Vehicle front side window flow field excitation rapid prediction method based on machine learning

Abstract

The invention discloses an automobile front side window flow field excitation rapid prediction method based on machine learning, and the method comprises the following steps: 1), carrying out the processing of the time domain flow field data of the surface of an automobile front side window through a wave number frequency spectrum decomposition method, and obtaining the frequency domain flow field excitation data of the surface of the automobile front side window; 2) measuring geometric parameter data of the automobile, and forming mapping with frequency domain flow field excitation data on the surface of the front side window of the automobile; 3) constructing a sample data set taking automobile geometric parameter data as input and frequency domain flow field excitation data as output; 4) training the neural network model by using the sample data set to obtain a front side window flow field excitation prediction model; and 5) measuring geometric parameter data of a to-be-measured vehicle, and inputting the geometric parameter data into the front side window flow field excitation prediction model to obtain a front side window flow field excitation prediction value of the to-be-measured vehicle. According to the invention, rapid prediction of front side window excitation can be realized.

Description

Technical Field

[0001] The present invention relates to the fields of aerodynamics and machine learning, and specifically to a fast prediction method for the flow field excitation of the front side window of a vehicle based on machine learning. Background Art

[0002] As an indispensable tool in modern human life and work, the comfort inside a vehicle has gradually attracted the attention of users. With the innovation of industrial technologies and the popularization of electric vehicles, the transmission system noise and tire noise have been effectively controlled. Aerodynamic noise has gradually become an important factor affecting the interior noise level of a vehicle, especially under the condition of high-speed driving, where the aerodynamic noise is particularly obvious. Therefore, the research on issues related to aerodynamic noise has always received extensive attention. Computational Fluid Dynamics (CFD) simulation technology is currently the most mainstream technical method for studying the interior wind noise of a vehicle.

[0003] However, there is an unavoidable problem with the CFD simulation method, that is, the trade-off between the solution efficiency and the solution accuracy. Its solution time is positively correlated with the number of model grids, and as the number of grids increases, the solution time will increase exponentially. However, an insufficient number of grids will result in insufficient accuracy of the solution results or even incorrect solutions. Therefore, the balance between solution efficiency and solution accuracy has always been a difficult problem to solve. Summary of the Invention

[0004] The purpose of the present invention is to provide a fast prediction method for the flow field excitation of the front side window of a vehicle based on machine learning, including the following steps:

[0005] 1) Using computational fluid dynamics simulation technology to calculate the time-domain flow field data on the surface of the front side window of the vehicle;

[0006] 2) Using the wave number-frequency spectrum decomposition method to process the time-domain flow field data on the surface of the front side window of the vehicle to obtain the frequency-domain flow field excitation data on the surface of the front side window of the vehicle;

[0007] 3) Measuring the vehicle geometric parameter data and forming a mapping with the frequency-domain flow field excitation data on the surface of the front side window of the vehicle;

[0008] 4) Repeating steps 1) to 3) to construct a sample data set with the vehicle geometric parameter data as the input and the frequency-domain flow field excitation data as the output;

[0009] 5) Using machine learning technology to establish a neural network model;

[0010] 6) Using the sample data set to train the neural network model to obtain a front side window flow field excitation prediction model;

[0011] 7) Obtain the geometric parameter data of the vehicle to be tested and input it into the front side window flow field excitation prediction model to obtain the front side window flow field excitation prediction value of the vehicle to be tested.

[0012] Further, in step 1), the steps of calculating the time-domain flow field data on the front side window surface of the vehicle include:

[0013] 1.1) Establish a virtual numerical wind tunnel in the computational fluid dynamics simulation software. The virtual numerical wind tunnel is a cuboid, and two of its walls are used as the velocity inlet wall and the pressure outlet wall;

[0014] 1.2) Construct a closed model of the vehicle to be tested;

[0015] 1.3) Place the model of the vehicle to be tested in the virtual numerical wind tunnel and close to the velocity inlet wall;

[0016] 1.4) Set a grid-dense area in the front side window area of the model of the vehicle to be tested;

[0017] 1.5) Set the incoming flow velocity on the velocity inlet wall, set the pressure on the pressure outlet wall to 0 Pa, and set the remaining four walls and the vehicle body wall as non-slip walls;

[0018] 1.6) Run the simulation software to obtain the time-domain flow field data on the front side window surface of the vehicle.

[0019] Further, the length of the virtual numerical wind tunnel is 15 times the length of the vehicle body of the vehicle to be tested, the width is 13 times the width of the vehicle body, and the height is 7 times the height of the vehicle body.

[0020] Further, the distance between the model of the vehicle to be tested and the velocity inlet wall is 3 times the length of the vehicle body, and the distance between both sides of the vehicle body and both side walls is 6 times the width of the vehicle body.

[0021] Further, in step 2), the frequency-domain flow field excitation data on the front side window surface of the vehicle includes the sound pressure excitation on the front side window surface of the vehicle and the hydrodynamic pressure excitation on the front side window surface of the vehicle.

[0022] Further, in step 2), the steps of processing the time-domain flow field data on the front side window surface of the vehicle include:

[0023] 2.1) Perform Fourier transform on the time-domain flow field data p(x, y, t) on the front side window surface of the vehicle to obtain:

[0024] p * (x, y, ω) = ∫p(x, y, t)e (-iωt) dt (1)

[0025] where ω is the circular frequency, x, y are spatial quantities, t is time, e is the natural constant, and i is the imaginary number; p * (x, y, ω) is the flow field data after Fourier transform;

[0026] 2.2) Perform a second Fourier transform on the pressure pulsations in the frequency domain in the x and y directions to obtain the pressure data in the wavenumber domain, i.e.:

[0027]

[0028] where k x is the wavenumber in the x direction, and k y is the wavenumber in the y direction;

[0029] 2.3) In the two-dimensional wavenumber domain (k x , k y ), separate the acoustic energy and the hydrodynamic energy according to the difference in the propagation speeds of the acoustic energy and the hydrodynamic energy;

[0030] Among them, the distribution of the acoustic energy is as follows:

[0031]

[0032] where c is the speed of sound; f is the frequency;

[0033] The distribution of the hydrodynamic energy is as follows:

[0034]

[0035] where u is the air flow velocity.

[0036] 2.4) Use the decomposed acoustic energy as the acoustic pressure excitation on the front side window surface of the vehicle, and the decomposed hydrodynamic energy as the hydrodynamic pressure excitation on the front side window surface of the vehicle.

[0037] Furthermore, the geometric parameters of the vehicle are measured by geometric measurement means.

[0038] Furthermore, the geometric parameters of the vehicle include, but are not limited to: vehicle width, front break angle, A-pillar width, A-pillar deflection, A-pillar side step height, A-pillar side inclination angle, A-pillar included angle, rearview mirror handle thickness, rearview mirror inner side included angle, rearview mirror outer side included angle, rearview mirror upper part deflection, rearview mirror bottom deflection, rearview mirror distance from the side window, rearview mirror height, rearview mirror width, etc., which are parameters that can describe the geometric shape of the A-pillar and the rearview mirror.

[0039] Furthermore, in step 6), the steps of training the neural network model using the sample data set include:

[0040] 6.1) Normalize the data in the sample data set, and divide the normalized sample data set into a test set and a training set;

[0041] 6.2) Select the training function of the neural network model and the activation functions of each hidden layer;

[0042] 6.3) Train the neural network model using the training set;

[0043] 6.4) Test the neural network model using the test set. If the error is less than the preset value, output the front side window flow field excitation prediction model; otherwise, reconstruct the sample data set.

[0044] Furthermore, the neural network model includes a backpropagation neural network model.

[0045] The technical effect of the present invention is beyond doubt. The present invention proposes a fast prediction method for the front side window flow field excitation of an automobile based on machine learning. This method establishes the relationship between the geometric parameters of the automobile and the side window flow field excitation based on machine learning, and then realizes the fast prediction of the front side window flow field excitation by inputting the geometric parameters of the automobile. It has the advantages of high prediction efficiency, simple process, and accurate results. It has great application value and broad application prospects in the fields of automotive aerodynamic shape optimization design and automotive wind noise prediction. Description of the Drawings

[0046] Figure 1 It is a flowchart of a fast prediction method for the front side window flow field excitation of an automobile based on machine learning;

[0047] Figure 2 It is a schematic diagram of the size of the virtual numerical wind tunnel;

[0048] Figure 3 It is a schematic diagram of the dense area of the front side window grid;

[0049] Figure 4 It is a diagram of the frequency domain sound pressure excitation and hydrodynamic pressure excitation data obtained by decomposing the time domain flow field data using the WFS method;

[0050] Figure 5 It is a schematic diagram of the position for collecting the geometric parameters of the automobile;

[0051] Figure 6 It is a schematic diagram of the BPNN neural network structure;

[0052] Figure 7 It is a schematic diagram of the BPNN neural network structure adopted in Embodiment 2;

[0053] Figure 8 It is a comparison of the predicted sound pressure excitation and the expected sound pressure excitation obtained by this method in Test Set 1;

[0054] Figure 9 It is a comparison of the predicted sound pressure excitation and the expected sound pressure excitation obtained by this method in Test Set 2;

[0055] Figure 10 It is a comparison of the predicted hydrodynamic pressure excitation and the expected sound pressure excitation obtained by this method in Test Set 1;

[0056] Figure 11 This is the comparison between the predicted hydrodynamic pressure excitation and the expected sound pressure excitation obtained by this method in Test Set 2;

[0057] Figure 12 This is the rapid prediction process of the flow field excitation of the front side window of the car. Specific implementation mode

[0058] The present invention will be further described below in conjunction with embodiments, but it should not be understood that the above-mentioned subject scope of the present invention is limited to the following embodiments. Without departing from the above-mentioned technical idea of the present invention, various substitutions and changes made according to ordinary technical knowledge and customary means in the art shall be included within the protection scope of the present invention.

[0059] Embodiment 1:

[0060] See Figures 1 to 12 , a rapid prediction method for the flow field excitation of the front side window of a car based on machine learning, comprising the following steps:

[0061] 1) Use computational fluid dynamics simulation technology to calculate the time-domain flow field data on the surface of the front side window of the car;

[0062] 2) Use the wave number-frequency spectrum decomposition method to process the time-domain flow field data on the surface of the front side window of the car to obtain the frequency-domain flow field excitation data on the surface of the front side window of the car;

[0063] 3) Measure the geometric parameter data of the car and form a mapping with the frequency-domain flow field excitation data on the surface of the front side window of the car;

[0064] 4) Repeat steps 1) to 3) to construct a sample data set with the geometric parameter data of the car as the input and the frequency-domain flow field excitation data as the output;

[0065] 5) Use machine learning technology to establish a neural network model;

[0066] 6) Use the sample data set to train the neural network model to obtain a front side window flow field excitation prediction model;

[0067] 7) Obtain the geometric parameter data of the vehicle to be tested and input it into the front side window flow field excitation prediction model to obtain the front side window flow field excitation prediction value of the vehicle to be tested. The acquisition methods of the geometric parameter data of the vehicle to be tested include: 1. Obtain through design data. 2. Obtain through geometric measurement of the three-dimensional model of the vehicle to be tested.

[0068] In step 1), the steps of calculating the time-domain flow field data on the surface of the front side window of the car include:

[0069] 1.1) Establish a virtual numerical wind tunnel in the hydrodynamic simulation software. The virtual numerical wind tunnel is a cuboid, with two of its walls serving as the velocity inlet wall and the pressure outlet wall;

[0070] 1.2) Construct a closed vehicle model to be tested;

[0071] 1.3) Place the vehicle model to be tested inside the virtual numerical wind tunnel and close to the velocity inlet wall;

[0072] 1.4) Set a region with a dense grid in the front side window area of the vehicle model to be tested;

[0073] 1.5) Set the incoming flow velocity on the velocity inlet wall, set the pressure on the pressure outlet wall to 0 Pa, and set the remaining four walls and the vehicle body wall as non-slip walls;

[0074] 1.6) Run the simulation software to obtain the time-domain flow field data on the front side window surface of the vehicle.

[0075] The length of the virtual numerical wind tunnel is 15 times the length of the vehicle body to be tested, the width is 13 times the width of the vehicle body, and the height is 7 times the height of the vehicle body.

[0076] The distance between the vehicle model to be tested and the velocity inlet wall is 3 times the length of the vehicle body, and the distances between both sides of the vehicle body and the two side walls are 6 times the width of the vehicle body.

[0077] In step 2), the frequency-domain flow field excitation data on the front side window surface of the vehicle includes the sound pressure excitation and the hydrodynamic pressure excitation on the front side window surface of the vehicle.

[0078] In step 2), the steps for processing the time-domain flow field data on the front side window surface of the vehicle include:

[0079] 2.1) Perform a Fourier transform on the time-domain flow field data p(x, y, t) on the front side window surface of the vehicle to obtain:

[0080] p * (x, y, ω) = ∫p(x, y, t)e (-iωt) dt (1)

[0081] where ω is the circular frequency, x and y are spatial quantities, t is time, e is the natural constant, and i is the imaginary number; p * (x, y, ω) is the flow field data after Fourier transform;

[0082] 2.2) Perform a second Fourier transform on the pressure pulsation in the x and y directions in the frequency domain to obtain the pressure data in the wavenumber domain, that is:

[0083]

[0084] where k xis the wave number in the x - direction, k y is the wave number in the y - direction;

[0085] 2.3) In the two - dimensional wave number domain (k x , k y ), separate the acoustic energy and hydrodynamic energy according to the difference in the propagation speeds of the acoustic energy and the hydrodynamic energy;

[0086] Among them, the distribution of the acoustic energy is as follows:

[0087]

[0088] In the formula, c is the speed of sound; f is the frequency;

[0089] The distribution of the hydrodynamic energy is as follows:

[0090]

[0091] In the formula, u is the air flow velocity.

[0092] 2.4) Use the decomposed acoustic energy as the acoustic pressure excitation on the front side window surface of the vehicle, and the decomposed hydrodynamic energy as the hydrodynamic pressure excitation on the front side window surface of the vehicle.

[0093] The geometric parameters of the vehicle are measured by geometric measurement means.

[0094] The geometric parameters of the vehicle include but are not limited to: vehicle width, front break - wind angle, A - pillar width, A - pillar deflection, A - pillar side - step height, A - pillar side - inclination angle, A - pillar included angle, rear - view mirror handle thickness, included angle of the inner side of the rear - view mirror, included angle of the outer side of the rear - view mirror, upper deflection of the rear - view mirror, bottom deflection of the rear - view mirror, distance between the rear - view mirror and the side window, rear - view mirror height, rear - view mirror width, etc., which are parameters that can describe the geometric shape of the A - pillar and the rear - view mirror.

[0095] In step 6), the steps of training the neural network model using the sample data set include:

[0096] 6.1) Normalize the data in the sample data set and divide the normalized sample data set into a test set and a training set;

[0097] 6.2) Select the training function of the neural network model and the activation functions of each hidden layer;

[0098] 6.3) Use the training set to train the neural network model;

[0099] 6.4) Use the test set to test the neural network model. If the error is less than the preset value, output the front side - window flow - field excitation prediction model; otherwise, reconstruct the sample data set.

[0100] The neural network model includes a back - propagation neural network model.

[0101] Example 2:

[0102] A fast prediction method for the flow field excitation of the front side window of a vehicle based on machine learning, comprising the following steps:

[0103] 1) Use computational fluid dynamics simulation technology to calculate the time-domain flow field data on the surface of the front side window of the vehicle;

[0104] 2) Use the wave number-frequency spectrum decomposition method to process the time-domain flow field data on the surface of the front side window of the vehicle to obtain the frequency-domain flow field excitation data on the surface of the front side window of the vehicle;

[0105] 3) Measure the geometric parameter data of the vehicle and form a mapping with the frequency-domain flow field excitation data on the surface of the front side window of the vehicle;

[0106] 4) Repeat steps 1) to 3) to construct a sample data set with the geometric parameter data of the vehicle as the input and the frequency-domain flow field excitation data as the output;

[0107] 5) Use machine learning technology to establish a neural network model;

[0108] 6) Use the sample data set to train the neural network model to obtain a front side window flow field excitation prediction model;

[0109] 7) Obtain the geometric parameter data of the vehicle to be tested and input it into the front side window flow field excitation prediction model to obtain the front side window flow field excitation prediction value of the vehicle to be tested.

[0110] Example 3:

[0111] A fast prediction method for the flow field excitation of the front side window of a vehicle based on machine learning, the technical content is the same as that of Example 2. Further, in step 1), the steps of calculating the time-domain flow field data on the surface of the front side window of the vehicle include:

[0112] 1.1) Establish a virtual numerical wind tunnel in the fluid dynamics simulation software. The virtual numerical wind tunnel is a cuboid, and two of its walls are used as the velocity inlet wall and the pressure outlet wall;

[0113] 1.2) Construct a closed vehicle model to be tested;

[0114] 1.3) Place the vehicle model to be tested in the virtual numerical wind tunnel and close to the velocity inlet wall;

[0115] 1.4) Set a grid-dense area in the front side window area of the vehicle model to be tested;

[0116] 1.5) Set the incoming flow velocity on the velocity inlet wall, set the pressure on the pressure outlet wall to 0 Pa, and set the remaining four walls and the vehicle body wall as non-slip walls;

[0117] 1.6) Run the simulation software to obtain the time-domain flow field data on the surface of the front side window of the vehicle.

[0118] Example 4:

[0119] A fast prediction method for the flow field excitation of the front side window of a vehicle based on machine learning, the technical content is the same as any one of Examples 2-3. Further, the length of the virtual numerical wind tunnel is 15 times the length of the body of the vehicle to be measured, the width is 13 times the width of the body, and the height is 7 times the height of the body.

[0120] Example 5:

[0121] A fast prediction method for the flow field excitation of the front side window of a vehicle based on machine learning, the technical content is the same as any one of Examples 2-4. Further, the distance between the model of the vehicle to be measured and the velocity inlet wall surface is 3 times the length of the body, and the distances between both sides of the body and the side wall surfaces are 6 times the width of the body.

[0122] Example 6:

[0123] A fast prediction method for the flow field excitation of the front side window of a vehicle based on machine learning, the technical content is the same as any one of Examples 2-5. Further, in step 2), the frequency-domain flow field excitation data on the surface of the front side window of the vehicle includes the acoustic pressure excitation on the surface of the front side window of the vehicle and the hydrodynamic pressure excitation on the surface of the front side window of the vehicle.

[0124] Example 7:

[0125] A fast prediction method for the flow field excitation of the front side window of a vehicle based on machine learning, the technical content is the same as any one of Examples 2-6. Further, in step 2), the steps for processing the time-domain flow field data on the surface of the front side window of the vehicle include:

[0126] 2.1) Perform Fourier transform on the time-domain flow field data on the surface of the front side window of the vehicle to obtain:

[0127] p * (x,y,ω) = ∫p(x,y,t)e (-iωt) dt (1)

[0128] In the formula, ω is the circular frequency, x and y are spatial quantities, t is time, e is the natural constant, and i is the imaginary number;

[0129] 2.2) Perform a second Fourier transform on the pressure pulsation in the frequency domain in the x and y directions to obtain the pressure data in the wave number domain, that is:

[0130]

[0131] In the formula, k x is the wave number in the x direction, and k y is the wave number in the y direction;

[0132] 2.3) In the two-dimensional wavenumber domain (k x , k y ), separate the acoustic energy and hydrodynamic energy according to the difference in the propagation speeds of the acoustic energy and hydrodynamic energy;

[0133] Among them, the distribution of the acoustic energy is as follows:

[0134]

[0135] In the formula, c is the speed of sound; f is the frequency;

[0136] The distribution of the hydrodynamic energy is as follows:

[0137]

[0138] In the formula, u is the air flow velocity.

[0139] 2.4) Use the decomposed acoustic energy as the acoustic pressure excitation on the front side window surface of the vehicle, and the decomposed hydrodynamic energy as the hydrodynamic pressure excitation on the front side window surface of the vehicle.

[0140] Example 8:

[0141] A fast prediction method for the flow field excitation of the front side window of a vehicle based on machine learning, the technical content is the same as any one of Examples 2-7. Further, the geometric parameters of the vehicle are measured by geometric measurement means.

[0142] Example 9:

[0143] A fast prediction method for the flow field excitation of the front side window of a vehicle based on machine learning, the technical content is the same as any one of Examples 2-8. Further, the geometric parameters of the vehicle include but are not limited to: vehicle width, front break angle, A-pillar width, A-pillar deflection, A-pillar side step height, A-pillar side inclination angle, A-pillar included angle, rearview mirror handle thickness, inner side included angle of the rearview mirror, outer side included angle of the rearview mirror, upper deflection of the rearview mirror, bottom deflection of the rearview mirror, distance between the rearview mirror and the side window, rearview mirror height, rearview mirror width, etc., which can describe the geometric shape of the A-pillar and the rearview mirror.

[0144] Example 10:

[0145] A fast prediction method for the flow field excitation of the front side window of a vehicle based on machine learning, the technical content is the same as any one of Examples 2-9. Further, in step 6), the steps of training the neural network model using the sample data set include:

[0146] 6.1) Normalize the data in the sample data set, and divide the normalized sample data set into a test set and a training set;

[0147] 6.2) Select the training function of the neural network model and the activation functions of each hidden layer;

[0148] 6.3) Train the neural network model using the training set;

[0149] 6.4) Test the neural network model using the test set. If the error is less than the preset value, output the front side window flow field excitation prediction model; otherwise, reconstruct the sample data set.

[0150] Example 11:

[0151] A fast prediction method for the flow field excitation of the front side window of an automobile based on machine learning, the technical content is the same as any one of Examples 2-10. Further, the neural network model includes a backpropagation neural network model.

[0152] Example 12:

[0153] A fast prediction method for the flow field excitation of the front side window of an automobile based on machine learning, the steps are as follows:

[0154] A fast prediction method for the flow field excitation of the front side window of an automobile based on machine learning, including the following steps:

[0155] 1) Use computational fluid dynamics (CFD) simulation technology to obtain the time-domain flow field data on the surface of the front side window of the automobile;

[0156] 2) Use the wavenumber frequency spectrum decomposition method (WFS) to process the time-domain flow field data obtained in step 1) on the surface of the front side window of the automobile to obtain the frequency-domain flow field excitation data on the surface of the front side window of the automobile;

[0157] 3) Use geometric measurement means to measure the geometric parameter data of the automobile;

[0158] 4) Use machine learning technology to establish a backpropagation neural network (BPNN) model;

[0159] 5) Use the sample of the geometric parameter data of the automobile measured in step 3) as the input sample set of the BPNN model, and use the frequency-domain flow field excitation data calculated in step 2) on the surface of the front side window of the automobile as the BPNN output sample set to train the neural network model and establish a neural network model for quickly predicting the flow field excitation of the front side window;

[0160] 6) Measure the geometric parameter data of the vehicle to be tested, input it into the front side window flow field excitation quick prediction model established in step 5), and obtain the front side window flow field excitation prediction value of the vehicle to be tested.

[0161] The steps for calculating the time-domain flow field data on the front side window surface of a vehicle using CFD simulation technology include:

[0162] 1.1) Establish a virtual numerical wind tunnel in the CFD simulation software. The virtual numerical wind tunnel is approximately 15 times the length of the vehicle body to be measured, about 13 times the width of the vehicle body, and about 7 times the height of the vehicle body. The virtual numerical wind tunnel consists of a velocity inlet wall surface, a pressure outlet wall surface, and the remaining four wall boundaries;

[0163] 1.2) Process the vehicle model to be measured, retaining only the outermost geometric surface and sealing it;

[0164] 1.3) Place the vehicle to be measured at the end of the virtual numerical wind tunnel near the velocity inlet wall surface, with the front of the vehicle about 3 times the length of the vehicle body away from the velocity inlet wall surface, and the two sides of the vehicle about 6 times the width of the vehicle body away from the two side wall surfaces;

[0165] 1.4) Set up 8 layers of prism grids with a boundary layer thickness of 2 mm in the front side window area of the vehicle to be measured, and the grid thickness growth rate between adjacent layers is 1.5. Set up a grid-dense area to improve the accuracy of obtaining side window flow field data;

[0166] 1.5) Set the incoming flow velocity on the velocity inlet wall surface, set the pressure outlet wall surface to 0 Pa, and set the remaining four wall surfaces and the vehicle body wall surface to non-slip wall surfaces;

[0167] 1.6) Run the simulation software to obtain the time-domain flow field data of the vehicle to be measured and save it.

[0168] The steps for solving the frequency-domain flow field excitation on the front side window surface of a vehicle using the WFS method include:

[0169] 2.1) Perform Fourier transform on the flow field data of the front side window of the vehicle obtained by solving in the CFD simulation software (i.e., the pressure pulsation p(x, y, t) in the time domain) to convert it to the frequency domain:

[0170] p * (x, y, ω) = ∫p(x, y, t)e (-iωt) dt (1)

[0171] In the formula, ω is the circular frequency, x and y are spatial quantities, t is time, e is the natural constant, and i is the imaginary number.

[0172] 2.2) Perform Fourier transform on the pressure pulsation in the frequency domain in the x and y directions to obtain the pressure data in the wave number domain:

[0173]

[0174] In the formula, k x is the wave number in the x direction, and k y is the wave number in the y direction.

[0175] 2.3) In the two-dimensional wavenumber domain (k x , k y ), the two kinds of energy are separated according to the difference in the propagation speeds of acoustic energy and hydrodynamic energy.

[0176] The distribution of acoustic energy is:

[0177]

[0178] The distribution of hydrodynamic energy is:

[0179]

[0180] Where c is the speed of sound, f is the frequency, and u is the air flow velocity.

[0181] 2.4) The decomposed acoustic energy is the acoustic pressure excitation on the front side window surface of the vehicle, and the decomposed hydrodynamic energy is the hydrodynamic pressure excitation on the front side window surface of the vehicle.

[0182] The geometric parameters of the vehicle measured by geometric measurement means include but are not limited to: vehicle width, front break angle, A-pillar width, A-pillar deflection, A-pillar side step height, A-pillar side inclination angle, A-pillar included angle, rearview mirror handle thickness, included angle between the inner side surface of the rearview mirror, included angle between the outer side surface of the rearview mirror, upper deflection of the rearview mirror, bottom deflection of the rearview mirror, distance between the rearview mirror and the side window, rearview mirror height, rearview mirror width, etc., which are parameters that can describe the geometric shape of the A-pillar and the rearview mirror.

[0183] The steps of establishing a BPNN model using machine learning technology include:

[0184] 4.1) Determine the basic structure of the neural network, that is, one input layer, single or multiple hidden layers, and one output layer;

[0185] 4.2) Determine the number of neurons in each layer of the neural network. The number of neurons in the input layer is the same as the number of selected geometric parameters, and the number of neurons in each hidden layer is more than the number of neurons in the input layer;

[0186] 4.3) Select the training function of the neural network. The training function is the trainlm function, that is, the Levenberg-Marquardt (LM) algorithm. For the following non-linear least squares problem:

[0187]

[0188] Where H(x) is the sum of squared residuals, P i ′ is the predicted vector of the network model, P i is the desired output vector, n is the number of samples, e i (x) is the residual, and a is the vector composed of network weights and thresholds.

[0189] The iterative step size of the LM algorithm is:

[0190] I LM = -(J T (a)J(a)+μI) -1 J T (a)e(a) (6)

[0191] Where: μ is the damping factor, I is the identity matrix, and J(a) is the Jacobian matrix.

[0192] 4.4) Select the activation function for each hidden layer of the neural network. The activation function is the Sigmoid function. Its mathematical definition is:

[0193]

[0194] Where f(b) is the output value of the Sigmoid function, e is the natural constant, and b is the input value of the previous layer of neurons.

[0195] 4.5) Determine that the representation of the performance of the neural network model is the mean squared error (Mean Squared Error, MSE):

[0196]

[0197] Where n is the number of samples, P i ′ is the predicted vector of the network model, and P i is the desired output vector.

[0198] Using the geometric parameter sample data as the input and the sound pressure excitation and hydrodynamic pressure excitation sample data as the output, the steps for training the neural network model include:

[0199] 5.1) Normalize the input and output sample data. Use the maximum-minimum normalization method to normalize the data to the interval [0, 1]:

[0200]

[0201] Where is the value after normalization, i represents the i-th type of sample, i = 1, 2,..., p, m represents the m-th feature in the same type of sample, m = 1, 2,..., n, and α im represents the m-th feature value in the i-th type of sample.

[0202] 5.2) Divide the sample data into a training sample set and a test sample set. The training sample set is used to train the network model. The test sample set is used to verify the model prediction results and shall not be used for network training;

[0203] 5.3) Use the geometric parameter sample data in the training sample set as the input of the network model, and the sample data of the acoustic pressure excitation and hydrodynamic pressure excitation on the front side window surface of the vehicle as the output of the network model to train the network model;

[0204] 5.4) Use the test sample set to test the model prediction effect;

[0205] Obtain a fast prediction model for the flow field excitation of the front side window of the vehicle based on machine learning, and the acoustic pressure excitation and hydrodynamic pressure excitation on the front side window surface can be quickly predicted by inputting the corresponding geometric parameters of the vehicle.

[0206] Example 13:

[0207] See Figures 1 to 11 , a fast prediction method for the flow field excitation of the front side window of a vehicle based on machine learning, including the following steps:

[0208] 1) Use computational fluid dynamics (CFD) simulation technology to obtain the time-domain flow field data on the front side window surface of the vehicle;

[0209] The steps of using CFD simulation technology to calculate the flow field data on the front side window surface of the vehicle include:

[0210] 1.1) Establish a virtual numerical wind tunnel in the CFD simulation software. The virtual numerical wind tunnel is 60m long, 24m wide, and 12m high. The virtual numerical wind tunnel is composed of a velocity inlet wall surface, a pressure outlet wall surface, and the remaining four wall boundaries;

[0211] 1.2) Process the vehicle model to be measured, only retain the outermost geometric surface and seal it;

[0212] 1.3) Place the vehicle to be measured at the end of the virtual numerical wind tunnel close to the velocity inlet wall surface. The distance from the front of the vehicle to the velocity inlet wall surface is about 3 times the vehicle body length, and the distances from both sides of the vehicle body to the two side wall surfaces are equal;

[0213] 1.4) Set 8 layers of prismatic grids with a boundary layer thickness of 2mm in the front side window area of the vehicle to be measured. The grid thickness growth rate between adjacent layers is 1.5. Set a grid dense area to improve the accuracy of obtaining the side window flow field data;

[0214] 1.5) Set the incoming flow velocity on the velocity inlet wall surface to 120 km / h, set the pressure outlet wall surface to 0 Pa, and set the remaining four wall surfaces and the vehicle body wall surface to non-slip wall surfaces;

[0215] 1.6) The turbulence model used in the numerical simulation calculation is the two-equation SST k-ω model, and the semi-implicit algorithm for pressure linked equations (SIMPLE) of the pressure equation is used to solve the pressure-velocity coupling term. In the transient simulation, detached eddy simulation is adopted. Some other settings of the solver are shown in the following table.

[0216]

[0217] 1.7) The simulation software was run, and the flow field data of the front side windows of 6 different vehicle models, totaling 46 groups of different shape conditions, were solved.

[0218] 2) Use the Wavenumber Frequency Spectrum Decomposition (WFS) method to process the time-domain flow field data on the surface of the front side window of the vehicle obtained in step 1) to obtain the frequency-domain flow field excitation data on the surface of the front side window of the vehicle;

[0219] The steps of using the WFS method to solve the frequency-domain flow field excitation on the surface of the front side window of the vehicle include:

[0220] 2.1) Perform Fourier transform on the flow field data of the front side window of the vehicle obtained by the CFD simulation software (i.e., the pressure pulsation p(x, y, t) in the time domain) to convert it to the frequency domain:

[0221] p * (x, y, ω) = ∫p(x, y, t)e (-iωt) dt (1)

[0222] In the formula, ω is the circular frequency, x and y are spatial quantities, t is the time, e is the natural constant, and i is the imaginary number.

[0223] 2.2) Perform Fourier transform on the pressure pulsation in the frequency domain in the x and y directions to obtain the pressure data in the wavenumber domain:

[0224]

[0225] In the formula, k x is the wavenumber in the x direction, and k y is the wavenumber in the y direction.

[0226] 2.3) In the two-dimensional wavenumber domain (k x , k y ), separate the two kinds of energy according to the difference in the propagation speeds of acoustic energy and hydrodynamic energy.

[0227] The distribution of acoustic energy is:

[0228]

[0229] The distribution of water kinetic energy is as follows:

[0230]

[0231] In the formula, c is the speed of sound and u is the air flow velocity.

[0232] 2.4) The sound energy obtained by decomposition is the sound pressure excitation on the front side window surface of the vehicle, and the water kinetic energy obtained by decomposition is the hydrodynamic pressure excitation on the front side window surface of the vehicle. The data is saved in the form of one-third octave results.

[0233] 3) Use geometric measurement means to measure the geometric parameter data of the vehicle.

[0234] The geometric parameters of the vehicle measured by geometric measurement means are: vehicle width, front wind-breaking angle, A-pillar width, A-pillar deflection, A-pillar side step height, A-pillar side inclination angle, A-pillar included angle, rearview mirror handle thickness, inner side included angle of the rearview mirror, outer side included angle of the rearview mirror, upper deflection of the rearview mirror, bottom deflection of the rearview mirror, distance between the rearview mirror and the side window, rearview mirror height, rearview mirror width. There are a total of 15 geometric parameters.

[0235] 4) Use machine learning technology to establish a Back Propagation Neural Network (BPNN) model.

[0236] The steps of establishing a BPNN model using machine learning technology include:

[0237] 4.1) Determine the basic structure of the neural network, that is, one input layer, three hidden layers and one output layer;

[0238] 4.2) Determine the number of neurons in each layer of the neural network. The number of neurons in the input layer is 15, and the number of neurons in each hidden layer is 40;

[0239] 4.3) Select the training function of the neural network. The training function is the trainlm function. Select the activation function for each hidden layer of the neural network. The activation function is the Sigmoid function. Determine that the representation method of the performance of the neural network model is the Mean Squared Error (MSE).

[0240] 5) The steps of training the neural network model using geometric parameter sample data as input and sound pressure excitation and hydrodynamic pressure excitation sample data as output include:

[0241] 5.1) Normalize the input and output sample data. Use the maximum-minimum normalization method to normalize the data to the interval [0, 1].

[0242] 5.2) Divide the sample data into a training sample set and a test sample set, where the number of the training sample set is 44 groups and the number of the test sample set is 2 groups. The training sample set is used to train the network model. The test sample set is used to verify the prediction results of the model and shall not be used for network training;

[0243] 5.3) Use the geometric parameter sample data in the training sample set as the input of the network model, and the sample data of the surface acoustic pressure excitation and hydrodynamic pressure excitation of the front side window of the vehicle as the output of the network model to train the network model;

[0244] 5.4) Input the geometric parameters of the vehicle in the two groups of test sample sets to obtain the corresponding excitation data predicted by the model. Compare the predicted data with the expected data of the test sample set, and calculate the average sound pressure level error and the relative error. The relative error is defined as follows:

[0245]

[0246] 6) Compare the errors between the predicted data of the front side window excitation obtained by the method proposed in the present invention and the expected data, as shown in the following table.

[0247]

[0248] It can be seen that whether it is the predicted acoustic pressure excitation or the hydrodynamic pressure excitation, the average sound pressure level error is within 2 dB, and the relative error is even controlled within 3%. It further verifies the rapidity and accuracy of a fast prediction method for the flow field excitation of the front side window of a vehicle based on machine learning.

Claims

1. A method for rapid prediction of flow field excitation of automobile front side windows based on machine learning, characterized in that: The following steps are involved: 1) Computational fluid dynamics simulation technology is used to calculate the time domain flow field data of the front side window surface of the car. 2) The time domain flow field data of the front side window surface of the automobile is processed by using the wave number frequency spectrum decomposition method to obtain the frequency domain flow field excitation data of the front side window surface of the automobile; 3) Measure the vehicle geometric parameter data and map it with the frequency domain flow field excitation data of the vehicle front side window surface; 4) Repeat steps 1) to 3) to construct a sample data set with vehicle geometric parameter data as input and frequency domain flow field excitation data as output; 5) Use machine learning technology to build a neural network model; 6) Using the sample data set to train the neural network model, a front side window flow field excitation prediction model is obtained; 7) Obtaining geometric parameter data of the vehicle to be tested and inputting them into the front side window flow field excitation prediction model to obtain the front side window flow field excitation prediction value of the vehicle to be tested.

2. The method for rapid prediction of flow field excitation of a front side window of an automobile based on machine learning according to claim 1, characterized in that: In step 1), the step of calculating the time domain flow field data of the front side window surface of the automobile includes: 1.1) A virtual numerical wind tunnel is established in the fluid dynamics simulation software. The virtual numerical wind tunnel is a cuboid, in which two walls serve as the velocity inlet wall and the pressure outlet wall; 1.2) Build a closed model of the vehicle to be tested; 1.3) Place the vehicle model to be tested in the virtual numerical wind tunnel and close to the velocity inlet wall; 1.4) Set a dense grid area in the front side window area of ​​the vehicle model to be tested; 1.5) The velocity inlet wall is set to the incoming wind speed, the pressure outlet wall is set to 0Pa, and the remaining four walls and the body wall are set to no-slip walls; 1.6) Run the simulation software to obtain the time domain flow field data of the front side window surface of the car.

3. The method for rapid prediction of flow field excitation of a front side window of an automobile based on machine learning according to claim 2 is characterized in that: The length of the virtual numerical wind tunnel is 15 times the length of the vehicle to be tested, the width is 13 times the width of the vehicle, and the height is 7 times the height of the vehicle.

4. The method for rapid prediction of flow field excitation of a front side window of an automobile based on machine learning according to claim 1, characterized in that: The distance between the vehicle model to be tested and the speed inlet wall is 3 times the body length, and the distance between the two sides of the body and the two side walls is 6 times the body width.

5. The method for rapid prediction of flow field excitation of a front side window of an automobile based on machine learning according to claim 1, characterized in that: In step 2), the frequency domain flow field excitation data of the front side window surface of the automobile includes the acoustic pressure excitation of the front side window surface of the automobile and the hydrodynamic pressure excitation of the front side window surface of the automobile.

6. The method for rapid prediction of flow field excitation of a front side window of an automobile based on machine learning according to claim 1 is characterized in that: In step 2), the step of processing the time domain flow field data of the front side window surface of the automobile includes: 2.1) Perform Fourier transform on the time domain flow field data p(x, y, t) of the front side window surface of the car to obtain: p * (x,y,ω)=∫p(x,y,t)e (-iωt) dt (1) In the formula, ω is the circular frequency, x, y are spatial quantities, t is time, e is a natural constant, and i is an imaginary number; p * (x, y, ω) is the flow field data after Fourier transformation; 2.2) Perform a second Fourier transform on the pressure pulsation in the frequency domain in the x and y directions to obtain the pressure data P in the wave number domain * (k x ,k y ,ω), that is: In the formula, k x is the wave number in the x direction, k y is the wave number in the y direction; 2.3) In the two-dimensional wave number domain (k x ,k y ), separating the acoustic energy and the water kinetic energy according to the difference in propagation speeds of the acoustic energy and the water kinetic energy; The distribution of acoustic energy is as follows: In the formula, c is the speed of sound; f is the frequency; u is the air flow velocity; The distribution of water kinetic energy is shown below: Where u is the air flow velocity. 2.4) The decomposed sound energy is used as the sound pressure excitation on the surface of the front side window of the car, and the decomposed water kinetic energy is used as the water dynamic pressure excitation on the surface of the front side window of the car.

7. The method for rapid prediction of flow field excitation of a front side window of an automobile based on machine learning according to claim 1 is characterized in that: The vehicle geometric parameters are measured by geometric measurement means.

8. The method for rapid prediction of flow field excitation of a front side window of an automobile based on machine learning according to claim 1 is characterized in that: The automobile geometric parameters include but are not limited to: vehicle width, front wind breaking angle, A-pillar width, A-pillar deflection, A-pillar side step height, A-pillar roll angle, A-pillar angle, rearview mirror handle thickness, rearview mirror inner side angle, rearview mirror outer side angle, rearview mirror upper deflection, rearview mirror bottom deflection, rearview mirror distance from side window, rearview mirror height, rearview mirror width and other parameters that can describe the geometric shape of the A-pillar and rearview mirror.

9. The method for rapid prediction of flow field excitation of a front side window of an automobile based on machine learning according to claim 1, characterized in that: In step 6), the step of training the neural network model using the sample data set includes: 6.1) Normalize the data of the sample data set and divide the normalized sample data set into a test set and a training set; 6.2) Select the training function of the neural network model and the activation function of each hidden layer; 6.3) Use the training set to train the neural network model; 6.4) Use the test set to test the neural network model. If the error is less than the preset value, the front side window flow field excitation prediction model is output. Otherwise, the sample data set is reconstructed.

10. The method for rapid prediction of vehicle front side window flow field excitation based on machine learning according to claim 1, characterized in that: The neural network model includes a back-propagation neural network model.

Images