Deep learning-based aircraft geometrical characteristic parameter inversion modeling method

Abstract

The invention provides an aircraft geometrical characteristic parameter inverse derivation modeling method based on deep learning. The aircraft geometrical characteristic parameter inverse derivation modeling method comprises the following steps: step 1, classifying aircrafts; 2, aiming at different types of aircrafts, analyzing appearance characteristics and aerodynamic characteristics, and extracting a geometric characteristic parameter set; step 3, constructing a pneumatic database for neural network model training; 4, establishing a deep neural network model; step 5, training the deep neural network model, wherein network output is aircraft geometric feature parameters; and step 6, inputting the designed aerodynamic characteristic information into the trained deep neural network model to obtain aircraft geometric characteristic parameters meeting design requirements. According to the neural network model from the aerodynamic parameters to the geometric shape characteristic parameters, the network output is not geometric design parameters during parametric modeling, but a geometric characteristic parameter set capable of generalizing the shape of a certain type of aircraft, so that the network model has the shape generalization ability.

Description

Technical Field

[0001] The present invention relates to a method for inverse derivative modeling of aircraft geometric characteristic parameters based on deep learning. Background Art

[0002] At present, the demand for the flight quality, aerodynamic performance and flight efficiency of aircraft is increasing day by day, and multiple constraints are often generated in the aerodynamic shape design stage. Aircraft aerodynamic design is mainly divided into forward design and inverse design methods. Forward design obtains an aerodynamic shape that meets the design indicators by changing the aircraft shape multiple times and performing CFD calculations, which undoubtedly consumes a lot of computing resources and time. The inverse design method directly solves the shape that meets the design requirements through the designed aerodynamic characteristic information, and directly establishes the mapping of aerodynamic parameters to geometric shapes, which greatly improves the design efficiency. The inverse design method based on machine learning has become a current research hotspot. However, most of the current methods are suitable for airfoil inverse design, wing inverse design or single shape optimization problems. The input end of the machine learning model is the design parameters of the shape parameter modeling, rather than the geometric feature parameters. Therefore, it is not applicable to the inverse design problem of the new shape. Based on this, it is necessary to provide a method for inverse modeling of aircraft geometric feature parameters based on deep learning, and establish a mapping model from aerodynamic characteristic information to geometric feature parameters suitable for a certain type of aircraft, rather than limited to a single shape, that is, the constructed model has the ability to generalize the shape. Summary of the invention

[0003] Purpose of the invention: The technical problem to be solved by the present invention is to provide a method for inverse modeling of aircraft geometric characteristic parameters based on deep learning in view of the deficiencies of the prior art. The present invention directly establishes a mapping model from the aerodynamic characteristics of a three-dimensional aircraft to the geometric shape characteristic parameters through a deep learning method. The model can predict the geometric shape characteristic parameters of the aircraft that meet the design requirements based on the designed aerodynamic characteristic information, thereby shortening the overall shape design cycle of the aircraft.

[0004] The method of the present invention comprises the following steps:

[0005] Step 1, classify the aircraft according to the appearance characteristics and aerodynamic characteristics; currently, there are various ways to classify aircraft, such as civil aircraft and military aircraft according to their purpose, subsonic aircraft and supersonic aircraft according to their speed, propeller aircraft and jet aircraft according to their engine type, etc. Since the present invention aims to perform inverse derivative modeling on the geometric characteristic parameters of the aircraft, and since aircraft with different aerodynamic layouts have great differences in appearance, they need to be classified according to the characteristics of the geometric appearance and the characteristics of the aerodynamic characteristics, in order to ensure that the geometric characteristic parameters extracted in step 2 are suitable for this type of aircraft. The classification methods are currently artificially divided. For example, for hypersonic aircraft, they are divided into lifting bodies, wing-body combinations, rotating bodies, and waveriders according to their aerodynamic layout characteristics.

[0006] Step 2: Analyze the appearance characteristics and aerodynamic characteristics of different types of aircraft and extract the set of geometric feature parameters;

[0007] Step 3, constructing a pneumatic database for neural network model training;

[0008] Step 4, building a deep neural network model;

[0009] Step 5, deep neural network training and testing: Use the aerodynamic database constructed in step 3 to train the deep neural network model established in step 4, the network input is the aerodynamic characteristics information of the aircraft, and the network output is the geometric characteristic parameters of the aircraft;

[0010] Step 6: Input the designed aerodynamic characteristics information into the trained deep neural network model to predict the geometric characteristic parameters of the aircraft that meet the design requirements.

[0011] In step 2,

[0012] The characteristic of geometric characteristic parameters is that they combine the geometric shape characteristics of the aircraft with the aerodynamic characteristics. For example, for the wing-body combination, the fuselage mostly adopts a slender body shape, and the wing adopts a large swept trapezoidal wing. For the trapezoidal wing, its relative plane shape can be determined by the aspect ratio, root-tip ratio, and leading edge sweep angle. In order to meet the requirements of high lift-to-drag ratio at high speed and high lift at low speed, the wing is accompanied by a winglet design, and geometric characteristic parameters such as wing aspect ratio, wing sweep angle, fuselage slenderness ratio, and winglet sweep angle can be extracted;

[0013] For a body of rotation, it is formed by a generatrix (smooth curve or broken line) rotating around a certain axis, and the axis is called the body axis of the body of rotation. Any cross section of the body of rotation is circular, and its appearance features are a pointed head, a large slenderness ratio, a large swept angle of the wing, and a small aspect ratio. A tailless aerodynamic layout is often used. In addition, the position of the wing has a more critical influence on the aerodynamic characteristics of the body of rotation. Therefore, geometric characteristic parameters such as slenderness ratio, contraction ratio, taper ratio, ratio of the maximum cross-sectional area of ​​the fuselage to the wing area, and ratio of the length from the leading edge of the wing root chord to the apex of the fuselage head to the average aerodynamic chord length of the wing can be extracted;

[0014] For the waverider, its shape design feature is to reversely solve its geometric shape through the known supersonic flow field. Therefore, for the waverider configuration, it is necessary to combine the flow field design and the geometric shape to extract the geometric characteristic parameters. The geometric characteristic parameters such as the aspect ratio, the height-to-width ratio of the symmetry plane, and the volume ratio can be extracted.

[0015] In step 2, for a lifting body aircraft, extract the Kuchman coefficient τ and volume fraction I vol , Y-direction projection area S Y With reference area S ref Ratio S Y / S ref , Z-direction projection area S Z Ratio to reference area S Z / S ref , Wetting area S wet With plane area S plan Ratio S wet / S plan As geometric characteristic parameters; the X-direction projection area of ​​the aircraft is taken as the reference area S ref , the Z-direction projection area is taken as the plane area S plan ; For the definition of the coordinate system, the origin of the coordinate system is the center of mass O of the aircraft, the direction from the center of mass to the nose is the positive direction of the X-axis, the Z-axis is located in the longitudinal symmetry plane of the aircraft, perpendicular to the X-axis, and the upward direction is the positive direction of the Z-axis. The Y-axis is perpendicular to the OXZ plane, and the positive direction of the Y-axis is determined according to the right-hand rule.

[0016] Step 2 also includes: using single parameter sensitivity analysis to analyze the extracted geometric feature parameter A, plotting the aerodynamic characteristics of the aircraft as the parameter A changes, using parameter A as the horizontal coordinate and the aerodynamic characteristics as the vertical coordinate, analyzing the impact of parameter A on the aerodynamic characteristics of the aircraft, such as whether the lift coefficient will increase or decrease, or whether it has no effect on the lift coefficient. If parameter A increases or decreases the lift coefficient, parameter A is extracted into the geometric feature parameter set; if there is no effect, it is not extracted.

[0017] Step 3 includes:

[0018] Step 3-1, establishing a geometric shape library: using a parametric modeling method to generate more than two aircraft shapes, that is, using more than two parameters to describe the shape when designing the geometric shape of the aircraft, such as length, width, height, etc., and then modifying the aircraft shape by changing the size of the parameter value, thereby generating more than two aircraft shapes; and extracting the geometric feature parameters of each aircraft according to the geometric feature parameter set extracted in step 2;

[0019] Step 3-2, establish an aerodynamic database: obtain the aerodynamic characteristic data of each aircraft, including lift coefficient, drag coefficient, and moment coefficient, through engineering estimation methods, CFD simulation or wind tunnel test methods.

[0020] Step 3-1 includes: firstly, establishing a reference model according to the geometric design parameters of the parametric modeling of the aircraft, the modeling method adopts the modeling method of the control section shape, dividing the aircraft from the head to the tail into more than two sections, and the sections are called control sections; firstly designing the height, aspect ratio and section geometry of each section to obtain a series of reference shapes; the control parameters of each control section are the design variables of the parametric modeling;

[0021] Secondly, since multiple control sections are used to parameterize the aircraft, multiple design variables will be generated. By changing the values ​​of each design variable, different aircraft shapes can be generated. If all permutations and combinations of design variables are used, the sample size generated will be too large. Therefore, the orthogonal test design method is used to generate the design space of samples. The core of the orthogonal test is the orthogonal table, which is used to generate the sample space. m (q n ) represents an orthogonal array, L m (q n ) means that for an orthogonal experiment with n design variables and q levels for each design variable, a total of m samples are designed and generated; the meaning of level is the number of all possible values ​​of a design variable, for example, the possible values ​​of the aspect ratio of section 1 are [0.3, 0.4, 0.5, 0.6, 0.7, 0.8], then the factor has six levels; further, since it is impossible to ensure that the number of levels of each design variable is the same, a mixed level orthogonal table is used to design the sample space, The expression of the mixed level orthogonal array is Lm(q1n 1 ×q2n 2 ), It means there are n1 design variables with q1 levels. It means there are n2 design variables with q2 levels, L 49 (4 6 ×7 1) as an example, it means that there are 6 design variables with 4 levels and 1 design variable with 7 levels, and a total of 49 samples are generated. The values ​​of the parameters of aircraft with different shapes are recorded in the orthogonal table, which is called the parameter design table;

[0022] The OpenVSP software is used to generate aircraft with different shapes. According to the obtained parameter design table, a parameter design file Design file (in the format of *.des) is generated to control the cross-sectional shape. The parameter design file is used to adjust the geometric shape of the benchmark model. The content includes the number of design parameters and the specific value of each design parameter. A script file Scriptfile (in the format of *.vspscript) is written and called by the OpenVSP software to batch generate more than two aircraft shapes.

[0023] In step 3-2, for a lifting body aircraft, the lifting body head cross-sectional shape, head cross-sectional height, head cross-sectional aspect ratio, tail cross-sectional shape, tail cross-sectional height, and tail cross-sectional aspect ratio are used as design parameters. The cross-sectional shapes include ellipse, triangle, and rhombus. More than two lifting body models are generated by changing the design parameters.

[0024] The aerodynamic characteristics of each lifting body shape are calculated using an engineering estimation method: the aircraft shape is divided into two or more surface elements (generally more surface elements), and the pressure coefficient C of each surface element is calculated by solving p , and then integrate to obtain the aerodynamic characteristics data of the entire aircraft. When solving the pressure on the surface element, first divide the surface element into the windward surface and the leeward surface according to the positive and negative impact angle. The impact angle calculation formula is:

[0025]

[0026] in is the outward normal vector of the surface element; is the incoming flow velocity vector; δ greater than 0 indicates the windward impact angle, and less than 0 indicates the leeward impact angle;

[0027] For the windward side, the selected pressure coefficient C p The calculation formula is the modified Newton theory:

[0028]

[0029] Where γ is the specific heat ratio, Ma ∞ is the Mach number;

[0030] For the leeward side, the selected pressure coefficient C p Calculation formula: Prandtl-Meyer formula:

[0031]

[0032] Then, the pressure coefficient C of each panel is used p , unit normal vector Area ΔA and aircraft reference area S ref , integrate to get the aerodynamic coefficients in the X, Y, and Z directions of the aircraft;

[0033] Unit Normal Vector The calculation formula is:

[0034]

[0035] Where |N| is The module length, N x for The component of the vector in the X direction, N y for The component of the vector in the Y direction, N z for The component of the vector in the Z direction; for The component in the X direction, for ni The component in the Y direction, for Component in the Z direction;

[0036] The integral formula is:

[0037] F x =∑C p n x ΔA / S ref

[0038] F y =∑C p n y ΔA / S ref

[0039] F z =∑C p n z ΔA / S ref

[0040] Among them, n x for The value of n y for The value of n z for The value of F x is the aerodynamic coefficient in the X direction, F y is the aerodynamic coefficient in the Y direction, F z is the aerodynamic coefficient in the Z direction.

[0041] Finally, the lift coefficient C L , drag coefficient C D , torque coefficient C M :

[0042] C L =-F x sinα+F z cosα

[0043] C D =F x cosαcosβ+F y sinβ+F z sinαcosβ

[0044]

[0045]

[0046] Where L ref is the characteristic length, the subscript cg represents the center of gravity, α is the angle of attack, and β is the sideslip angle; C Mx , C My , C Mz are the moment coefficients of the aircraft around the X-axis, Y-axis, and Z-axis respectively; x, y, z are the coordinates of the location of the surface element, x cg is the coordinate value of the center of gravity of the aircraft on the X axis, y cg is the coordinate value of the center of gravity of the aircraft on the Y axis, z cg is the coordinate value of the center of gravity of the aircraft on the Z axis;

[0047] Based on the engineering estimation method, the lift coefficient, drag coefficient and pitch moment coefficient of each aircraft in the geometric shape library are calculated at different Mach numbers and angles of attack, and all the calculated aerodynamic data are used as an aerodynamic database.

[0048] In step 4, a one-dimensional convolutional neural network is used as a deep neural network model. The one-dimensional convolutional neural network includes a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer can extract the features of the input data; the pooling layer is used to reduce the data scale, extract the information after convolution, and improve the calculation speed, reduce the number of parameters to prevent overfitting; the convolutional layer and the pooling layer appear alternately, continuously learn the features of the input data, and finally pass the features to the fully connected layer to obtain the network output.

[0049] In step 5, the aerodynamic database constructed in step 2 is randomly divided into a training set, a validation set, and a test set. The lift coefficient, drag coefficient, lift-to-drag ratio, pitch moment coefficient, and angle of attack of each shape are used as the input of the one-dimensional convolutional neural network, and the corresponding geometric characteristic parameters of the shape are used as the network output. During the network training process, the learning rate is set to 2×10-4 , the optimizer uses Adam, and the loss function uses mean square error MSE:

[0050]

[0051] Where n is the sample size, Y i is the true value, is the network prediction value.

[0052] The present invention also provides an electronic device, characterized in that it includes a processor and a memory, wherein the memory stores program code, and when the program code is executed by the processor, the processor executes the steps of the described method.

[0053] Beneficial effect: The network output of the neural network model from aerodynamic parameters to geometric shape characteristic parameters established by the present invention is not the geometric design parameters during parametric modeling, but a set of geometric characteristic parameters that can generalize the shape of a certain type of aircraft, which enables the network model to have shape generalization capabilities. Taking a lifting body as an example, after obtaining the geometric characteristic parameters of a lifting body aircraft that meet the aerodynamic design requirements through neural network prediction, the designer can establish a three-dimensional model of the lifting body aircraft through different parameterization methods, provided that the aerodynamic shape characteristics of the lifting body aircraft are met. The design variables can be determined by the design requirements, and it is only necessary to ensure that the geometric characteristic parameter values ​​of the model are consistent with the predicted geometric characteristic parameter values. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0055] Figure 1 It is a schematic diagram of coordinate system definition.

[0056] Figure 2 It is a schematic diagram of the lifting body model generated by the parametric method.

[0057] Figure 3 This is a schematic diagram of the surface pressure distribution of the X-37.

[0058] Figure 4a , Figure 4b , Figure 4c It is a schematic diagram comparing the aerodynamic characteristics estimated by engineering and the CDF results.

[0059] Figure 5 It is a one-dimensional convolutional neural network structure diagram.

[0060] Figure 6a , Figure 6b It is a diagram of network training loss.

[0061] Figure 7a , Figure 7b , Figure 7c , Figure 7d , Figure 7e It is a schematic diagram of the network's prediction results of geometric feature parameters for training samples and test samples.

[0062] Figure 8 It is a flowchart for constructing an aircraft shape library. DETAILED DESCRIPTION

[0063] The present invention provides a method for inverse modeling of aircraft geometric characteristic parameters based on deep learning, comprising the following steps:

[0064] Step 1, classify the aircraft according to the appearance characteristics and aerodynamic characteristics; currently, there are various ways to classify aircraft, such as civil aircraft and military aircraft according to their purpose, subsonic aircraft and supersonic aircraft according to their speed, propeller aircraft and jet aircraft according to their engine type, etc. Since the present invention aims to perform inverse derivative modeling on the geometric characteristic parameters of the aircraft, and since aircraft with different aerodynamic layouts have great differences in appearance, they need to be classified according to the characteristics of the geometric appearance and the characteristics of the aerodynamic characteristics, in order to ensure that the geometric characteristic parameters extracted in step 2 are suitable for this type of aircraft. The classification methods are currently artificially divided. For example, for hypersonic aircraft, they are divided into lifting bodies, wing-body combinations, rotating bodies, and waveriders according to their aerodynamic layout characteristics.

[0065] Step 2: Analyze the appearance characteristics and aerodynamic characteristics of different types of aircraft and extract the set of geometric feature parameters;

[0066] Step 3, constructing a pneumatic database for neural network model training;

[0067] Step 4, building a deep neural network model;

[0068] Step 5, deep neural network training and testing: Use the aerodynamic database constructed in step 3 to train the deep neural network model established in step 4, the network input is the aerodynamic characteristics information of the aircraft, and the network output is the geometric characteristic parameters of the aircraft;

[0069] Step 6: Input the designed aerodynamic characteristics information into the trained deep neural network model to predict the geometric characteristic parameters of the aircraft that meet the design requirements.

[0070] In step 2,

[0071] The characteristic of geometric characteristic parameters is that they combine the geometric shape characteristics of the aircraft with the aerodynamic characteristics. For example, for the wing-body combination, the fuselage mostly adopts a slender body shape, and the wing adopts a large swept trapezoidal wing. For the trapezoidal wing, its relative plane shape can be determined by the aspect ratio, root-tip ratio, and leading edge sweep angle. In order to meet the requirements of high lift-to-drag ratio at high speed and high lift at low speed, the wing is accompanied by a winglet design, and geometric characteristic parameters such as wing aspect ratio, wing sweep angle, fuselage slenderness ratio, and winglet sweep angle can be extracted;

[0072] For a body of rotation, it is formed by a generatrix (smooth curve or broken line) rotating around a certain axis, and the axis is called the body axis of the body of rotation. Any cross section of the body of rotation is circular, and its appearance features are a pointed head, a large slenderness ratio, a large swept angle of the wing, and a small aspect ratio. A tailless aerodynamic layout is often used. In addition, the position of the wing has a more critical influence on the aerodynamic characteristics of the body of rotation. Therefore, geometric characteristic parameters such as slenderness ratio, contraction ratio, taper ratio, ratio of the maximum cross-sectional area of ​​the fuselage to the wing area, and ratio of the length from the leading edge of the wing root chord to the apex of the fuselage head to the average aerodynamic chord length of the wing can be extracted;

[0073] For the waverider, its shape design feature is to reversely solve its geometric shape through the known supersonic flow field. Therefore, for the waverider configuration, it is necessary to combine the flow field design and the geometric shape to extract the geometric characteristic parameters. The geometric characteristic parameters such as the aspect ratio, the height-to-width ratio of the symmetry plane, and the volume ratio can be extracted.

[0074] In step 2, for a lifting body aircraft, extract the Kuchman coefficient τ and volume fraction I vol , Y-direction projection area S Y With reference area S ref Ratio S Y / S ref , Z-direction projection area S Z Ratio to reference area S Z / S ref , Wetting area S wet With plane area S plan Ratio S wet / S plan As geometric characteristic parameters; the X-direction projection area of ​​the aircraft is taken as the reference area S ref , the Z-direction projection area is taken as the plane area S plan ; For the definition of the coordinate system, the origin of the coordinate system is the center of mass O of the aircraft, the direction from the center of mass to the nose is the positive direction of the X-axis, the Z-axis is located in the longitudinal symmetry plane of the aircraft, perpendicular to the X-axis, and the upward direction is the positive direction of the Z-axis. The Y-axis is perpendicular to the OXZ plane, and the positive direction of the Y-axis is determined according to the right-hand rule.

[0075] Step 2 also includes: using single parameter sensitivity analysis to analyze the extracted geometric feature parameter A, plotting the aerodynamic characteristics of the aircraft as the parameter A changes, using parameter A as the horizontal coordinate and the aerodynamic characteristics as the vertical coordinate, analyzing the impact of parameter A on the aerodynamic characteristics of the aircraft, such as whether the lift coefficient will increase or decrease, or whether it has no effect on the lift coefficient. If parameter A increases or decreases the lift coefficient, parameter A is extracted into the geometric feature parameter set; if there is no effect, it is not extracted.

[0076] Step 3 includes:

[0077] Step 3-1, establishing a geometric shape library: using a parametric modeling method to generate more than two aircraft shapes, that is, using more than two parameters to describe the shape when designing the geometric shape of the aircraft, such as length, width, height, etc., and then modifying the aircraft shape by changing the size of the parameter value, thereby generating more than two aircraft shapes; and extracting the geometric feature parameters of each aircraft according to the geometric feature parameter set extracted in step 2;

[0078] Step 3-2, establish an aerodynamic database: obtain the aerodynamic characteristic data of each aircraft, including lift coefficient, drag coefficient, and moment coefficient, through engineering estimation methods, CFD simulation or wind tunnel test methods.

[0079] Step 3-1 includes: firstly, establishing a reference model according to the geometric design parameters of the parametric modeling of the aircraft, the modeling method adopts the modeling method of the control section shape, dividing the aircraft from the head to the tail into more than two sections, and the sections are called control sections; firstly designing the height, aspect ratio and section geometry of each section to obtain a series of reference shapes; the control parameters of each control section are the design variables of the parametric modeling;

[0080] Secondly, since multiple control sections are used to parameterize the aircraft, multiple design variables will be generated. By changing the values ​​of each design variable, different aircraft shapes can be generated. If all permutations and combinations of design variables are used, the sample size generated will be too large. Therefore, the orthogonal test design method is used to generate the design space of samples. The core of the orthogonal test is the orthogonal table, which is used to generate the sample space. m (q n ) represents an orthogonal array, L m (q n ) means that for an orthogonal experiment with n design variables and q levels for each design variable, a total of m samples are designed and generated; the meaning of level is the number of all possible values ​​of a design variable, for example, the possible values ​​of the aspect ratio of section 1 are [0.3, 0.4, 0.5, 0.6, 0.7, 0.8], then the factor has six levels; further, since it is impossible to ensure that the number of levels of each design variable is the same, a mixed level orthogonal table is used to design the sample space, The expression of the mixed level orthogonal array is It means there are n1 design variables with q1 levels. It means there are n2 design variables with q2 levels, L 49 (4 6 ×7 1 ) as an example, it means that there are 6 design variables with 4 levels and 1 design variable with 7 levels, and a total of 49 samples are generated. The values ​​of the parameters of aircraft with different shapes are recorded in the orthogonal table, which is called the parameter design table;

[0081] The OpenVSP software is used to generate aircraft with different shapes. Based on the obtained parameter design table, a parameter design file (in the *.des format) is generated to control the cross-sectional shape. The parameter design file is used to adjust the geometric shape of the reference model. The content includes the number of design parameters and the specific value of each design parameter. A script file (in the *.vspscript format) is written and called by the OpenVSP software to batch generate more than two aircraft shapes. Figure 8 The figure shows the process of building an aircraft shape library.

[0082] In step 3-2, for a lifting body aircraft, the lifting body head cross-sectional shape, head cross-sectional height, head cross-sectional aspect ratio, tail cross-sectional shape, tail cross-sectional height, and tail cross-sectional aspect ratio are used as design parameters. The cross-sectional shapes include ellipse, triangle, and rhombus. More than two lifting body models are generated by changing the design parameters.

[0083] The aerodynamic characteristics of each lifting body shape are calculated using an engineering estimation method: the aircraft shape is divided into two or more surface elements (generally more surface elements), and the pressure coefficient C of each surface element is calculated by solving p , and then integrate to obtain the aerodynamic characteristics data of the entire aircraft. When solving the pressure on the surface element, first divide the surface element into the windward surface and the leeward surface according to the positive and negative impact angle. The impact angle calculation formula is:

[0084]

[0085] in is the outward normal vector of the surface element; is the incoming flow velocity vector; δ greater than 0 indicates the windward impact angle, and less than 0 indicates the leeward impact angle;

[0086] For the windward side, the selected pressure coefficient C p The calculation formula is the modified Newton theory:

[0087]

[0088] Where γ is the specific heat ratio, Ma ∞ is the Mach number;

[0089] For the leeward side, the selected pressure coefficient C p Calculation formula: Prandtl-Meyer formula:

[0090]

[0091] Then, the pressure coefficient C of each panel is used p , unit normal vector Area ΔA and aircraft reference area S ref , integrate to get the aerodynamic coefficients in the X, Y, and Z directions of the aircraft;

[0092] Unit Normal Vector The calculation formula is:

[0093]

[0094] Where |N| is The module length, N x for The component of the vector in the X direction, N y for The component of the vector in the Y direction, N z for The component of the vector in the Z direction; for The component in the X direction, for The component in the Y direction, for Component in the Z direction;

[0095] The integral formula is:

[0096] F x =∑C p n x ΔA / S ref

[0097] F y =∑C p n y ΔA / S ref

[0098] F z =∑C p n z ΔA / S ref

[0099] Among them, n x for The value of n y for The value of n z for The value of F x is the aerodynamic coefficient in the X direction, F y is the aerodynamic coefficient in the Y direction, F z is the aerodynamic coefficient in the Z direction.

[0100] Finally, the lift coefficient C L , drag coefficient C D , torque coefficient C M :

[0101] C L =-F x sinα+F z cosα

[0102] C D =F x cosαcosβ+F y sinβ+F z sinαcosβ

[0103]

[0104] Where L ref is the characteristic length, the subscript cg represents the center of gravity, α is the angle of attack, and β is the sideslip angle; C Mx , C My , C Mz are the moment coefficients of the aircraft around the X-axis, Y-axis, and Z-axis respectively; x, y, z are the coordinates of the location of the surface element, x cg is the coordinate value of the center of gravity of the aircraft on the X axis, y cg is the coordinate value of the center of gravity of the aircraft on the Y axis, z cg is the coordinate value of the center of gravity of the aircraft on the Z axis;

[0105] Based on the engineering estimation method, the lift coefficient, drag coefficient and pitch moment coefficient of each aircraft in the geometric shape library are calculated at different Mach numbers and angles of attack, and all the calculated aerodynamic data are used as an aerodynamic database.

[0106] In step 4, a one-dimensional convolutional neural network is used as a deep neural network model. The one-dimensional convolutional neural network includes a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer can extract the features of the input data; the pooling layer is used to reduce the data scale, extract the information after convolution, and improve the calculation speed, reduce the number of parameters to prevent overfitting; the convolutional layer and the pooling layer appear alternately, continuously learn the features of the input data, and finally pass the features to the fully connected layer to obtain the network output.

[0107] In step 5, the aerodynamic database constructed in step 2 is randomly divided into a training set, a validation set, and a test set. The lift coefficient, drag coefficient, lift-to-drag ratio, pitch moment coefficient, and angle of attack of each shape are used as the input of the one-dimensional convolutional neural network, and the corresponding geometric characteristic parameters of the shape are used as the network output. During the network training process, the learning rate is set to 2×10 -4 , the optimizer uses Adam, and the loss function uses mean square error MSE:

[0108]

[0109] Where n is the sample size, Y i is the true value, is the network prediction value.

[0110] In a specific embodiment of the present invention, a method for inverse modeling of aircraft geometric characteristic parameters based on deep learning is provided, comprising the following steps:

[0111] Step 1: Taking the lifting body aircraft as an example, extract the Kuchman coefficient τ and volume fraction I vol , the ratio of the Y-direction projection area to the reference area S Y / S ref , the ratio of the Z-direction projection area to the reference area S Z / S ref , the ratio of wetted area to plane area S wet / S plan As the geometric feature parameter set, the X-direction projection area of ​​the aircraft is taken as the reference area S ref , the Z-direction projection area is taken as the plane area S plan . Figure 1 This is the coordinate system definition method of this embodiment.

[0112] Step 2, establishing a lifting body shape database by a parameterized method. In this embodiment, the lifting body head cross-sectional shape, head cross-sectional height, head cross-sectional aspect ratio, tail cross-sectional shape, tail cross-sectional height, and tail cross-sectional aspect ratio are used as design parameters. The cross-sectional shapes include ellipse, triangle, diamond, etc. Multiple lifting body models are generated by changing the design parameters. Figure 2 The generated partial lifting body model is shown.

[0113] This embodiment uses an engineering estimation method to calculate the aerodynamic characteristics data of each lifting body shape. That is, the aircraft shape is divided into several surfaces, and the aerodynamic characteristics data of the entire aircraft are obtained by solving the local surface pressure coefficient and then integrating. For the windward surface, the pressure coefficient calculation formula selected in this embodiment is the modified Newton theory:

[0114]

[0115] Where γ is the specific heat ratio, Ma ∞ is the Mach number, and δ is the impact angle. The Prandtl-Meyer formula is selected for the leeward side:

[0116]

[0117] Figure 3 This is the surface pressure distribution of the X-37 calculated using the above formula. In the figure, Cp_ep represents the pressure coefficient obtained by the engineering estimation method. Figure 4a , Figure 4b , Figure 4c The comparison between the engineering estimate and CFD results of X-37. EEM stands for engineering estimate, C L , C D , L / D are lift coefficient, drag coefficient and lift-to-drag ratio respectively, and AOA is angle of attack. It can be seen that the engineering estimation is highly consistent with the CFD results, so this paper adopts the engineering estimation method to calculate the aerodynamic characteristics of the lifting body.

[0118] In this embodiment, a total of 450 lifting body shapes are generated, and the lift coefficient, drag coefficient, lift-to-drag ratio and pitch moment coefficient of each lifting body shape are calculated. The Mach number of the calculation condition is 3, the angle of attack ranges from -10° to 20°, and the step length is 2°.

[0119] Step 3: Building a deep neural network model. This embodiment uses a one-dimensional convolutional neural network as a deep neural network model. Figure 5 The one-dimensional convolutional neural network consists of a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer can extract the features of the input data. The pooling layer reduces the data size, extracts the main information after convolution, and increases the operation speed, reduces the number of parameters to prevent overfitting. The convolutional layer and pooling layer in CNN usually appear alternately, continuously learning the features of the input data, and finally passing the features to the fully connected layer to obtain the network output.

[0120] Step 4, neural network training. The aerodynamic database constructed in step 2 is divided into training set, validation set and test set at a ratio of 8:1:1, that is, 360 training shapes, 45 validation set shapes and 45 test set shapes. The lift coefficient, drag coefficient, lift-to-drag ratio, pitch moment coefficient and angle of attack of each shape are used as the input of the one-dimensional convolutional neural network, and the corresponding geometric characteristic parameters of the shape are used as the network output. During the network training process, the learning rate is set to 2×10 -4 , the optimizer uses Adam, and the loss function uses Mean Square Error (MSE):

[0121]

[0122] Where n is the sample size, Y i is the true value, is the network prediction value. Figure 6a , Figure 6b is the training set loss and validation set loss of network training. Epoch in the figure is the number of training iterations.

[0123] Step 5, network training results and testing. The network trained in step 4 has converged, so it can be used for testing. Use the test set data divided in step 4 to test the network, and input the lift coefficient, drag coefficient, lift-to-drag ratio, pitch moment coefficient and angle of attack information of each shape. Figure 7a , Figure 7b , Figure 7c , Figure 7d , Figure 7e The network's prediction results for the geometric characteristic parameters of the lifting body shape of the training set and the test set are shown. The left side shows the network's prediction results for the training samples, and the right side shows the network's prediction results for the test samples. Figure 7a , Figure 7b , Figure 7c , Figure 7d , Figure 7e The horizontal axis is the true value, the vertical axis is the predicted value, each point represents a sample, the closer the point is to the 45° line, the closer the predicted value is to the true value, and Kuchemann's tau is the Kuchemann coefficient. For the data in the training set, all sample points are on the 45° line. For the test data, some points deviate from the 45° line. Among all the test samples, the average relative errors for the prediction of each geometric parameter are 0.889%, 0.489%, 2.322%, 0.821%, and 0.754%, and the maximum relative errors are 5.341%, 3.301%, 8.093%, 6.011%, and 2.414%, respectively, all of which are less than 10%. It can be seen that the network has a high prediction accuracy for the test samples.

[0124] In another embodiment of the present invention, in step 3, an aerodynamic database is constructed from historical aerodynamic data. When constructing the shape database, a free-form deformation method (FFD) can also be used to generate different shapes, and a CFD method and a wind tunnel test method can be used to obtain aerodynamic characteristic data.

[0125] In another embodiment of the present invention, in step 4, other proxy models are used to replace the deep neural network model, such as a polynomial response surface model, a Kriging model, a radial basis function, a support vector machine, etc.

[0126] The present invention provides a method for inverse modeling of aircraft geometric feature parameters based on deep learning. There are many methods and ways to implement the technical solution. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention. All components not specified in this embodiment can be implemented using existing technologies.

Claims

1. A method for inverse modeling of aircraft geometric characteristic parameters based on deep learning, characterized in that: The following steps are involved: Step 1, classify the aircraft according to their appearance and aerodynamic characteristics; Step 2: Analyze the appearance characteristics and aerodynamic characteristics of different types of aircraft and extract the set of geometric feature parameters; Step 3, constructing a pneumatic database for neural network model training; Step 4, building a deep neural network model; Step 5, deep neural network training and testing: Use the aerodynamic database constructed in step 3 to train the deep neural network model established in step 4, the network input is the aerodynamic characteristics information of the aircraft, and the network output is the geometric characteristic parameters of the aircraft; Step 6: Input the designed aerodynamic characteristics information into the trained deep neural network model to predict the geometric characteristic parameters of the aircraft that meet the design requirements.

2. The method according to claim 1, characterized in that In step 2, for the wing-body assembly, the wing aspect ratio, wing sweep angle, fuselage slenderness ratio, and wing sweep angle are extracted as geometric feature parameters; For the rotational body, the slenderness ratio, contraction ratio, taper ratio, ratio of the maximum cross-sectional area of ​​the fuselage to the wing area, and the ratio of the length from the leading edge of the wing root chord to the top of the fuselage head to the average aerodynamic chord length of the wing are extracted as geometric characteristic parameters; For the waverider, the aspect ratio, the height-to-width ratio of the symmetry plane, and the volume ratio are extracted as geometric characteristic parameters.

3. The method according to claim 2, characterized in that In step 2, for a lifting body aircraft, extract the Kuchman coefficient τ and volume fraction I vol , Y-direction projection area S Y With reference area S ref Ratio S Y / S ref , Z-direction projection area S Z Ratio to reference area S Z / S ref , Wetting area S wet With plane area S plan Ratio S wet / S plan As geometric feature parameters; The X-direction projection area of ​​the aircraft is taken as the reference area S ref , the Z-direction projection area is taken as the plane area S plan ; For the definition of the coordinate system, the origin of the coordinate system is the center of mass O of the aircraft, the direction from the center of mass to the nose is the positive direction of the X-axis, the Z-axis is located in the longitudinal symmetry plane of the aircraft, perpendicular to the X-axis, and the upward direction is the positive direction of the Z-axis. The Y-axis is perpendicular to the OXZ plane, and the positive direction of the Y-axis is determined according to the right-hand rule.

4. The method according to claim 3, characterized in that Step 2 also includes: using single parameter sensitivity analysis to analyze the extracted geometric characteristic parameter A, plotting the aerodynamic characteristics of the aircraft as the parameter A changes, using parameter A as the horizontal coordinate and the aerodynamic characteristics as the vertical coordinate, analyzing the impact of parameter A on the aerodynamic characteristics of the aircraft, if parameter A increases or decreases the lift coefficient, then extracting parameter A into the geometric characteristic parameter set, if there is no impact, then not extracting it.

5. The method according to claim 4, characterized in that Step 3 includes: Step 3-1, establishing a geometric shape library: using a parametric modeling method to generate more than two aircraft shapes, that is, using more than two parameters to describe the shape when designing the geometric shape of the aircraft, and then modifying the aircraft shape by changing the size of the parameter value, thereby generating more than two aircraft shapes; and extracting the geometric feature parameters of each aircraft according to the geometric feature parameter set extracted in step 2; Step 3-2, establish an aerodynamic database: obtain the aerodynamic characteristic data of each aircraft, including lift coefficient, drag coefficient, and moment coefficient.

6. The method according to claim 5, characterized in that Step 3-1 includes: establishing a reference model according to the geometric design parameters of the parametric modeling of the aircraft, the modeling method adopts the modeling method of the control section shape, dividing the aircraft from the head to the tail into more than two sections, and the sections are called control sections; firstly designing the height, aspect ratio and section geometry of each section to obtain a series of reference shapes; the control parameters of each control section are the design variables of the parametric modeling; By changing the values ​​of various design variables, different aircraft shapes are generated: the orthogonal experimental design method is used to generate the design space of the samples, and L m (q n ) represents an orthogonal array, L m (q n ) means that for an orthogonal experiment with n design variables and q levels for each design variable, a total of m samples are generated; The sample space is designed using a mixed level orthogonal array. The expression of the mixed level orthogonal array is It means there are n1 design variables with q1 levels. It means there are n2 design variables with q2 levels. The values ​​of the parameters of aircrafts with different shapes are recorded in the orthogonal table, which is called the parameter design table; OpenVSP software is used to generate aircraft with different shapes. According to the obtained parameter design table, a parameter design file Design file for controlling the cross-sectional shape is generated. The parameter design file is used to adjust the geometric shape of the benchmark model. The content includes the number of design parameters and the specific value of each design parameter. A script file Script file is written and called by OpenVSP software to batch generate more than two aircraft shapes.

7. The method according to claim 6, characterized in that In step 3-2, for a lifting body aircraft, the lifting body head cross-sectional shape, head cross-sectional height, head cross-sectional aspect ratio, tail cross-sectional shape, tail cross-sectional height, and tail cross-sectional aspect ratio are used as design parameters. The cross-sectional shapes include ellipse, triangle, and rhombus. More than two lifting body models are generated by changing the design parameters. The aerodynamic characteristics of each lifting body shape are calculated using an engineering estimation method: the aircraft shape is divided into two or more surface elements, and the pressure coefficient C of each surface element is calculated. p , and then integrate to obtain the aerodynamic characteristics data of the entire aircraft. When solving the pressure on the surface element, first divide the surface element into the windward surface and the leeward surface according to the positive and negative impact angle. The impact angle calculation formula is: in is the outward normal vector of the surface element; is the incoming flow velocity vector; δ greater than 0 indicates the windward impact angle, and less than 0 indicates the leeward impact angle; For the windward side, the selected pressure coefficient C p The calculation formula is the modified Newton theory: Where γ is the specific heat ratio, Ma ∞ is the Mach number; For the leeward side, the selected pressure coefficient C p Calculation formula: Prandtl-Meyer formula: Using the pressure coefficient C of each surface element p , unit normal vector Area ΔA and aircraft reference area S ref , integrate to get the aerodynamic coefficients in the X, Y, and Z directions of the aircraft; Unit Normal Vector The calculation formula is: Where |N| is The module length, N x for The component of the vector in the X direction, N y for The component of the vector in the Y direction, N z for The component of the vector in the Z direction; for The component in the X direction, for The component in the Y direction, for Component in the Z direction; The integral formula is: Among them, n x for The value of n y for The value of n z for The value of F x is the aerodynamic coefficient in the X direction, F y is the aerodynamic coefficient in the Y direction, F z is the aerodynamic coefficient in the Z direction; Finally, the lift coefficient C L , drag coefficient C D , torque coefficient C M : C L =-F x sinα+F z cosα C D =F x cosαcosβ+F y sinβ+F z sinαcosβ Where L ref is the characteristic length, the subscript cg represents the center of gravity, α is the angle of attack, and β is the sideslip angle; C Mx , C My , C Mz are the moment coefficients of the aircraft around the X-axis, Y-axis, and Z-axis respectively; x, y, z are the coordinates of the location of the surface element, x cg is the coordinate value of the center of gravity of the aircraft on the X axis, y cg is the coordinate value of the center of gravity of the aircraft on the Y axis, z cg is the coordinate value of the center of gravity of the aircraft on the Z axis; Based on the engineering estimation method, the lift coefficient, drag coefficient and pitch moment coefficient of each aircraft in the geometric shape library are calculated at different Mach numbers and angles of attack, and all the calculated aerodynamic data are used as an aerodynamic database.

8. The method according to claim 7, characterized in that In step 4, a one-dimensional convolutional neural network is used as a deep neural network model. The one-dimensional convolutional neural network includes a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer can extract the features of the input data; the pooling layer is used to reduce the data scale, extract the information after convolution, and improve the calculation speed, reduce the number of parameters to prevent overfitting; the convolutional layer and the pooling layer appear alternately, continuously learn the features of the input data, and finally pass the features to the fully connected layer to obtain the network output.

9. The method according to claim 8, characterized in that In step 5, the aerodynamic database constructed in step 2 is randomly divided into a training set, a validation set, and a test set. The lift coefficient, drag coefficient, lift-to-drag ratio, pitch moment coefficient, and angle of attack of each shape are used as the input of the one-dimensional convolutional neural network, and the corresponding geometric characteristic parameters of the shape are used as the network output. During the network training process, the learning rate is set to 2×10 -4 , the optimizer uses Adam, and the loss function uses mean square error MSE: Where n is the sample size, Y i is the true value, is the network prediction value.

10. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores program codes, and when the program codes are executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 9.

Images