Wind tunnel dynamics modeling method considering the blockage of test model

Abstract

The present invention discloses a wind tunnel dynamics modeling method considering the blockage of a test model, which relates to the field of aerospace aerodynamic test research, including: S1, establishing a wind tunnel NARX mathematical model to select key parameters in a wind tunnel database; S2, standardizing and labeling the key parameters to obtain labeled wind tunnel parameters, and extracting blockage data F from the test model digital model corresponding to the test vehicle; S3, standardizing the blockage data F to construct an autoencoder neural network for extracting blockage code Code; S4, splicing the wind tunnel parameters and Code as a training set, and obtaining the following wind tunnel dynamics model by constructing a neural network model based on GRU. The present invention provides a wind tunnel dynamics modeling method considering the blockage of a test model, which considers the blockage of the test model in the modeling process, can be applied to most low-speed wind tunnels and high-speed wind tunnels, and can more accurately predict the evolution of the flow field in a wind tunnel test.

Description

Technical Field

[0001] The present invention relates to the field of aerospace aerodynamic test research, and more specifically, to a wind tunnel dynamics modeling method taking into account the blockage of a test model. Background Art

[0002] When a wind tunnel is facing a test of a new model or under new working conditions, a large amount of pre-test air blowing debugging is required to ensure that the target flow field is established stably, accurately and safely. For high-energy-consuming equipment such as wind tunnels, the cost of each air blowing debugging is very expensive, and the debugging is heavily dependent on the experience of aerodynamic experts and has uncertainty. Therefore, establishing a wind tunnel dynamics model to replace air blowing debugging can effectively shorten the wind tunnel test operation time, reduce test gas consumption, improve wind tunnel test capabilities, and significantly reduce wind tunnel test costs. An effective wind tunnel dynamics modeling method is an important task in promoting the digital construction of wind tunnels.

[0003] The commonly used method for wind tunnel flow field dynamics modeling is the data-driven method, which usually uses deep learning algorithms to describe the wind tunnel operation mechanism. The key difficulties are mainly in two aspects. First, the real flow field in the wind tunnel test section has the characteristics of frequent disturbances and strong time variation. The data sampling points during the test are high, and the time point that the dynamic model needs to predict is very long. Therefore, the conventional modeling method is prone to error accumulation, resulting in the divergence of the flow field of the dynamic model evolution, and even time offset from the real flow field. Second, the neural network-based modeling method itself does not distinguish between the key parameters of the wind tunnel, but for the wind tunnel system, there are two parts: the key parameters of flow field control and the key parameters of flow field state. If these two parts of parameters are not distinguished, the operation mechanism of the wind tunnel cannot be accurately modeled. The existing wind tunnel dynamics modeling methods also have problems such as not considering the blockage of the test model, the single test condition, and the strict applicable conditions. Therefore, they are only applicable to specific conditions or when the blockage of the model can be ignored. However, under the conditions of large-scale test models, the blocking effect of the model cannot be ignored, and the blowing debugging of large-scale models is more difficult, and the test cost is higher. Summary of the invention

[0004] An object of the present invention is to solve at least the above problems and / or disadvantages and to provide at least the advantages which will be described hereinafter.

[0005] In order to achieve these objectives and other advantages of the present invention, a wind tunnel dynamics modeling method considering the blockage of a test model is provided, comprising:

[0006] S1. Establish the NARX mathematical model of the wind tunnel to select the key parameters in the wind tunnel database;

[0007] S2, standardizing and labeling the key parameters obtained in S1 to obtain labeled wind tunnel parameters, and extracting the blockage data F corresponding to the test model from the test model digital model corresponding to the test vehicle number;

[0008] S3, standardize the blockage data F to construct an autoencoder neural network, and extract the blockage code Code through transfer learning;

[0009] S4. The wind tunnel parameters and Code are concatenated as training sets to construct a neural network model based on GRU. The neural network model is trained in a supervised manner to obtain a wind tunnel dynamics model. The results of the dynamics model evolution prediction are denormalized to obtain a flow field state with real physical meaning:

[0010]

[0011] In the above formula, Y represents the flow field state with physical significance after anti-normalization, Y ′ represents the predicted output of the kinetic model, Min and Max Indicates the minimum and maximum values ​​when the wind tunnel flow field state is standardized.

[0012] Preferably, it also includes: S5, under the condition of given wind tunnel valve system parameters, setting the initial flow field state, gradually predicting the flow field, and feeding back the predicted state to the input of the dynamic model each time to achieve evolution prediction of the entire blowing test process.

[0013] Preferably, in S1, the wind tunnel NARX mathematical model is constructed in the following manner:

[0014] S100, based on the calculation of the flow field state and the wind tunnel control mechanism parameters, to obtain a mutual information value between the two;

[0015] S110, determining the estimated parameters of the neighboring mutual information value based on the cross-validation method and the permutation test method k , to balance the systematic error and statistical error in mutual information estimation;

[0016] S120, based on estimated parameters k Calculate the mutual information value of each variable of the flow field state-valve system;

[0017] S130. Extend the NARX model to multi-dimensional situations and describe all the flow field state variables of the test section by the following formula:

[0018]

[0019] In the above formula, Y tIndicates the flow field state of the test section that needs to be predicted in the wind tunnel. U T Indicates the valve system parameters that affect the flow field state, Y T represents the past value of the flow field state in the wind tunnel test section, represents the implicit function relationship of the wind tunnel flow field, N is the order of the flow field NARX model, t represents the wind tunnel operation time, T - N Indicates wind tunnel operation time t - N,t - N -1,… t In-range sequence data;

[0020] The key parameters include valve system parameters U T And the flow field state of the test section of the wind tunnel needs to be predicted.

[0021] Preferably, the estimated parameter k The way to obtain is:

[0022] S111. Randomly divide the data set of a vehicle wind tunnel test into K non-overlapping and approximately equal subsets, removing one subset at a time, and the remaining subsets constitute a new sample set Ω;

[0023] S112. Assume that the Mach number of the wind tunnel flow field state is the flow field output variable Y , valve system i variables as feature input X , find each parameter k ′, the mutual information value of the sample set Ω is calculated again after randomly permuting the Mach numbers;

[0024] S113, using the z-test method to calculate the test statistic to measure the difference between the mutual information values ​​of the two groups before and after the Mach number permutation;

[0025] S114. Calculate all values ​​of the test statistic of the wind tunnel flow field state and valve system parameters within the range of parameter k′=1~25, and take the largest corresponding value in the test statistic k ′ is the estimated parameter k The value of .

[0026] Preferably, in S2, the blockage degree data F The acquisition process includes:

[0027] S200, measure the length from the front end point of the test model to the rear end face to obtain the test model scale L ;

[0028] S210, defining the installation position of the test model in the wind tunnel based on the test model scale to obtain the applied position parameters L ′;

[0029] S220, along the fuselage axis of the test model, use n The test model is divided into equal parts by 2 cross sections, and the ratio of the area of ​​each cross section to the nozzle outlet area is taken as the overall envelope of the test model. The blockage data of the test model is F It can be expressed by the following formula:

[0030]

[0031] In the above formula, S i Indicates i The cross-sectional area of ​​the cross section, and i =1,2,… n , S is the cross-sectional area of ​​the nozzle outlet.

[0032] Preferably, in S3, the autoencoder neural network removes the decoder part of the trained autoencoder through unsupervised learning, and then obtains the Code by fixing the learnable parameters.

[0033] Preferably, in S4, the GRU-based neural network model includes an input layer, a GRU layer, a driving network layer, a fully connected layer and an output layer;

[0034] Among them, the driving network layer is used to distinguish the opening of the wind tunnel main pressure regulating valve, the opening of the pilot pressure regulating valve and the flow field state parameters, and the input valve system parameters of the driving network layer;

[0035] The input of the output layer is the valve system parameter features extracted by the driving network layer and the flow field state parameters output by the fully connected layer, and the output of the output layer is the flow field state obtained by linear combination.

[0036] The present invention has at least the following beneficial effects:

[0037] First, by systematically and theoretically analyzing the key parameters of the wind tunnel, the empirical selection of valve system parameters that affect the flow field state of the wind tunnel is avoided.

[0038] Secondly, a new definition of the blockage degree of the test model is proposed, which effectively enriches the test conditions that can be covered by the dynamic model and improves the scope of application of the dynamic modeling method.

[0039] Thirdly, a method to distinguish the wind tunnel flow field control parameters and flow field state is proposed. A driving network layer is designed based on the GRU neural network, which effectively improves the modeling accuracy and proposes an evolutionary prediction of the dynamic model evolution.

[0040] Other advantages, objectives and features of the present invention will be embodied in part through the following description, and in part will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of the process of the GRU-based dynamic modeling method in the present invention;

[0042] Figure 2 It is a schematic diagram of calculation results of the test statistic of the Z-test method of Mach number-main pressure regulating valve opening adopted in the present invention;

[0043] Figure 3 A schematic diagram of extracting the blockage degree of a wind tunnel model using the method of the present invention;

[0044] Figure 4 It is a neural network structure diagram based on GRU in the present invention;

[0045] Figure 5 In order to use the method of the present invention to perform dynamic modeling on a wind tunnel, a schematic diagram of the comparison between the predicted value of the Mach number and the actual value in the flow field evolution results of the test section under the test condition of the target Mach number of 0.6 is shown;

[0046] Figure 6 In order to use the method of the present invention to perform dynamic modeling on a wind tunnel, a schematic diagram of the comparison between the predicted total pressure value and the actual value in the flow field evolution results of the test section under the test condition of a target Mach number of 0.6 is shown;

[0047] Figure 7 In order to use the method of the present invention to perform dynamic modeling on a wind tunnel, a schematic diagram of the comparison between the predicted value and the actual value of the nozzle static pressure in the test section flow field evolution results under the test condition of a target Mach number of 0.6 is shown;

[0048] Figure 8 In order to use the method of the present invention to perform dynamic modeling on a wind tunnel, a schematic diagram is shown comparing the predicted value and the actual value of the static pressure of the test cabin in the flow field evolution results of the test section under the test condition with a target Mach number of 0.6. DETAILED DESCRIPTION

[0049] The present invention is further described in detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0050] The present invention introduces cutting-edge algorithms in the field of deep learning into the field of wind tunnel testing, and proposes a wind tunnel dynamics modeling method based on GRU, which is suitable for most low-speed wind tunnels and high-speed wind tunnels, and can more accurately predict the evolution process of the flow field in wind tunnel tests.

[0051] (1) The estimated parameter k of the neighbor mutual information estimation was determined based on the cross-validation method and the permutation test method. The nonlinear auto-regressive (NARX) mathematical model of the wind tunnel was established using the neighbor mutual information estimation method, and the key parameters in the wind tunnel database were selected.

[0052] S100, calculating the mutual information value between the flow field state and the wind tunnel control mechanism parameters.

[0053] S110, use the permutation test method and cross-validation method to reduce the approximate bias, determine the optimal value of the estimation parameter k of the neighboring mutual information, and balance the systematic error and statistical error in the mutual information estimation. The number of samples generated during a wind tunnel test is approximately 300 to 3000. The data of a test needs to be divided and combined into different data sets, that is, the cross-validation method is used for permutation test.

[0054] Specifically, the selection of the estimated parameter k includes:

[0055] S111. Randomly divide the data set of a train wind tunnel test (only the start-up to shutdown phase of the test) into K non-overlapping and approximately equal subsets, remove one subset each time, and the remaining subsets constitute a new sample set Ω.

[0056] S112, Assuming that the Mach number of the wind tunnel flow field state is the flow field output variable Y, the valve system i variables as feature input X, find each parameter k ′, and then recalculate the mutual information value of the sample set Ω after randomly permuting the Mach numbers.

[0057] S113. The z-test method is used to calculate the test statistic to measure the difference between the mutual information values ​​of the two groups before and after the Mach number substitution.

[0058] S114. Calculate all values ​​of the test statistics of the wind tunnel flow field state and the valve system parameters in the range of parameter k′=1~25. The k′ value corresponding to the maximum test statistic is the best value of the estimated parameter k.

[0059] S120, calculating the mutual information value of the flow field state-valve system variables under the optimal estimated parameter k value.

[0060] S121. Determine a mutual information threshold, and select valve system variables under the condition that the mutual information is greater than the threshold.

[0061] S130. Extend the NARX model to multi-dimensional situations and use a NARX mathematical model to describe all the flow field state variables of the test section, as shown in the following formula:

[0062]

[0063] In the formula, Y t Indicates the flow field state of the test section that needs to be predicted in the wind tunnel. U T Indicates the valve system parameters that affect the flow field state, Y T represents the past value of the flow field state in the wind tunnel test section, represents the implicit function relationship of the wind tunnel flow field, N is the order of the flow field NARX model, t represents the wind tunnel operation time, T - N Indicates wind tunnel operation time t - N,t - N -1,… t Range sequence data.

[0064] S131. In S121, the valve system variable whose mutual information is greater than the threshold is U T The specific parameters of the wind tunnel, the test section flow field state and valve system parameters that need to be predicted U T Together they constitute the key parameters of the wind tunnel.

[0065] (2) The key parameters of the wind tunnel are standardized and labeled to the maximum and minimum values, and the model blockage parameters are extracted from the test model digital model corresponding to the test vehicle.

[0066] S200, the maximum and minimum values ​​of key parameters of the wind tunnel are standardized.

[0067] S210, labeling the standardized key wind tunnel parameters, where the data labels of the key wind tunnel parameters in the range of running time 0 to t are the wind tunnel flow field states at time t+1, from the issuance of the wind tunnel test start-up command to N times before the end of the wind tunnel test, where N is the order of the wind tunnel NARX model.

[0068] S220, the test model has two main effects on the flow field. One is the blocking effect of the model itself and the size of its cross section on the flow field. The other is the effect of the installation position of the model in the test chamber. The specific method for extracting the blocking degree from the test model digital model is as follows:

[0069] S221. Measure the total length of the digital model of the test model, that is, the length from the front end point to the rear end face of the test model, in millimeters.

[0070] S222. Define the installation position of the test model in the wind tunnel. Due to the difference in the scale of the test model, the model has a relative position in the wind tunnel test chamber, and extract the position parameters.

[0071] S223, extract the overall envelope of the test model. O x The axis refers to the direction of the fuselage axis. The test model is divided into 10 equal sections. The ratio of the area of ​​each section to the nozzle outlet area is used as the overall envelope of the test model. The number of equally divided sections can be increased or decreased depending on the specific application.

[0072] S224, so use a set of parameters F To define the blocking degree of the test model, as shown in the following formula:

[0073]

[0074] in, L Indicates the overall length of the test model in millimeters. L ′ represents the installation position parameter of the test model, in millimeters. S i It means that the test model is divided into 10 equal cross-sectional areas along the Z axis of the body coordinate system from the front end to the tail end, in square meters. i =1,2,… 10 . S is the cross-sectional area of ​​the nozzle outlet, in square meters.

[0075] (3) The blockage data of the experimental model is trained based on the autoencoder, and transfer learning is performed to extract the blockage code Code as part of the input of the neural network model based on the GRU neural network.

[0076] S300, the blocking degree data F The data is standardized to the maximum and minimum values, and an autoencoder neural network is constructed.

[0077] S310, unsupervised learning is performed on the autoencoder neural network, that is, the input and output of the autoencoder network are both standardized blocking degree data F .

[0078] S320, remove the decoder part of the trained autoencoder, fix the learnable parameters of the encoder, and convert the blocking degree data F Input to the autoencoder to get the Code.

[0079] (4) The labeled wind tunnel parameters and codes constitute a training set, and a neural network model based on the gated recurrent unit (GRU) is constructed. The wind tunnel dynamics model is obtained through training.

[0080] S400 and Code are concatenated with the labeled wind tunnel parameters in the feature dimension to form the input of the neural network model.

[0081] S410, constructing a neural network model based on GRU, performing supervised training on the neural network, and the neural network model after reaching the training times is a dynamic model of the wind tunnel.

[0082] (5) Kinetic model verification and evaluation

[0083] The dynamic modeling of the wind tunnel flow field is to set the initial flow field state under the condition of given wind tunnel valve system parameters, and gradually predict the flow field. Each time, the predicted state is fed back to the input of the dynamic model to achieve the evolution prediction of the entire blowing test process.

[0084] S500: Given valve system parameters under the target train number, the initial flow field state of the wind tunnel is set to a common indoor flow field, and the valve system parameters and the initial flow field state are input into the dynamic model for single-step prediction.

[0085] S510, feeding back the single-step prediction result as the flow field state part in the dynamic model input, forming a new input for recursive prediction.

[0086] S520, the result of the kinetic evolution prediction is denormalized, and the calculation formula is as follows:

[0087]

[0088] In the above formula, Y represents the flow field state with physical significance after anti-normalization, Y ′ represents the predicted output of the kinetic model, Min and Max Indicates the minimum and maximum values ​​when the wind tunnel flow field state is standardized.

[0089] Example:

[0090] Figure 1 The following is a flow chart of a wind tunnel dynamics modeling method based on GRU according to an embodiment of the present application. The modeling method includes:

[0091] (1) The estimated parameter k of the neighbor mutual information was determined based on the cross-validation method and the permutation test method. The wind tunnel NARX data model was established using the neighbor mutual information estimation method, and the key parameters in the wind tunnel database were selected.

[0092] The key parameters in the wind tunnel database include valve system parameters and flow field state parameters. The flow field state parameters are variables that need to be detected and controlled when conducting tests in specific application wind tunnels. The flow field state parameters of the wind tunnel in this example are Mach number, total pressure, nozzle outlet static pressure and test chamber static pressure. The valve mechanisms of the wind tunnel are the main pressure regulating valve and the pilot pressure regulating valve. It is necessary to select the specific parameters that have the greatest impact on the flow field state of the wind tunnel from these two control mechanisms.

[0093] S100, calculating the mutual information value between the flow field state and the wind tunnel control mechanism parameters, as shown in the following formula:

[0094]

[0095] In the formula, the estimated parameter k represents the sample point x i The kth nearest neighbor of , N represents the number of samples of wind tunnel control mechanism parameter X, Y represents the wind tunnel flow field state, n x ( i ) represents the sample point in the sample set of X x i is the center of the circle, and the estimated parameter k has a neighbor distance of all neighbors within the circle with a diameter; n y ( i ) indicates that the sample point in the sample set of Y y i is the center of the circle, Ψ Represents the Digamma function, which is the derivative of the logarithm of the Gamma function. The definition of the Gamma function and Ψ The calculation recursion formula is as follows:

[0096]

[0097] In the above formula, x is the function variable, γ is the Euler constant, and its approximate value is 0.5775216. r is the independent variable of integration.

[0098] S110, the choice of parameter k has a great influence on mutual information estimation. The permutation test method and cross-validation method are used to reduce the approximate deviation, determine the optimal value of the k-neighbor mutual information estimation parameter k, and balance the systematic error and statistical error in mutual information estimation. The number of samples generated during a wind tunnel test is approximately 300~3000. The data of a test needs to be divided and combined into different data sets, that is, the cross-validation method is used for permutation test.

[0099] Specifically, the estimated parameter k is obtained by:

[0100] S111. Randomly divide the data set Ω of a train wind tunnel test (only the test start-up to shutdown phase) into K non-overlapping and approximately equal subsets Ω1, Ω2, …Ω K , each time removing a subset Ω i , the remaining subsets constitute a new sample set Ω′=Ω / Ω i .

[0101] S112, Assuming that the Mach number of the wind tunnel flow field state is the flow field output variable Y, the valve system i variables as feature input X, find the mutual information value of sample set Ω′ under each parameter k′ I k (X,Y), randomly permutate the Mach numbers and calculate the mutual information again I k (X,Y′).

[0102] S113. The z-test method is used to calculate the test statistic to measure the difference between the mutual information values ​​of the two groups before and after the Mach number replacement. The calculation method is shown in the following formula:

[0103]

[0104] Among them, Z i,k is the test statistic, μ and σ ' 2 Respectively represent 10-fold cross validation I k The mean and variance of (X,Y), μ 'and σ ' 2 Respectively I k The mean and variance of (X,Y′). i,k Maximum time mutual information I k (X,Y) and I k There is a significant difference between (X, Y′), and the estimated parameter k is the optimal value.

[0105] S114. Calculate the test statistic Z of the wind tunnel flow field state and valve system parameters i,k For all values ​​of parameter k=1~25, Z i,k The corresponding maximum k′ value is the optimal value of the estimated parameter k.

[0106] Figure 2 The z-test method for Mach number-main pressure regulating valve opening is used, and the optimal value of parameter k is 5. The test statistics of other flow field states-valve system parameters are calculated using the same method. The optimal value of parameter k is as follows: Figure 2 As shown, it is the horizontal axis value corresponding to the highest point in the histogram.

[0107] S120, calculate the mutual information value of each variable of the flow field state-valve system under the optimal parameter k value. The wind tunnel main pressure regulating valve and the inlet pressure regulating valve of this embodiment respectively contain 3 variables, and the variables related to the main pressure regulating valve are respectively recorded as X M1 , X M2 , XM3 , the variable related to the pressure regulating valve is recorded as X R1 , X R2 , X R3 .

[0108] S121, determine that the mutual information threshold is 1, and the valve system variable under the condition that the mutual information is greater than the threshold is X M1 and X R1 .

[0109] Table 1 is a summary of the mutual approximate estimates of Mach number, total pressure, nozzle outlet static pressure, test chamber static pressure, and main pressure regulating valve related variables / pilot pressure regulating valve related variables. M1 , X M2 , X M3 The mutual trust values ​​are in the 1st to 3rd columns of the 1st row of Table 1, where the Mach number and X M1 The mutual trust value of is significantly greater than that of the other two main valve variables, indicating that the Mach number of the wind tunnel flow field is closely related to X M1 The total pressure, nozzle outlet static pressure and test chamber static pressure are closely related to X M1 and X R1 The mutual trust value is large, so the specific parameter of the main pressure regulating valve and the pilot pressure regulating valve that has the greatest influence on the wind tunnel flow field state is X M1 and X R1 , corresponding to the opening variables of the main pressure regulating valve and the pilot pressure regulating valve respectively.

[0110] Table 1

[0111]

[0112] S130. Extend the NARX model to multi-dimensional situations and use a NARX mathematical model to describe all the flow field state variables of the test section, as shown in the following formula:

[0113]

[0114] In the formula, Y t Indicates the flow field state of the test section that needs to be predicted in the wind tunnel: Mach number, total pressure, nozzle exit static pressure and test chamber static pressure, U T Indicates valve system parameters that affect the flow field state: main pressure regulating valve opening, pilot pressure regulating valve opening, represents the implicit function relationship of the wind tunnel flow field, N is the order of the flow field NARX model, t represents the wind tunnel operation time, T - N Indicates wind tunnel operation time t - N,t - N -1,… t Range sequence data.

[0115] S131, Mach number, total pressure, nozzle outlet static pressure, test chamber static pressure, main pressure regulating valve opening, and pilot pressure regulating valve opening together constitute the key parameters of the wind tunnel.

[0116] (2) The key parameters of the wind tunnel are standardized and labeled to the maximum and minimum values, and the model blockage parameters are extracted from the test model digital model corresponding to the test vehicle.

[0117] S200, wind tunnel key parameters are standardized to the maximum and minimum values, as shown in the following formula:

[0118]

[0119] Where min(X) and max(X) represent the minimum and maximum values ​​of the key wind tunnel parameter X, respectively.

[0120] S210, labeling the standardized data, as shown in the following formula:

[0121] in, u 1 indicates the main pressure regulating valve opening, u 2 indicates the opening of the pressure regulating valve. P 0 indicates the total pressure of the test chamber. P s1 Indicates the static pressure of the test chamber, P s2 represents the static pressure at the nozzle outlet, Ma represents the Mach number of the test chamber, N represents the order of the wind tunnel NARX model, that is, the sequence length of the key flow field parameters, t represents the wind tunnel operation time, and --- represents the label correspondence. Starting from time point 0, every N Time point, timestamp data The corresponding label is N +Flow state at 1 time point , and loop in sequence until all the data of the entire train are labeled.

[0122] S220, the test model has two main effects on the flow field. One is the blocking effect of the model itself and the size of its cross section on the flow field. The other is the effect of the installation position of the model in the test chamber. The specific method for extracting the blocking degree from the test model digital model is as follows:

[0123] S221. Measure the total length of the digital model of the test model, that is, the length from the front end point to the rear end face of the test model, in millimeters.

[0124] S222. Define the installation position of the test model in the wind tunnel. Due to the difference in the scale of the test model, the model has a relative position in the wind tunnel test chamber, and extract the position parameters.

[0125] S223, extract the overall envelope of the test model. O x The axis refers to the direction of the fuselage axis. The test model is divided into 10 equal sections. The ratio of the area of ​​each section to the nozzle outlet area is used as the overall envelope of the test model. The number of equally divided sections can be increased or decreased depending on the specific application.

[0126] S224, so use a set of parameters F To define the blocking degree of the test model, as shown in the following formula:

[0127]

[0128] in, L Indicates the overall length of the test model in millimeters. L ′ represents the installation position parameter of the test model, in millimeters. S i It means that the test model is divided into 10 equal cross-sectional areas along the Z axis of the body coordinate system from the front end to the tail end, in square meters. i =1,2,… 10 . S is the cross-sectional area of ​​the nozzle outlet, in square meters.

[0129] Figure 3 This is a specific method for extracting the blockage degree in a wind tunnel used in this embodiment. L To measure the total length from the front end of the test model to the rear end of the model, this embodiment uses the installation position parameters of the test model in the wind tunnel L′ There are three items, namely: the vertical distance L1 from the front end of the test model to the lower end of the nozzle center A, the horizontal distance from the front end of the test model to the nozzle center L 2. The horizontal distance L3 from the front end of the collector B to the nozzle outlet section. The test model is divided into 10 sections along the technical axis ( Figure 3 The test model is divided into four sections for demonstration purposes). The area of ​​each section is measured. S i .

[0130] (3) The blocking degree data of the experimental model is trained based on the autoencoder, and transfer learning is performed to extract the Code as part of the input of the neural network model based on the GRU neural network.

[0131] S300, the blockage degree data F data is normalized to the maximum and minimum values, and an autoencoder neural network is constructed. The encoder and decoder structures of the autoencoder are consistent, both of which are set to a multi-layer perceptron structure (MLP). The number of neurons in the middle layer of the MLP is set to 256, and the encoding length of the encoder blockage degree is N, which is consistent with the length of the key flow field parameter sequence labeled in 210.

[0132] S310, unsupervised learning is performed on the autoencoder neural network, that is, the input and output of the autoencoder network are both standardized blocking degree data F The optimizer for autoencoder training is set to the adaptive moment estimation method (Adam), and the learning rate (lr) is set to 0.005.

[0133] S320, remove the decoder part of the trained autoencoder, fix the learnable parameters of the encoder, and input the blocking degree data F into the autoencoder to obtain Code.

[0134] (4) The labeled wind tunnel parameters and codes constitute a training set, and a GRU-based neural network model is constructed. The wind tunnel dynamics model is obtained through training.

[0135] S400 and Code are concatenated with the labeled wind tunnel parameters in the feature dimension to form the input of the neural network model. As shown in the following formula:

[0136] In the formula, X T represents the neural network model input, Fcode Indicates the encoder output Code, Represents the sequence data from t-10 to t. The other parameters are derived from S210 and have the same meaning.

[0137] S410, constructing a neural network model based on GRU, performing supervised training on the neural network, and the neural network model after reaching the training times is a dynamic model of the wind tunnel.

[0138] Figure 4 The specific structure of the GRU-based neural network model designed for this embodiment. The neural network includes several modules: input layer, GRU layer, driving network layer, fully connected layer and output layer. The input shape of the neural network model is a 3D array, and the output flow field state is Mach number, total pressure, nozzle outlet static pressure and test chamber static pressure.

[0139] In S400 , the input dimension of the GRU-based neural network model is [tN, 10, 7], the model input layer is set to 10, the GRU layer is set to 256, and the output layer is set to 128.

[0140] The driving network layer is used to distinguish the opening of the wind tunnel main pressure regulating valve, the opening of the pilot pressure regulating valve and the flow field state parameters. The valve system parameter part of the input data is input to the driving network layer. Finally, the valve system parameter features extracted by the driving network and the flow field state parameter features output by the fully connected layer are input to the output layer together. The output layer is a simple linear layer, which is used to linearly combine the two parts of the features to obtain the flow field state.

[0141] The driving network uses a two-layer network structure, the middle layer of the network has a Sigmoid activation function, the neurons between the layers are connected in a fully connected manner, the number of neurons in the first layer of the driving network is set to 256, and the number of neurons in the second layer is set to 128. The opening of the main pressure regulating valve and the opening of the pilot pressure regulating valve act again with the output of the fully connected layer after passing through the driving network, which increases the importance of the control parameters of the wind tunnel valve system and plays a driving role for the entire dynamic model, which is in line with the operation process of the real wind tunnel test.

[0142] (5) Kinetic model verification and evaluation

[0143] The dynamic modeling of the wind tunnel flow field is to set the initial flow field state under the condition of given wind tunnel valve system parameters, and gradually predict the flow field. Each time, the predicted state is fed back to the input of the dynamic model to achieve the evolution prediction of the entire blowing test process.

[0144] S500: Given valve system parameters under the target train number, the initial flow field state of the wind tunnel is set to a common indoor flow field, and the valve system parameters and the initial flow field state are input into the dynamic model for single-step prediction.

[0145] The input of the network is in the form of a 3D array. Each predicted flow field state is fed back to the last node of the input time dimension to form a new input (such as Figure 4 The feedback dotted line in the figure shows the evolution of the main pressure regulating valve opening and the pilot pressure regulating valve opening to achieve the purpose of dynamic modeling and prediction.

[0146] S510, feeding back the single-step prediction result as the flow field state part in the dynamic model input, forming a new input for recursive prediction.

[0147] S520, the result of the kinetic evolution prediction is denormalized, and the calculation formula is shown as follows:

[0148]

[0149] In the formula, Y represents the flow field state with physical significance after anti-normalization, Y ′ represents the predicted output of the kinetic model, Min and Max Indicates the minimum and maximum values ​​when the wind tunnel flow field state is standardized.

[0150] Table 2 Flow field prediction error, and Figure 5-Figure 8 , which is the prediction effect of the kinetic model trained according to this scheme.

[0151] Table 2

[0152]

[0153] Specifically, Figure 5-Figure 8 The solid line in the middle is the actual flow field state, and the dotted line is the flow field state predicted by the dynamic model. The actual flow field is very close to the flow field predicted by the dynamic model, which demonstrates the effectiveness of this scheme.

[0154] Table 2 shows the mean square error (MSE) and maximum absolute error (MAE) between the calculated real flow field and the predicted flow field. The calculation formula is shown as follows:

[0155]

[0156] y Represents the real flow field state, y′ Indicates the flow field state predicted by the dynamic model. Table 2 shows the prediction errors of the dynamic model under the condition of Mach number 0.47 for test model 1 and Mach number 0.6 for test model 2. The MSE value of the Mach number is around 0.007, and the maximum absolute deviation of the prediction is less than 0.08. The prediction accuracy of the dynamic model is very high.

[0157] The above solution is only an illustration of a preferred embodiment, but is not limited thereto. When implementing the present invention, appropriate replacement and / or modification can be performed according to user needs.

[0158] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and the embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily realized. Therefore, without departing from the general concept defined by the claims and equivalent scope, the present invention is not limited to the specific details and the illustrations shown and described here.

Claims

1. A wind tunnel dynamics modeling method considering the blockage of the test model, characterized in that: include: S1. Establish the NARX mathematical model of the wind tunnel to select the key parameters in the wind tunnel database; S2, standardizing and labeling the key parameters obtained in S1 to obtain labeled wind tunnel parameters, and extracting the blockage data F corresponding to the test model from the test model digital model corresponding to the test vehicle number; S3, standardize the blockage data F to construct an autoencoder neural network, and extract the blockage code Code through transfer learning; S4. The wind tunnel parameters and the code are concatenated as a training set to construct a neural network model based on the gated recurrent unit GRU. The neural network model is supervised for training to obtain the wind tunnel dynamics model. The results of the dynamics model evolution prediction are denormalized to obtain a flow field state with real physical meaning: In the above formula, Y represents the flow field state with physical significance after anti-normalization, Y ′ represents the predicted output of the kinetic model, Min and Max Indicates the minimum and maximum values ​​when the wind tunnel flow field state is standardized.

2. The wind tunnel dynamics modeling method considering the blockage of the test model according to claim 1, characterized in that: Also includes: S5. Under the condition of given wind tunnel valve system parameters, the initial flow field state is set, and the flow field is predicted step by step. Each time, the predicted state is fed back to the input of the dynamic model to achieve the evolution prediction of the entire blowing test process.

3. The wind tunnel dynamics modeling method considering the blockage of the test model according to claim 1, characterized in that: In S1, the wind tunnel NARX mathematical model is constructed as follows: S100, based on the calculation of the flow field state and the wind tunnel control mechanism parameters, to obtain a mutual information value between the two; S110, determining the estimated parameters of the neighboring mutual information value based on the cross-validation method and the permutation test method k , to balance the systematic error and statistical error in mutual information estimation; S120, based on estimated parameters k Calculate the mutual information value of each variable of the flow field state-valve system; S130. Extend the NARX model to multi-dimensional situations and describe all the flow field state variables of the test section by the following formula: In the above formula, Y t Indicates the flow field state of the test section that needs to be predicted in the wind tunnel. U T Indicates the valve system parameters that affect the flow field state, Y T represents the past value of the flow field state in the wind tunnel test section, represents the implicit function relationship of the wind tunnel flow field, N is the order of the flow field NARX model, t represents the wind tunnel operation time, T - N Indicates wind tunnel operation time t - N,t - N -1,… t In-range sequence data; The key parameters include valve system parameters U T And the flow field state of the test section of the wind tunnel needs to be predicted.

4. The wind tunnel dynamics modeling method considering the blockage of the test model according to claim 3 is characterized in that: The estimated parameters k The way to obtain is: S111. Randomly divide the data set of a vehicle wind tunnel test into K non-overlapping and approximately equal subsets, removing one subset at a time, and the remaining subsets constitute a new sample set Ω; S112. Assume that the Mach number of the wind tunnel flow field state is the flow field output variable Y , valve system i variables as feature input X , find each parameter k ′, the mutual information value of the sample set Ω is calculated again after randomly permuting the Mach numbers; S113, using the z-test method to calculate the test statistic to measure the difference between the mutual information values ​​of the two groups before and after the Mach number permutation; S114. Calculate all values ​​of the test statistic of the wind tunnel flow field state and valve system parameters within the range of parameter k′=1~25, and take the largest corresponding value in the test statistic k ′ is the estimated parameter k The value of .

5. The wind tunnel dynamics modeling method considering the blockage of the test model according to claim 1, characterized in that: In S2, the blocking degree data F The acquisition process includes: S200, measure the length from the front end point of the test model to the rear end face to obtain the test model scale L ; S210, defining the installation position of the test model in the wind tunnel based on the test model scale to obtain the applied position parameters L ′; S220, along the fuselage axis of the test model, use n The test model is divided into equal parts by 2 cross sections, and the ratio of the area of ​​each cross section to the nozzle outlet area is taken as the overall envelope of the test model. The blockage data of the test model is F It can be expressed by the following formula: In the above formula, S i Indicates i The cross-sectional area of ​​the cross section, and i =1,2,… n , S is the cross-sectional area of ​​the nozzle outlet.

6. The wind tunnel dynamics modeling method considering the blockage of the test model according to claim 1, characterized in that: In S3, the autoencoder neural network removes the decoder part of the trained autoencoder through unsupervised learning, and then obtains the Code by fixing the learnable parameters.

7. The wind tunnel dynamics modeling method considering the blockage of the test model according to claim 1, characterized in that: In S4, the GRU-based neural network model includes an input layer, a GRU layer, a driving network layer, a fully connected layer, and an output layer; Among them, the driving network layer is used to distinguish the opening of the wind tunnel main pressure regulating valve, the opening of the pilot pressure regulating valve and the flow field state parameters, and the input valve system parameters of the driving network layer; The input of the output layer is the valve system parameter features extracted by the driving network layer and the flow field state parameters output by the fully connected layer, and the output of the output layer is the flow field state obtained by linear combination.

Images