate first falls below a preset decay threshold are selected, and these feature values are used as the initial number of principal component dimensions. When the decay rate first falls below the preset decay threshold, it indicates that the changes in subsequent feature values tend to be gradual, and the corresponding feature vector contains relatively little information. Determining the initial number of principal component dimensions in this way can retain key information while avoiding the introduction of too many noisy dimensions.
[0033] For the airflow shear force data in the original dataset, the shear force gradient values at each grid node on the aircraft surface are calculated, and the proportion of grid nodes with shear force gradient values greater than a preset gradient threshold is counted out of the total number of grid nodes. This proportion is used as the flow field complexity index. The flow field complexity index reflects the degree of spatial variation of airflow shear force on the aircraft surface. When complex flow phenomena such as shock waves and separated flows exist in the flow field, the airflow shear force will generate a large gradient in local regions, leading to an increase in the flow field complexity index. Since the first and second airflow shear force data correspond to different operating conditions, it is necessary to calculate the flow field complexity index for the two types of data separately, and take the larger value as the final flow field complexity index to ensure that the number of principal component dimensions can meet the feature extraction requirements of the two types of data.
[0034] The initial number of principal component dimensions is adjusted based on the flow field complexity index to obtain the target number of principal component dimensions. When the flow field complexity index exceeds a preset complexity threshold, a preset number of dimensionality compensation values are added to the initial number of principal component dimensions. The physical significance of this adjustment mechanism is that for complex flow fields, the spatial distribution of airflow shear force is more complex and variable, requiring more principal component dimensions to fully capture flow field features; while for relatively simple flow fields, fewer principal component dimensions are sufficient to meet feature extraction requirements. The specific calculation process for the target number of principal component dimensions is as follows: The contribution rate of the target cumulative variance The calculation formula is: ; This indicates the number of dimensions of the target principal components; Indicates the number of candidate dimensions, which is a positive integer; Represents the set of positive integers; This indicates the number of items sorted in descending order of numerical value. One eigenvalue; This represents the total number of the feature values; Indicates the summation index variable; The complexity index represents the flow field complexity. The dynamically changing cumulative variance contribution rate of the target; This represents the preset baseline cumulative variance contribution rate; This represents the preset maximum cumulative variance contribution rate; This represents the maximum reference value of the flow field complexity index; This means that the cumulative variance contribution rate is greater than or equal to Under the given conditions, take the smallest positive integer. As .
[0035] Target cumulative variance contribution rate With flow field complexity index It increases linearly with the increase of [variable name]. When the flow field complexity is low, the objective cumulative variance contribution rate is close to the baseline value. The required number of principal component dimensions is relatively small; when the flow field complexity is high, the objective cumulative variance contribution rate approaches its maximum value. More principal component dimensions are needed to meet higher information retention requirements. This dynamic adjustment mechanism enables principal component analysis to adapt to flow fields of varying complexity, controlling computational complexity while ensuring feature extraction quality.
[0036] The top-ranked eigenvectors of the target principal component dimension are extracted as principal component eigenvectors. The first and second airflow shear force data from the original dataset are then projected onto the space formed by these principal component eigenvectors, respectively, to calculate the corresponding principal component coefficients. These principal component eigenvectors form a new low-dimensional coordinate system. Projecting the original high-dimensional airflow shear force data onto this coordinate system yields a set of principal component coefficients for each sample. The first airflow shear force data, after projection, yields the first principal component coefficient data, and the second airflow shear force data, after projection, yields the second principal component coefficient data. These principal component coefficients represent the projected components of the original data along each principal component direction and serve as input features for the subsequent neural network model. Through principal component analysis, the originally high-dimensional airflow shear force spatial distribution data is compressed into low-dimensional principal component coefficient vectors, significantly reducing the data dimensionality while preserving the core features of the flow field.
[0037] (3) Training of neural network models In this embodiment, a neural network model is constructed and trained using first and second operating condition parameters as inputs and the corresponding principal component coefficient data as outputs to obtain a trained neural network surrogate model. The network structure of the neural network model includes an input layer, at least two hidden layers, and an output layer. The input layer receives externally input feature data, and its number of nodes matches the total dimension of the first and second operating condition parameters. The hidden layer, located between the input and output layers, performs feature extraction and nonlinear transformation on the input data. By setting at least two hidden layers, the neural network model's ability to fit complex nonlinear mapping relationships is enhanced. The neurons in the hidden layers are configured with activation functions to introduce nonlinear factors, enabling the neural network model to handle highly nonlinear relationships between the operating condition parameters and the principal component coefficient data. The output layer outputs the final prediction result of the neural network model, and its number of nodes matches the dimension of the principal component coefficient data.
[0038] The first and second operating condition parameters are input into the input layer of the neural network model, and the data undergoes forward propagation within the network. After receiving the operating condition parameters, the input layer passes them to the first hidden layer. The hidden layer performs a weighted summation operation on the received data and performs a nonlinear mapping through an activation function, passing the mapped result to the next hidden layer. After at least two layers of layer-by-layer nonlinear mapping, the data is finally passed to the output layer, which outputs the predicted principal component coefficients. These predicted principal component coefficients represent the neural network model's response to the input operating condition parameters based on the current weight parameters.
[0039] After obtaining the predicted principal component coefficients, the mean squared error (MSE) between the predicted and actual principal component coefficients is calculated, and a loss function is constructed based on this MSE. The actual principal component coefficients are those calculated through principal component analysis in the previous step. The MSE quantifies the degree of deviation between the predicted and actual values by calculating the mean of the sum of the squares of the differences. The loss function, built around the MSE, is used to evaluate the predictive performance of the neural network model under the current weight parameters. The smaller the value of the loss function, the smaller the deviation between the predicted and actual principal component coefficients, and the better the predictive performance of the neural network model.
[0040] Based on the constructed loss function, the backpropagation algorithm and the adaptive moment estimation optimizer are used to update the weight parameters of the neural network model. The backpropagation algorithm starts from the output layer, calculates the gradient for each weight parameter according to the loss function, and propagates the gradient information layer by layer from back to front along the network structure to obtain the gradient value for each weight parameter. The adaptive moment estimation optimizer combines momentum and adaptive learning rate adjustment mechanisms, dynamically adjusting the learning rate for each weight parameter using the first and second moment estimates of the gradient. This optimization method can accelerate the convergence speed in the early stages of training and maintain a stable parameter update step size in the later stages of training, helping the neural network model find optimal solutions in a complex parameter space. Based on the calculated gradients and the update step size determined by the adaptive moment estimation optimizer, the weight parameters and bias parameters of each layer in the neural network model are iteratively updated.
[0041] The aforementioned process of calculating the loss function via forward propagation, calculating the gradient via backpropagation, and updating the weight parameters is iteratively executed on the training set. During each iteration, the neural network model continuously adjusts its internal weight parameters, causing the value of the loss function to gradually decrease. When the value of the loss function stabilizes, or when the preset maximum number of iterations is reached, the loss function is considered to have converged, and iterative training stops, resulting in the trained neural network surrogate model. The trained neural network surrogate model establishes a nonlinear mapping relationship between the first and second condition parameters and the principal component coefficient data, enabling it to predict the corresponding principal component coefficient data based on the input condition parameters.
[0042] (4) Iterative optimization of operating parameters In this embodiment, the airflow shear force under actual flight conditions is determined as the target value. This target value is derived from CFD simulation results or flight test measurement data under actual flight conditions. Wind tunnel test parameters are used as decision variables, including parameters affecting the airflow shear force distribution such as incoming Mach number, angle of attack, Reynolds number, and sideslip angle. An initial population is randomly generated within the feasible region of the wind tunnel test parameters. The initial population consists of multiple individuals, each representing a combination of wind tunnel test parameter values. Individuals are encoded using real-number encoding, directly using each component of the wind tunnel test parameters as gene values. This encoding method avoids the precision loss caused by binary encoding and facilitates the handling of continuous parameter variables. The gene value of each individual is randomly generated within its corresponding parameter value range, ensuring the representativeness of the initial population distribution in the parameter space.
[0043] When evaluating the fitness of each individual in the population, the wind tunnel test parameters represented by the individual are input into a trained neural network surrogate model, which then outputs predicted principal component coefficient data. The predicted principal component coefficient data is then multiplied with the principal component eigenvectors to obtain a reconstructed one-dimensional vector of airflow shear force. This reconstructed one-dimensional vector of airflow shear force is mapped back to the aircraft surface grid according to a preset grid node order, yielding the reconstructed wind tunnel test airflow shear force. The error between the reconstructed wind tunnel test airflow shear force and the airflow shear force under actual flight conditions is calculated. The error is calculated by taking the mean of the sum of the squares of the differences between the shear force values at each corresponding grid node. This error is then monotonically decreased and used as the fitness value of the individual in the genetic algorithm. A smaller error indicates a higher fitness value, signifying a better equivalent airflow shear force effect for that individual.
[0044] Based on the fitness assessment results, a selection operation is performed on the population using a tournament selection method. Several individuals are randomly selected from the current population for comparison, and the individual with the highest fitness value is chosen as the parent. This selection process is repeated until a predetermined number of parent individuals are selected. A crossover operation is then performed on the selected parent individuals, exchanging some gene values between the two parent individuals to produce offspring. A mutation operation is then performed on the offspring individuals with a preset mutation probability, randomly perturbing some gene values within their range to generate a new generation of the population, which then enters the next iteration.
[0045] The aforementioned fitness assessment, selection, crossover, and mutation processes are iteratively executed. When the error between the reconstructed wind tunnel test airflow shear force and the airflow shear force under actual flight conditions is less than a preset error threshold, or when the preset maximum number of iterations is reached, the iteration stops. The wind tunnel test parameters corresponding to the individual with the highest fitness value in the current population are output as the optimal wind tunnel test parameters. These optimal wind tunnel test parameters represent the equivalent airflow shear force, indicating that under these wind tunnel test conditions, the wind tunnel test can reproduce an airflow shear force distribution similar to that of actual flight conditions.
[0046] Obtain surface shear force data measured through wind tunnel testing. This surface shear force data is obtained by measuring the shear force distribution on the surface of the aircraft model under optimal wind tunnel test parameters. The surface shear force data is compared with the equivalent airflow shear force predicted and reconstructed by a neural network surrogate model under optimal wind tunnel test parameters. The deviation between the two at each grid node is calculated, and the overall error is statistically analyzed.
[0047] If the error between the surface shear force data and the equivalent airflow shear force result is greater than or equal to a preset error threshold, it indicates a significant deviation between the prediction results of the neural network surrogate model and the actual wind tunnel test measurement results, requiring supplementary training data to improve the model's generalization ability. At this point, return to step S2 and add new CFD simulation sample points within the feasible region of the wind tunnel test parameters to expand the original dataset. The supplemented CFD simulation sample points are preferentially selected for intensive sampling in the parameter region near the current optimal solution to improve the fitting ability of the neural network surrogate model in that parameter region. Then, use the expanded original dataset to re-execute step S3 to update and train the neural network model, and repeat step S4 for iterative optimization and verification until the error between the surface shear force data and the equivalent airflow shear force result is less than the preset error threshold. Through this iterative verification and training update mechanism, the predictive ability of the neural network surrogate model can be gradually improved, ensuring that the final output equivalent airflow shear force result meets the requirements.
[0048] Please refer to Figure 3 It can be seen that within the target working condition parameter range of 1~12km in altitude and 0~15° in angle of attack, the shear force equivalent of the wind tunnel test site can be achieved well, and the average equivalent error is controlled within 30%.
[0049] Example 2 Please refer to Figure 4 This embodiment 2 provides an equivalent system of airflow shear force based on PCA and neural networks, including: The original dataset construction unit is used to obtain the first operating condition parameters and first airflow shear force data of multiple real flight operating condition samples, as well as the second operating condition parameters and second airflow shear force data of multiple wind tunnel test operating condition samples, to construct the original dataset; The principal component feature extraction unit is used to perform principal component analysis on the first airflow shear force data and the second airflow shear force data in the original dataset, extract the principal component feature vectors, and calculate the corresponding principal component coefficient data. The neural network model training unit is used to construct a neural network model, taking the first working condition parameters and the second working condition parameters as inputs and the corresponding principal component coefficient data as outputs for training, to obtain a trained neural network surrogate model. The operating condition parameter iterative optimization unit is used to take the airflow shear force of the real flight condition as the target value, input the wind tunnel test operating condition parameters as decision variables into the trained neural network surrogate model, and perform iterative optimization through a genetic algorithm. When the error between the reconstructed wind tunnel test airflow shear force and the airflow shear force of the real flight condition meets the preset error threshold, the optimal wind tunnel test operating condition parameters are output to complete the airflow shear force equivalence.
[0050] Example 3 This embodiment 3 also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement any step of a PCA and neural network-based airflow shear force equivalent method.
[0051] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0052] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.
[0053] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for equivalence of airflow shear force based on PCA and neural networks, characterized in that, include: S1. Obtain the first operating condition parameters and first airflow shear force data of multiple real flight operating condition samples, as well as the second operating condition parameters and second airflow shear force data of multiple wind tunnel test operating condition samples, and construct the original dataset; S2. Perform principal component analysis on the first airflow shear force data and the second airflow shear force data in the original dataset, extract the principal component feature vectors, and calculate the corresponding principal component coefficient data; S3. Construct a neural network model, using the first working condition parameters and the second working condition parameters as inputs and the corresponding principal component coefficient data as outputs for training, to obtain a trained neural network proxy model; S4. Using the airflow shear force under actual flight conditions as the target value, the wind tunnel test conditions parameters are input as decision variables into the trained neural network surrogate model. The model is iteratively optimized using a genetic algorithm. When the error between the reconstructed wind tunnel test airflow shear force and the actual flight conditions airflow shear force meets the preset error threshold, the optimal wind tunnel test conditions parameters are output to complete the airflow shear force equivalence.
2. The method for equivalent airflow shear force based on PCA and neural networks according to claim 1, characterized in that, In S1, the actual flight condition is a low-density, high-Mach number condition; the wind tunnel test condition is a high-density, low-Mach number condition; the condition parameters include Mach number, density, temperature, and pressure.
3. The method for equivalent airflow shear force based on PCA and neural networks according to claim 1, characterized in that, In step S2, the principal component feature vectors are extracted, and the corresponding principal component coefficient data are calculated, including an adaptive processing mechanism, specifically: The airflow shear force data in the original dataset are subjected to eigenvalue decomposition to obtain multiple eigenvalues arranged in descending order of numerical value; The decay rate is calculated based on two adjacent feature values in the feature value sequence. The feature value index that the decay rate is first less than a preset decay threshold is selected and used as the initial principal component dimension number. For the airflow shear force data in the original dataset, calculate the shear force gradient value at each grid node on the aircraft surface, and count the proportion of the number of grid nodes with shear force gradient values greater than a preset gradient threshold to the total number of grid nodes. Use this proportion as an indicator of flow field complexity. The initial principal component dimension number is corrected based on the flow field complexity index to obtain the target principal component dimension number; wherein, when the flow field complexity index is greater than a preset complexity threshold, a preset number of dimension compensation values are added to the initial principal component dimension number. The number of feature vectors of the target principal component dimension that are first extracted and sorted are used as the principal component feature vectors. The airflow shear force data in the original dataset is projected onto the space formed by the principal component feature vectors to calculate the principal component coefficient data.
4. The method for equivalent airflow shear force based on PCA and neural networks according to claim 3, characterized in that, The method for calculating the attenuation rate is as follows: In the formula, Indicates the first The attenuation rate at each location; This indicates the number of items sorted in descending order of numerical value. One eigenvalue; Indicates the first One eigenvalue; This represents the total number of the feature values; Indicates the summation index variable; Represents the sum of all eigenvalues; It is a positive integer, and .
5. The method for equivalent airflow shear force based on PCA and neural networks according to claim 3, characterized in that, The process of correcting the initial principal component dimension number based on the flow field complexity index to obtain the target principal component dimension number is as follows: Among them, the target cumulative variance contribution rate The calculation formula is: In the formula, This indicates the number of dimensions of the target principal components; Indicates the number of candidate dimensions, which is a positive integer; Represents the set of positive integers; This indicates the number of items sorted in descending order of numerical value. One eigenvalue; This represents the total number of the feature values; Indicates the summation index variable; The complexity index represents the flow field complexity. The dynamically changing cumulative variance contribution rate of the target; This represents the preset baseline cumulative variance contribution rate; This represents the preset maximum cumulative variance contribution rate; This represents the maximum reference value of the flow field complexity index; This means that the cumulative variance contribution rate is greater than or equal to Under the given conditions, take the smallest positive integer. As .
6. The method for equivalent airflow shear force based on PCA and neural networks according to claim 1, characterized in that, The neural network model in S3 includes an input layer, at least two hidden layers, and an output layer; The training process, which uses the first operating condition parameters and the second operating condition parameters as inputs and the corresponding principal component coefficient data as outputs, includes: The first operating condition parameter and the second operating condition parameter are input into the neural network model, and nonlinear mapping is performed through the hidden layer. The predicted principal component coefficient data are output by the output layer. Calculate the mean squared error between the predicted principal component coefficient data and the actual principal component coefficient data, and construct a loss function based on the mean squared error; The backpropagation algorithm and adaptive moment estimation optimizer are used to update the weight parameters of the neural network model according to the loss function until the loss function converges, thus obtaining the trained neural network surrogate model.
7. The method for equivalent airflow shear force based on PCA and neural networks according to claim 1, characterized in that, In step S4, the process of obtaining the reconstructed wind tunnel test airflow shear force includes: The predicted principal component coefficient data output by the trained neural network surrogate model is multiplied with the principal component feature vector to obtain the reconstructed one-dimensional vector of airflow shear force. The reconstructed one-dimensional vector of airflow shear force is mapped back to the aircraft surface grid according to the preset grid node order to obtain the reconstructed wind tunnel test airflow shear force.
8. The method for equivalent airflow shear force based on PCA and neural networks according to claim 1, characterized in that, S4 further includes: Obtain surface shear force data measured through wind tunnel testing; compare the surface shear force data with the equivalent airflow shear force output by the neural network model, and calculate the error between the two; If the error is greater than or equal to a preset error threshold, then return to step S2 to supplement CFD simulation data, and repeat steps S3 and S4 to update and train the neural network model until the error is less than the preset error threshold.
9. An equivalent system of airflow shear force based on PCA and neural networks, characterized in that, include: The original dataset construction unit is used to obtain the first operating condition parameters and first airflow shear force data of multiple real flight operating condition samples, as well as the second operating condition parameters and second airflow shear force data of multiple wind tunnel test operating condition samples, to construct the original dataset; The principal component feature extraction unit is used to perform principal component analysis on the first airflow shear force data and the second airflow shear force data in the original dataset, extract the principal component feature vectors, and calculate the corresponding principal component coefficient data. The neural network model training unit is used to construct a neural network model, taking the first working condition parameters and the second working condition parameters as inputs and the corresponding principal component coefficient data as outputs for training, to obtain a trained neural network surrogate model. The operating condition parameter iterative optimization unit is used to take the airflow shear force of the real flight condition as the target value, input the wind tunnel test operating condition parameters as decision variables into the trained neural network surrogate model, and perform iterative optimization through a genetic algorithm. When the error between the reconstructed wind tunnel test airflow shear force and the airflow shear force of the real flight condition meets the preset error threshold, the optimal wind tunnel test operating condition parameters are output to complete the airflow shear force equivalence.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor as described in any one of claims 1-8, which is a method for equivalent airflow shear force based on PCA and neural networks.
