Parameter correction method for cascade wind tunnel test conditions based on deep neural network

Abstract

The present invention relates to a method for correcting working condition parameters of a cascade wind tunnel experiment based on a deep neural network, and belongs to the field of aerospace experimental measurement and artificial intelligence. Floating working condition parameters, constructing multiple sets of incoming flow parameters and performing numerical calculations of the cascade flow field; using the normalized working condition parameters after floating as input, and using the corresponding physical quantities of the measuring point positions in the front 50% of the blade chord length of the numerical flow field as output, to construct a deep neural network; using the mean square error of the deep neural network prediction results and the experimental measurement results as the loss function, using the automatic differentiation algorithm to backpropagate the loss function gradient to obtain the gradient of the mean square error with respect to the working condition parameters; using the gradient optimization algorithm to correct the working condition parameters. The present invention can effectively correct the incoming flow boundary conditions of the cascade wind tunnel experiment, can effectively overcome the separation zone prediction deviation generated by the turbulence model, does not require the special selection of the turbulence model, and is particularly suitable for the reconstruction and inversion of the experimental flow field using the flow field numerical simulation method.

Description

Technical Field

[0001] The invention belongs to the fields of aerospace experimental measurement and artificial intelligence, and relates to a method for correcting working condition parameters of a cascade wind tunnel experiment. Background Art

[0002] Cascade wind tunnel tests are a common method for measuring the performance of turbomachinery blade profiles. Operating parameters primarily refer to the incoming flow angle of attack and the incoming flow Mach number. The angle of attack is typically determined by rotating the cascade mounting dial to the appropriate angle. The incoming flow Mach number is determined by adjusting the wind tunnel air supply valve or fan speed so that the Mach number sensor at the front of the test piece displays a given Mach number. However, during actual wind tunnel tests, the airflow in the wind tunnel is affected by the tunnel walls, and the airflow experiences losses and deflections after passing through the Mach number sensor. The Mach number and angle of attack of the airflow actually hitting the cascade area are not the given operating parameters. Operating deviations are typically more pronounced at high angles of attack and Mach numbers, seriously impacting the effectiveness of cascade wind tunnel tests.

[0003] The existing method for correcting operating condition parameters uses the ensemble Kalman filter method. However, the ensemble Kalman filter method is a quasi-linear method that cannot capture the nonlinear information of the flow process in the wind tunnel. Its correction accuracy and scope of application are limited. The neural network has strong nonlinear fitting capabilities, can effectively capture nonlinear information, and has high modeling accuracy.

[0004] Therefore, there is currently a lack of an accurate and reliable method for correcting operating parameters in the field of cascade wind tunnel experiments. Summary of the Invention

[0005] The technical problems to be solved by the present invention are:

[0006] In order to avoid the deficiency of the existing technology that the ensemble Kalman filter method cannot capture the nonlinear information of the flow process in the wind tunnel, the present invention provides a method for correcting the working condition parameters of the blade wind tunnel experiment based on flow field numerical simulation and deep neural network.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0008] A method for correcting parameters of cascade wind tunnel test conditions based on flow field numerical simulation and deep neural network is characterized by the following steps:

[0009] Step 1: Floating experimental working condition parameters, constructing multiple groups of incoming flow parameter samples, and performing flow field numerical simulation for each group of parameters;

[0010] Step 2: Construct a neural network that describes the incoming flow parameters and flow field parameters. The input parameters of the neural network are the normalized incoming flow parameter samples in step 1, and the output parameters are the corresponding physical quantities of the measurement position in the front 50% of the blade chord length in the flow field calculated for each normalized sample.

[0011] Step 3: Use the experimentally measured operating parameters as the initial values; use the gradient optimization method to modify the incoming flow parameters and update the flow field input parameters:

[0012] Step 4: Use the flow field input parameters obtained from the current calculation as the corrected operating parameters of the modified blade cascade for the experimental working condition, and use the corrected incoming flow Mach number and incoming flow angle of attack to perform numerical calculation of the flow field. If the mean square error between the flow parameters of the measuring point in the numerical flow field and the experimental measurement data of the first 50% chord length area is less than the specified threshold, then the corrected parameters can be used as the final correction result; if the mean square error is greater than the specified threshold, then the numerical calculation data of this flow field is used as the training set for adding to the deep neural network, and repeat steps 1 to 4.

[0013] A further technical solution of the present invention: Step 1 is as follows: a floating range is set according to the control accuracy of the incoming flow parameters, a Latin hypercube method is used for random sampling, a sample of the incoming flow parameters is constructed, and a cascade flow field numerical calculation is performed on each sample in the turbulence model.

[0014] A further technical solution of the present invention: the corresponding physical quantities described in step 2 include the incoming flow Mach number, pressure coefficient, and static pressure.

[0015] A further technical solution of the present invention: Step 2 is as follows:

[0016] Step 2-1: The input parameters of the neural network are the normalized incoming flow parameters of each sample. For the incoming flow Mach number Ma in and the angle of attack α, the normalized formula is as follows:

[0017]

[0018]

[0019] Where, α max , α min 、Ma in，max 、Ma in，min , respectively represent the maximum and minimum values ​​of the incoming flow angle of attack and the incoming flow Mach number in the training data;

[0020] Step 2-2: The output parameter of the neural network is the physical quantity of the blade measurement point position corresponding to the first 50% chord length in the flow field calculation result of each sample after normalization. For the physical quantity of the jth measurement point The normalization formula is as follows:

[0021]

[0022] Where j = 1, 2, ..., M, M represents the number of blade measurement points corresponding to the first 50% of the chord length during the experiment. are the minimum and maximum values ​​of the physical quantities at the jth measuring point in the sample flow field;

[0023] Step 2-3: Construct a deep neural network DNN. The deep neural network DNN includes an input layer, several hidden layers and an output layer. The mathematical expression is as follows:

[0024]

[0025] Further technical solution of the present invention: Step 3 is as follows:

[0026] Step 3-1: Select the experimental working condition parameters as the initial value Ma in,0 ,α0;

[0027] Step 3-2: For step k, k = 0, 1, 2, ..., k max ,k max For the maximum number of iterations, neural network prediction is performed, and the input parameter Ma in,k and α k Normalize and obtain the DNN output result. Perform inverse transformation according to formula (3) to obtain the physical quantity corresponding to the measuring point. And calculate the experimental measurement data of the corresponding measurement points The mean square error is the loss function J(Ma in,k ,α k ), abbreviated as J k

[0028]

[0029] in, Indicates the measured physical quantity of the jth measurement point in the experimental measurement results. If J k+1 <ε or k equals k max , jump to step 4, where ε is the specified threshold;

[0030] Step 3-3: Use the automatic differentiation method to adjust the loss function J k Perform gradient back propagation to obtain the gradient of the loss function with respect to the incoming flow parameters

[0031]

[0032] Step 3-4: Use the gradient optimization method to correct the incoming flow parameters and update the flow field input parameters:

[0033]

[0034] Opt is the selected gradient optimization algorithm. The updated flow field input parameters are brought into step 3-2 for iteration.

[0035] A further technical solution of the present invention: the gradient optimization method includes gradient descent method, Adam, and AdaGrad.

[0036] A computer system, characterized in that it includes: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method.

[0037] A computer-readable storage medium is characterized by storing computer-executable instructions, which are used to implement the above method when executed.

[0038] The beneficial effects of the present invention are:

[0039] This invention provides a method for correcting parameters in cascade wind tunnel experiments based on flow field numerical simulation and deep neural networks. This method can effectively correct the incoming flow boundary conditions in cascade wind tunnel experiments and overcome the separation zone prediction bias caused by turbulence models in flow field numerical simulations during the correction process. This method eliminates the need for specific turbulence models and is particularly suitable for reconstructing and inverting experimental flow fields using flow field numerical simulation methods. The method is simple and produces highly accurate correction results.

[0040] The details are as follows:

[0041] 1. Because flow separation typically occurs in the trailing 50% of the blade, current turbulence models used in numerical flow simulations do not accurately predict flow separation. Consequently, the flow field prediction data for the trailing 50% of the blade is contaminated by the turbulence model's prediction bias, leading to correction errors. Therefore, the present invention utilizes data from the first 50% of the chord length, rather than the entire chord length. The first 50% represents attached flow without flow separation, and turbulence models generally provide very accurate predictions.

[0042] 2. Compared with the existing ensemble Kalman filter method, the deep neural network used in the present invention can better capture the nonlinear information of the flow process in the wind tunnel, with more accurate modeling and higher correction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0044] Figure 1 This is a flow chart of the method for correcting working condition parameters of cascade wind tunnel experiments based on flow field numerical simulation and deep neural network of the present invention;

[0045] Figure 2 This is a comparison chart of the flow field calculation results and experimental measurement results before and after the correction of the incoming flow parameters. DETAILED DESCRIPTION

[0046] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only intended to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0047] The present invention provides a method for correcting parameters of blade wind tunnel experimental working conditions based on flow field numerical simulation and deep neural network. The method comprises the following steps: floating the incoming flow Mach number and angle of attack of the experimental working conditions to be corrected, constructing multiple sets of incoming flow parameters and performing numerical calculation of the blade flow field; taking the floating incoming flow Mach number and angle of attack normalized as input, taking the corresponding physical quantities of the measuring point positions in the front 50% area of ​​the blade chord length of the numerical flow field normalized as output, and constructing a deep neural network; taking the mean square error of the deep neural network prediction result and the experimental measurement result as the loss function, using an automatic The differential algorithm backpropagates the gradient of the loss function to obtain the gradient of the mean square error with respect to the incoming flow Mach number and incoming flow angle of attack; the gradient optimization algorithm is used to correct the incoming flow Mach number and incoming flow angle of attack; the corrected incoming flow Mach number and incoming flow angle of attack are used to perform numerical calculations of the flow field. If the mean square error between the flow parameters of the measuring points in the numerical flow field and the experimental measurement results is less than the specified threshold, the corrected parameters can be used as the final correction result. If the mean square error is greater than the specified threshold, the numerical calculation data of the flow field is used as the training set to be added to the first step of the deep neural network, and the above process is repeated.

[0048] like Figure 1 As shown, the following steps are included:

[0049] Step 1: Floating operating parameters. The specific floating amplitude depends on the control accuracy of the incoming flow parameters. A typical floating range can be set to: incoming flow Mach number Ma in The floating angle is ±0.05, and the incoming flow angle α is floating ±1°. The Latin hypercube method is used for random sampling to construct samples of several incoming flow parameters. The appropriate turbulence model method is selected to perform numerical calculations of the cascade flow field for each sample.

[0050] Step 2: Construct a neural network that describes the incoming flow parameters and flow field parameters. The input parameters of the neural network are the normalized incoming flow parameter samples in step 1, and the output parameters are the corresponding physical quantities of the measurement position in the front 50% area of ​​the blade chord length in the flow field calculated for each normalized sample. The measured physical quantities include but are not limited to: surface isentropic Mach number, pressure coefficient, static pressure, etc.; the data of the front 50% chord length area is used here instead of all the data because flow separation generally occurs in the 50% area after the blade. The current turbulence model is not very accurate in predicting flow separation, so the data of the 50% after the blade is contaminated by the turbulence model prediction deviation, resulting in correction errors, while the 50% area before the blade is an attached flow without flow separation, which the turbulence model can generally predict very accurately.

[0051] Step 3: Use the experimentally measured operating parameters as the initial values; use the gradient optimization method to modify the incoming flow parameters and update the flow field input parameters:

[0052]

[0053] Where: J k is the loss function

[0054]

[0055] Where, is the physical quantity measured at the jth measuring point during the experiment, is the result of the jth measurement point predicted by the deep neural network, j = 1, 2, ..., M, M represents the number of measurement points in the front 50% chord length area of ​​the blade during the experiment, Ma in,k , α k is the incoming flow parameter of the kth iteration, k=0,1,2,…,k max ,k max is the maximum number of iterations, and the loss function J is modified using the automatic differentiation method. k Perform gradient back propagation to obtain the gradient of the loss function with respect to the incoming flow parameters

[0056]

[0057] Step 4: Use the flow field input parameters obtained by the current calculation as the corrected operating parameters of the modified blade in this experimental condition, recalculate the flow field numerically, and compare the corresponding data of the experimental measurement points of the calculation results with the experimental measurement data (the first 50% chord length area). If the deviation is significantly reduced compared to the previous one, generally it can be reduced by more than 50% after correction, then the operating parameter is output and the algorithm ends. If the deviation is not significantly reduced, add the calculated data to the training sample in step 1 and repeat steps 1 to 4.

[0058] Example 1:

[0059] Taking the MAN GHH 1-S1 compressor cascade as an example, its design Mach number is 0.62 and the design angle of attack is 0°. The distribution of the surface isentropic Mach number at 10 measurement points on the pressure and suction sides of the blade was experimentally measured. The incoming flow Mach number and incoming flow angle of attack of the wind tunnel test results of the design working condition were corrected. The numerical simulation and experimental measurement results of the blade surface isentropic Mach number before the correction parameters are shown in the figure. Figure 2 As shown, the specific implementation steps of this example are as follows:

[0060] Step 1: Floating operating parameters. The specific floating amplitude depends on the control accuracy of the incoming flow parameters. A typical floating range can be set as: floating incoming flow Mach number 0.62±0.05, incoming flow angle of attack floating 0°±1°. Latin hypercube and other methods can be used for random sampling to construct samples of incoming flow parameters. In this example, the number of samples is 32, and the cascade flow field is numerically calculated for each sample. The present invention is insensitive to the turbulence model in the numerical calculation of the flow field. It is sufficient to select an appropriate turbulence model. In this example, the SA turbulence model is selected.

[0061] Step 2: Construct a neural network to describe the incoming flow parameters and flow field parameters. The specific steps are as follows:

[0062] Step 2-1: The input parameters of the neural network are the normalized incoming Mach number and incoming flow angle of attack of each sample. in and the angle of attack α, the normalized formula is as follows:

[0063]

[0064]

[0065] Where, α max , α min 、Ma in，max 、Ma in，min , respectively represent the maximum and minimum values ​​of the incoming flow angle of attack and the incoming flow Mach number in the training data;

[0066] Step 2-2: The output parameter of the neural network is the surface isentropic Mach number of the blade measuring point corresponding to the first 50% chord length in the flow field calculation result of each sample after normalization. For the surface isentropic Mach number Ma of j measuring points is,j , the normalization formula is as follows:

[0067]

[0068] Where j = 1, 2, ..., M, M represents the number of blade measurement points corresponding to the first 50% of the chord length in the experimental measurement. In this example, M = 12, Ma is,j,min 、Ma is,j,maxare the minimum and maximum values ​​of the surface isentropic Mach number at the jth measuring point in the sample flow field. In this example, the surface isentropic Mach numbers of the pressure side and suction side of the blade are measured experimentally. In practical applications, the physical quantities measured on the blade surface can be: surface isentropic Mach number, pressure coefficient, static pressure, etc.

[0069] Step 2-3: Build a deep neural network (DNN). For this fully connected network, one input layer, one output layer, and three hidden layers basically meet the accuracy requirements. The mathematical expression is as follows:

[0070]

[0071] Step 3: Use the experimentally measured operating parameters as the initial values; use the gradient descent optimization method to correct the incoming flow parameters and update the flow field input parameters. The specific steps are as follows:

[0072] Step 3-1: For this example, the initial value of the operating condition parameter is Ma in,0 =0.62, α0=0°;

[0073] Step 3-2: For step k, k = 0, 1, 2, ..., k max ,k max For the maximum number of iterations, neural network prediction is performed, and the input parameter Ma in,k and α k Normalize according to formula (1) and (2) respectively to obtain the DNN output result. Perform inverse transformation according to formula (3) to obtain the surface isentropic Mach number Ma corresponding to the measuring point. is,j , and calculate the corresponding measurement point experimental measurement data Ma is,exp,j The mean square error is the loss function J(Ma in,k ,α k ), abbreviated as J k

[0074]

[0075] Among them, Ma is,exp,j represents the surface isentropic Mach number of the jth measuring point in the experimental measurement results. If J k+1 <ε or k equals k max , jump to step 4, where ε is the specified threshold;

[0076] Step 3-3: Use the automatic differentiation method to adjust the loss function J k Perform gradient back propagation to obtain the gradient of the loss function with respect to the incoming flow parameters

[0077]

[0078] Step 3-4: Use gradient optimization methods, such as gradient descent, Adam, AdaGrad, etc., to correct the incoming flow parameters and update the flow field input parameters:

[0079]

[0080] Opt is the selected gradient optimization algorithm. The updated flow field input parameters are brought into step 3-2 for iteration. Since the prediction uses a deep neural network, the calculation time is very small.

[0081] Step 4: Use the calculated flow field input parameters as the corrected operating parameters for the modified cascade for this experimental condition. Recalculate the flow field numerically and compare the calculated data corresponding to the experimental measurement points with the experimentally measured data for the first 50% of the chord length. If the deviation is significantly reduced compared to the previous value (the reduction depends on the specific situation, but generally can be reduced by approximately 50% after correction), the operating parameter is output and the algorithm terminates. If the deviation is not significantly reduced, add the calculated data to the training sample from Step 1 and repeat Steps 1 through 4. For this example and most other cases, performing Steps 1 through 4 once will yield satisfactory results. In this example, the corrected operating parameters are: an incoming Mach number of 0.6369 and an incoming angle of attack of -0.3259°.

[0082] The comparison of the corrected predicted isentropic Mach number curve and the experimental result is shown in the attached figure. Figure 2 It can be seen that after correcting the incoming flow parameters, the numerical simulation results of the flow field are in good agreement with the experimental results except for the trailing edge area. The deviation in the trailing edge area is due to the limited prediction ability of the turbulence model for the separation zone, but the turbulence model has a high accuracy in predicting the attached flow. When performing parameter correction, the present invention only uses the experimental measurement data of the front 50% chord length area of ​​the blade, which is the attached flow, thereby avoiding the influence of the turbulence model prediction deviation.

[0083] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present invention, and these modifications or replacements should all be included in the scope of protection of the present invention.

Claims

1. A method for correcting working condition parameters of cascade wind tunnel experiments based on flow field numerical simulation and deep neural network, characterized in that: Here are the steps: Step 1: Floating experimental working condition parameters, constructing multiple groups of incoming flow parameter samples, and performing flow field numerical simulation for each group of parameters; Step 2: Construct a neural network that describes the incoming flow parameters and flow field parameters. The input parameters of the neural network are the normalized incoming flow parameter samples from step 1, and the output parameters are the corresponding physical quantities of the measurement position in the front 50% of the blade chord length in the flow field calculated for each normalized sample. The details are as follows: Step 2-1: The input parameters of the neural network are the normalized incoming flow parameters of each sample. For the incoming flow Mach number and angle of attack , the normalization formula is as follows: (1) (2) Where, 、 、 、 Respectively represent the maximum and minimum values ​​of the incoming flow angle of attack and the incoming flow Mach number in the training data; Step 2-2: The output parameter of the neural network is the physical quantity of the blade measurement point position corresponding to the first 50% chord length in the flow field calculation result of each sample after normalization. j Physical quantity of a measuring point , the normalization formula is as follows: (3) Where, Indicates the number of blade measurement points corresponding to the first 50% of the chord length during the experiment, 、 are respectively the j The minimum and maximum values ​​of the physical quantities at each measuring point; Step 2-3: Construct a deep neural network DNN. The deep neural network DNN includes an input layer, several hidden layers and an output layer. The mathematical expression is as follows: (4) Step 3: Use the experimentally measured operating parameters as initial values; use the gradient optimization method to modify the incoming flow parameters and update the flow field input parameters; the details are as follows: Step 3-1: Select experimental parameters as initial values ; Step 3-2: For k step, For the maximum number of iterations, perform neural network prediction and input parameters and Normalize and obtain the DNN output result. Perform inverse transformation according to formula (3) to obtain the physical quantity corresponding to the measuring point. , and calculate the experimental measurement data of the corresponding measurement points The mean square error is the loss function , abbreviated as (5) in, Indicates the experimental measurement results j If the measured physical quantity of a measuring point is or k equal , jump to step 4, where ε is the specified threshold; Step 3-3: Use automatic differentiation method to adjust the loss function Perform gradient back propagation to obtain the gradient of the loss function with respect to the incoming flow parameters : (6) Step 3-4: Use the gradient optimization method to correct the incoming flow parameters and update the flow field input parameters: (7) Opt is the selected gradient optimization algorithm. The updated flow field input parameters are brought into step 3-2 for iteration. Step 4: Use the flow field input parameters obtained from the current calculation as the corrected operating parameters of the modified blade in this experimental condition, and use the corrected incoming flow Mach number and incoming flow angle of attack to perform numerical calculation of the flow field. If the mean square error between the flow parameters of the measuring point in the numerical flow field and the experimental measurement data of the first 50% chord length area is less than the specified threshold, the corrected incoming flow Mach number and incoming flow angle of attack are used as the final correction results. If the mean square error is greater than the specified threshold, the numerical calculation data of this flow field is used as the training set for adding to the deep neural network, and steps 1 to 4 are repeated.

2. The method for correcting working condition parameters of cascade wind tunnel experiments based on flow field numerical simulation and deep neural network according to claim 1, characterized in that: Step 1 is as follows: set the floating range according to the control accuracy of the incoming flow parameters, use the Latin hypercube method for random sampling, construct a sample of the incoming flow parameters, and perform numerical calculation of the blade flow field for each sample in the turbulence model.

3. The method for correcting working condition parameters of cascade wind tunnel experiments based on flow field numerical simulation and deep neural network according to claim 1 is characterized in that: The corresponding physical quantities described in step 2 include the incoming flow Mach number, pressure coefficient, and static pressure.

4. The method for correcting working condition parameters of cascade wind tunnel experiments based on flow field numerical simulation and deep neural network according to claim 1, characterized in that: The gradient optimization methods include gradient descent method, Adam, and AdaGrad.

5. A computer system, characterized in that include: One or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the method of claim 1.

6. A computer-readable storage medium, characterized in that Computer-executable instructions are stored, and when the instructions are executed, they are used to implement the method of claim 1.

Images