A method for predicting distributed loads on control surfaces in flow fields based on deformation

Abstract

The present invention provides a method for predicting the distributed load of a rudder surface in a deformation-based flow field, which relates to the technical field of aircraft design and solves the problem that the control torque and distributed load of a thin rudder surface cannot be measured using traditional force balance technology. The method includes: calculating the aerodynamic load field of the aircraft's rudder surface based on a wing-rudder fusion model, and establishing a typical rudder surface aerodynamic load database; using thin plate control equations or finite element method to determine the correspondence between the deformation displacement field and the aerodynamic load field of the rudder surface, and establishing a rudder surface load-displacement database; obtaining a deep convolutional neural network and completing corresponding training to construct a rudder surface displacement-load solver; in engineering applications, inputting a measured rudder surface displacement cloud map into the rudder surface displacement-load solver, and the solver correspondingly outputs a rudder surface load cloud map to achieve load prediction. The present invention makes it possible to measure the refined load of thin-walled rudder surfaces, providing key data support for structural strength analysis and aerodynamic optimization.

Description

Technical Field

[0001] The present invention relates to the technical field of aircraft design, and is applied to a process of acquiring aircraft control surface loads, and in particular to a method for predicting control surface distributed loads in a deformation-based flow field. Background Art

[0002] In modern aircraft design, the load characteristics of control surfaces directly impact the reliability of flight control systems. Accurately measuring control surface hinge torque under different flight conditions and matching it with appropriate servos is a core component of aircraft design. With increasing demands for stealth performance and maneuverability, the trend toward integrated wing and control surface designs is becoming increasingly prominent. This has led to a continuous reduction in the structural thickness of control surfaces and their mounting areas, posing a significant challenge to traditional physical-based hinge torque testing methods.

[0003] Currently, three main technical means are used in the engineering field to obtain rudder load data. The first is the computational fluid dynamics method, which obtains flow field and surface load information by establishing a full-scale model of the aircraft and discretizing and solving the flow control equations. This method has achieved high accuracy in lift calculations under simple shapes and high Mach number conditions, and the relative error can be controlled within 3%. However, when faced with complex flow phenomena in the rudder area - including physical processes such as shock wave boundary layer interference, gap flow, and unsteady flow separation, the prediction deviation of the local rudder load increases significantly due to the limitations of turbulence model accuracy and grid resolution. In particular, the dynamic separation flow and shock wave oscillation phenomena in the leeward area of the rudder are difficult to accurately capture with existing computational models, resulting in insufficient credibility in load assessments in key areas.

[0004] The second pressure measurement integration technique uses discrete pressure sensors placed on the rudder surface and a reconstruction algorithm to invert the surface pressure distribution. This method has a certain adaptability in testing special-shaped rudder surfaces, but due to the spatial limitations of thin-walled structures, there are technical difficulties in optimizing the density and position of the pressure measurement points. Engineering practice has shown that reconstruction algorithms based on area weighting or radial basis functions have significant errors in complex flow areas, with systematic deviations from strain balance measured data reaching the order of 10%. In addition, the difficulty in laying out internal piping on thin rudder surfaces further limits the applicable scenarios of this technology.

[0005] The third traditional method, strain gauge force measurement, relies on the principles of elastic mechanics to infer load parameters by measuring structural strain. While its theoretical framework is comprehensive and widely used in conventional structural testing, it faces inherent drawbacks such as long design cycles and high testing costs. When applied to ultra-thin rudder surfaces, the conflict between controlling structural deformation and the space required to install the force gauge becomes particularly pronounced. On the one hand, the excessively thin structure's stiffness makes it difficult to meet the mechanical requirements for the force gauge's installation. On the other hand, the reduced thickness of the rudder surface increases the additional moment on the force gauge, directly impacting measurement accuracy. For ultra-thin rudder surfaces with thicknesses below a critical value, existing force gauges are unable to effectively measure these surfaces due to physical size limitations.

[0006] Considering existing technologies, computational fluid dynamics methods suffer from insufficient accuracy in simulating complex flows, pressure integration techniques are limited by sensor layout and reconstruction algorithm errors in thin-walled structures, and traditional strain gauge balances face spatial adaptation challenges in ultra-thin structures. These technical bottlenecks severely restrict the ability to assess control surface loads on modern aircraft as they develop stealth and high maneuverability, necessitating the development of new measurement methods to overcome these limitations. Summary of the Invention

[0007] The purpose of the present invention is to solve the problem that some aircraft have thin control surfaces that cannot be measured using traditional force balance techniques to obtain their control torque and distributed load. Therefore, a method for predicting the distributed load of the control surface in a deformation-based flow field is proposed. The present invention uses binocular vision to measure the deformation of the control surface in real time after being loaded, and reconstructs the distributed load of the control surface based on the displacement measurements of key points. In this way, the present invention avoids the difficulty of installing measurement equipment due to space limitations, and can accurately obtain the distributed load of the control surface, filling the gap in the existing technology for effectively obtaining distributed loads.

[0008] The present invention adopts the following technical solutions to achieve the purpose:

[0009] A method for predicting distributed loads on control surfaces in a deformation-based flow field comprises the following steps:

[0010] S1. Based on the wing-rudder fusion model, calculate the aerodynamic load field of the aircraft's rudder surface under different flight conditions and establish a typical rudder surface aerodynamic load database;

[0011] S2. Using the thin plate governing equations or the finite element method, determine the correspondence between the deformation displacement field of the control surface and the aerodynamic load field, establish a control surface load-displacement database, and obtain a training data set from it;

[0012] S3. Obtain a deep convolutional neural network, complete training of the deep convolutional neural network using the training data set, and construct a rudder displacement-load solver;

[0013] S4. In engineering applications, the measured rudder surface displacement cloud map is input into the rudder surface displacement-load solver, and the rudder surface displacement-load solver outputs the rudder surface load cloud map to realize the prediction of the rudder surface distributed load.

[0014] Specifically, in step S1, the wing-rudder fusion model consists of a wing surface of preset area, shape and size and a rudder surface hinged to the wing surface; when calculating the rudder surface aerodynamic load field, the corresponding calculation conditions are designed based on the flight envelope parameters of the aircraft, and CFD calculations are performed to obtain the aerodynamic load distribution of the rudder surface within the flight envelope, which serves as the basis for establishing a typical rudder surface aerodynamic load database.

[0015] Preferably, when performing CFD calculations, a second-order method based on finite volume method discretization is used to solve the Reynolds-averaged Navier-Stokes equations under the calculation conditions, and the k-ω SST turbulence model is used to achieve the closure of the equation; the grid type used in the calculation is a polyhedral unstructured grid, free far-field conditions are used all around the calculation domain, and adiabatic wall conditions are used for the wing and rudder surfaces.

[0016] Specifically, in the wing-rudder fusion model, the aerodynamic forces acting on the rudder surface include the pressure perpendicular to any surface of the rudder surface and the friction parallel to any surface of the rudder surface. After performing CFD calculations for different flight conditions of the aircraft, the CFD calculation results are extracted to obtain the upper surface pressure distribution acting on the rudder surface. and the lower surface pressure distribution , forming a rudder pressure matrix corresponding to each flight condition; by traversing the rudder pressure matrix corresponding to each flight condition, a typical rudder aerodynamic load database can be formed.

[0017] Furthermore, in step S2, thin plate control equations or finite element method are used to process the two forms of the control surface;

[0018] The first form is that the upper and lower surfaces of the rudder are both symmetrical planes, and the thickness change rate between the symmetrical planes is less than the preset thickness threshold. This form is processed using the thin plate control equation;

[0019] The second form is that the thickness change rate between the upper and lower surfaces of the rudder surface is greater than or equal to a preset thickness threshold, or the curvature of the upper / lower surface of the rudder surface is greater than or equal to a preset curvature threshold. This form is processed using the finite element method.

[0020] Specifically, for the first type of rudder surface, based on Kirchhoff thin plate theory, with deflection as the unknown quantity, the thin plate control equation is established as follows:

[0021]

[0022] Where, is the bending stiffness of the rudder surface, is the distributed load on the rudder surface, is the deflection distribution of the rudder surface, coordinate Indicates the horizontal and vertical coordinates in the plane coordinate system established based on the plane where the upper / lower surface of the rudder is located. is the gradient operator; The expression is as follows:

[0023]

[0024] Where, is the elastic modulus of the rudder surface, is the thickness of the rudder surface, is Poisson's ratio; when the above thin plate control equation is applied to the rudder, when there is a distributed load in the rudder , then the rudder surface will have a deflection distribution at the same time , so for each flight condition, there is ; By distributing the load After expressing the net load acting on the rudder surface, the control equation of the thin plate is solved to obtain the deflection distribution corresponding to the rudder surface , forming the deformation displacement field of the rudder surface, that is, determining the corresponding relationship between the deformation displacement field of the rudder surface and the aerodynamic load field.

[0025] Specifically, for the second type of rudder, the finite element method is directly used to solve it. The deformation displacement field of the rudder and the aerodynamic load field are related through a set of partial differential equations that control the elastic deformation of the solid, forming the Lame-Navier equation shown below:

[0026]

[0027] In the equation, is the elastic modulus of the rudder surface, is the gradient operator, is Poisson's ratio; 、 、 are the corresponding coordinates in the three-dimensional coordinate system; 、 、 Respectively represent the 、 、 Displacement in the direction; 、 、 Respectively represent 、 、 Directional body forces; is the volume strain of the rudder surface, and its value is equal to the first invariant of the strain tensor, as shown below:

[0028]

[0029] The above partial differential equations are a linear system. For any distributed load on the rudder surface By giving boundary conditions and load conditions, the deformation displacement field of the linear system can be obtained, thereby determining the corresponding relationship between the deformation displacement field of the rudder surface and the aerodynamic load field.

[0030] Preferably, in step S3, the training data set includes a rudder displacement cloud map as a training input, and a rudder load cloud map corresponding to the rudder displacement cloud map obtained based on the rudder load-displacement database; the deep convolutional neural network includes multiple branch paths, each branch path includes multiple convolution layers, multiple activation function layers and a fully connected layer; the input of some convolution layers in a preset number of branch paths comes from the output of the activation function layer in any other branch path; the output of each branch path is used as the output of the deep convolutional neural network after the output fusion operation.

[0031] Specifically, the constructed rudder displacement-load solver is expressed as follows:

[0032]

[0033] Where, is the parameter set of the deep convolutional neural network; is the displacement cloud map of the rudder surface, which serves as the input of the solver; The rudder surface load cloud map is the output of the solver and is used to represent the distributed load on the rudder surface.

[0034] In a deep convolutional neural network, the convolution layer and activation function layer of each branch path make the input features undergo convolution operation and nonlinear transformation, and the output is as follows:

[0035]

[0036] Where, is the input feature map of the convolutional layer, is the convolution operator, is a trainable convolution kernel, is the bias parameter, is a nonlinear activation function, is the output feature map of the activation function layer; the nonlinear activation function in the activation function layer The ReLU activation function is used, and its operation is as follows:

[0037]

[0038] Where, Represents the input value of the activation function layer, Represents the input value and 0, select the larger value;

[0039] Loss Functions for Deep Convolutional Neural Networks As follows:

[0040]

[0041] Where, represents the distributed load on the rudder surface obtained by CFD calculation, represents the rudder surface distributed load predicted by the deep convolutional neural network, Represents the total number of feature points; , represents the trainable convolution kernel and its corresponding weight A collection of L2 regularization coefficient to prevent overfitting of deep convolutional neural networks; Represents the system of differential equations that govern the displacement and stress of the control surface.

[0042] Specifically, in step S4, a high-definition camera is used to photograph and monitor the aircraft's rudder surface, and the deformation of the rudder surface after being loaded is measured in real time through binocular vision to form a rudder surface displacement cloud map; the rudder surface displacement-load solver outputs the corresponding rudder surface load cloud map based on the measured rudder surface displacement cloud map, thereby realizing the prediction of the rudder surface distributed load.

[0043] In summary, due to the adoption of this technical solution, the beneficial effects of the present invention are as follows:

[0044] The method of the present invention utilizes structural deformation characteristics as a key point, pioneering non-contact measurement of distributed loads on rudder surfaces. Unlike the traditional pressure integration method, which relies on the limitations of discrete pressure point reconstruction, this method combines high-precision displacement field measurements with mechanical constitutive equations to directly obtain continuous load distribution on the rudder surface, providing key data support for structural strength analysis and aerodynamic optimization. This method also represents a breakthrough in resolving the industry's challenge of insufficient installation space for ultra-thin rudder surfaces, preventing the installation of force balances or pressure measurement pipelines. This makes refined load measurement possible on thin-walled rudder surfaces.

[0045] At the engineering application level, the method of the present invention also demonstrates good technical transferability. In addition to being effectively applied in the field of aircraft design, its core principles can also be extended to the health monitoring process in the fields of building cantilever structures, large-span bridge components, etc. For example, in order to solve the problem of direct stress distribution measurement of large-area cantilever panels in buildings, the structural displacement field can be obtained through a machine vision system, and then the method of the present invention can be combined to achieve non-contact load reconstruction, thereby significantly improving the assessment capability of the structural safety of hidden parts. Compared with the discrete point pressure interpolation mode of the traditional pressure measurement integration method and the inherent defect of the strain balance that can only obtain concentrated force, the method of the present invention has the dual advantages of global load distribution measurement and non-contact measurement.

[0046] Comparisons between numerical simulations and typical operating conditions have demonstrated that the load prediction accuracy of this method meets practical engineering requirements. The magnitude and spatial distribution of the distributed loads are highly consistent with the true values. The deviations between the rudder surface concentrated force and hinge moment parameters obtained through integral calculations and the baseline values are consistently controlled within 5%. This method not only effectively overcomes the computational distortion inherent in traditional measurement methods in complex flow regions, but also provides reliable technical support for the refined design of thin-walled rudder surfaces for next-generation aircraft. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A schematic diagram briefly describing the overall process of the method of the present invention;

[0048] Figure 2 is a schematic diagram of the wing-rudder fusion model used in the method of the present invention;

[0049] Figure 3 is an overall schematic diagram of the computational domain grid used in the method of the present invention;

[0050] Figure 4 Schematic diagram of the calculation grid and boundary layer of the wing-rudder fusion model in the method of the present invention;

[0051] Figure 5 This is an example diagram of the rudder surface pressure cloud diagram in the method of the present invention;

[0052] Figure 6 Schematic diagram of a thin plate model based on Kirchhoff thin plate theory in the method of the present invention;

[0053] Figure 7 This is the aerodynamic load distribution diagram of the windward surface of the second type of rudder surface in the method of the present invention;

[0054] Figure 8 The aerodynamic load distribution diagram of the leeward side of the second type of rudder surface in the method of the present invention is shown in FIG.

[0055] Figure 9 Schematic diagram of the deformation displacement of the steel rudder surface in the method of the present invention;

[0056] Figure 10 Schematic diagram of the deformation displacement of the aluminum alloy rudder surface in the method of the present invention;

[0057] Figure 11 Schematic diagram of the structure of the deep convolutional neural network used in the method of the present invention. DETAILED DESCRIPTION

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0059] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0060] Example

[0061] A method for predicting distributed loads on control surfaces in a deformation-based flow field. Figure 1 The overall process of the method is briefly described here; the various steps of the method are summarized as follows:

[0062] S1. Based on the wing-rudder fusion model, calculate the aerodynamic load field of the aircraft's rudder surface under different flight conditions and establish a typical rudder surface aerodynamic load database;

[0063] S2. Using the thin plate governing equations or the finite element method, determine the correspondence between the deformation displacement field of the control surface and the aerodynamic load field, establish a control surface load-displacement database, and obtain a training data set from it;

[0064] S3. Obtain a deep convolutional neural network, complete training of the deep convolutional neural network using the training data set, and construct a rudder displacement-load solver;

[0065] S4. In engineering applications, the measured rudder surface displacement cloud map is input into the rudder surface displacement-load solver, and the rudder surface displacement-load solver outputs the rudder surface load cloud map to realize the prediction of the rudder surface distributed load.

[0066] This embodiment will introduce the details or preferred contents of each step in detail according to the above step sequence.

[0067] In step S1, according to the characteristics of the thin rudder surface, the following can be established: Figure 2 The wing-rudder fusion model shown, Figure 2 The example dimensioning is given in the figure and can be used as a modeling reference. The wing-rudder fusion model consists of a wing with a preset area, shape and size, and a rudder hinged to the wing.

[0068] For common aircraft flight conditions, the outflow Mach number, inflow static pressure, inflow static temperature, flight angle of attack, and flight sideslip angle can be determined based on the flight envelope parameters. Corresponding calculation conditions can then be designed. Through CFD calculations, the aerodynamic load distribution of the control surfaces within the flight envelope can be derived, serving as the basis for establishing a typical control surface aerodynamic load database. Because the wing-rudder fusion model in this embodiment is a simplified model, it can avoid the computational complexity required by the complex geometry of a real aircraft. At the same time, this typical wing-rudder combination can best reflect the pressure distribution patterns on the control surfaces.

[0069] In this embodiment, a second-order method based on the finite volume method (FVM) discretization is used during CFD calculations to solve the Reynolds-averaged Navier-Stokes (RANS) equations for the calculation conditions. The k-ω SST turbulence model is also used to achieve closure of the equations. The k-ω SST turbulence model has been validated through extensive engineering calculations and is well-suited for simulations with adverse pressure gradients and separation zones. Therefore, it is suitable for the flow characteristics of the wing-rudder fusion model of this embodiment in the leeward region of the large rudder.

[0070] The mesh type used in the calculation of this embodiment is a polyhedral unstructured mesh. The mesh used in the calculation domain can be found in Figure 3 The computational grid and boundary layer after application to the wing-rudder fusion model can be seen in Figure 4 The recommended number of mesh elements is about 2 million. Free far-field conditions are used around the computational domain, while adiabatic wall conditions are used on the wing and rudder surfaces.

[0071] In the wing-rudder fusion model of this embodiment, the aerodynamic forces acting on the rudder surface include the pressure perpendicular to any surface of the rudder surface and the friction parallel to any surface of the rudder surface. Compared with the surface pressure, the surface friction is a small quantity. This embodiment will use magnitude analysis to determine it. First, the surface friction coefficient of the plate is It can be determined as follows:

[0072]

[0073] Where, Represents the Reynolds number; during flight, the Reynolds number at high altitude is generally taken as 10 7 to 10 8 In this embodiment, the rudder surface area is 0.2m 2For example, using the atmospheric parameters at an altitude of 10 km to calculate the surface friction of an aircraft with a flight Mach number of 0.6, the following is true:

[0074]

[0075] The direction of this surface friction force is parallel to the rudder surface. If the thickness of the rudder surface is assumed to be 10 mm, the torque on its rudder shaft will be less than 0.04 Nm. This torque value is small compared to the hinge torque generated by the surface pressure (usually 100 Nm), so it can be ignored in this embodiment.

[0076] This embodiment performs CFD calculations for different flight conditions of the aircraft, extracts corresponding CFD calculation results, and obtains the pressure distribution on the control surface, for example: Figure 5 An example of this is shown; when the rudder action deviates, Figure 5 The left side of the middle is the pressure cloud diagram of the windward side of the rudder, and the right side is the pressure cloud diagram of the leeward side of the rudder. In this way, the pressure distribution on the upper surface of the rudder can be obtained. and the lower surface pressure distribution , forming a pressure matrix of the control surface corresponding to each flight condition. In addition, in actual engineering applications, since there are many grid cells on the control surface, it is not necessary to extract pressure for each grid point. Instead, a series of feature points can be pre-designed for sampling. These feature points can be obtained through the local coordinate system corresponding to the control surface, such as the standard three-axis coordinate system. In this way, by traversing the control surface pressure matrix corresponding to each flight condition, a typical control surface aerodynamic load database can be formed.

[0077] In step S2, for each flight state of the aircraft, there is a corresponding rudder distributed load based on the typical rudder aerodynamic load database. , which can be based on the upper surface pressure distribution of the rudder and the lower surface pressure distribution In this embodiment, the deformation and displacement simulation process of the rudder surface is processed by using the thin plate control equation or the finite element method according to the two forms of the rudder surface.

[0078] The first form is one in which the upper and lower surfaces of the control surface are both symmetrical planes, and the thickness change rate between the symmetrical planes is less than a preset thickness threshold. This form is processed using the thin plate control equation. Because the symmetrical plane dimensions of modern aircraft control surfaces are typically much larger than their thickness dimensions, that is, the ratio of thickness to width is between 1 / 80 and 1 / 5, which meets the assumptions of Kirchhoff's thin plate theory, this embodiment simplifies the control surface of the first form into a thin plate form, and its deformation displacement is mainly concentrated within the symmetrical plane, thereby ignoring deformation in the other two directions. Based on Kirchhoff's thin plate theory, with deflection as the unknown quantity, the thin plate control equation is established as follows:

[0079]

[0080] Where, is the bending stiffness of the rudder surface, is the distributed load on the rudder surface, is the deflection distribution of the rudder surface, coordinate Indicates the plane coordinate system established based on the plane where the upper / lower surface of the rudder is located The horizontal and vertical coordinates in is the gradient operator; The expression is as follows:

[0081]

[0082] Where, is the elastic modulus of the rudder surface, is the thickness of the rudder surface, is Poisson's ratio; when the above thin plate control equation is applied to the rudder, when there is a distributed load in the rudder , then the rudder surface will have a deflection distribution at the same time , so for each flight condition, there is ; By distributing the load After expressing the net load acting on the rudder surface, the control equation of the thin plate is solved (the solution of this equation can also be achieved using the finite element method) to obtain the deflection distribution corresponding to the rudder surface , forming the deformation displacement field of the rudder surface, that is, determining the corresponding relationship between the deformation displacement field of the rudder surface and the aerodynamic load field.

[0083] The thin plate model under Kirchhoff thin plate theory can be found in Figure 6 In this embodiment, a rectangular plate is used to represent the shape of the rudder surface. Coordinate system, the rudder length is , with a width of , thickness is ; There is a rudder axis at the center line in the width direction ( Figure 6In this established rudder coordinate system, the rudder shape function can be expressed as , on its upper surface there is pressure , the pressure is distributed on the lower surface .

[0084] After the above simplification, the force on the first type of rudder in the air can be simplified to have a distributed load A rectangular thin plate that acts and rotates around a fixed axis. , which represents the net load acting on the rudder surface. After solving the thin plate governing equations under Kirchhoff theory, the deformation displacement field of the first form of the rudder surface is obtained.

[0085] The second form is that the thickness change rate between the upper and lower surfaces of the control surface is greater than or equal to a preset thickness threshold, or the curvature of the upper / lower surface of the control surface is greater than or equal to a preset curvature threshold; this form is usually seen on the trailing edge control surface or V-tail control surface of the aircraft. Since this type of control surface does not meet the thin plate theory assumptions, this embodiment will directly use the finite element method to process it.

[0086] When the finite element method is used for solution, the deformation displacement field of the rudder surface and the aerodynamic load field are related through a set of partial differential equations that control the elastic deformation of the solid, forming the Lame-Navier equations shown below:

[0087]

[0088] In the equation, is the elastic modulus of the rudder surface, is the gradient operator, is Poisson's ratio; 、 、 are the corresponding coordinates in the three-dimensional coordinate system; 、 、 Respectively represent the 、 、 Displacement in the direction; 、 、 Respectively represent 、 、 Directional body forces; is the volume strain of the rudder surface, and its value is equal to the first invariant of the strain tensor, as shown below:

[0089]

[0090] The above partial differential equations are a linear system. For any distributed load on the rudder surface By giving boundary conditions and load conditions, the deformation displacement field of the linear system can be obtained, thereby determining the corresponding relationship between the deformation displacement field of the rudder surface and the aerodynamic load field.

[0091] In this embodiment, the characteristic point representation method of the rudder surface can adopt the standard three-axis three-dimensional coordinate system. ; Combined with the measured distributed loads in the typical rudder aerodynamic load database , that is, the aerodynamic load field, can be combined with the deformation displacement field obtained by solving the above partial differential equations Take the rudder with this morphological feature as an example, after the aerodynamic load distribution is applied to the rudder surface, its distribution can be seen in Figure 7 and Figure 8 , where Figure 7 This is the pressure diagram of the windward side. The load is larger overall. Figure 8 This is an indication of the pressure on the leeward side, and the overall load is small.

[0092] After solving with the finite element method, the displacement changes of the rudder surfaces of different materials under the above aerodynamic loads can be found in Figure 9 The steel types shown and Figure 10 Therefore, after sampling the characteristic points (in actual application, the sampling characteristic points for deformation displacement and aerodynamic load can be distributed inconsistently, but if a consistent characteristic point distribution is adopted for sampling, the calculation process can be relatively simplified and the processing is facilitated), the deformation displacement field of the rudder surface is integrated with the typical rudder surface aerodynamic load database to obtain the rudder surface load-displacement database.

[0093] In step S3, the training dataset includes a rudder displacement cloud map as training input, and a rudder load cloud map corresponding to the rudder displacement cloud map, derived from a rudder load-displacement database. This embodiment uses a deep convolutional neural network to construct a rudder displacement-load solver. The rudder displacement cloud map and rudder load cloud map can be viewed as two-dimensional single-channel photographs. After undergoing uniform preprocessing operations such as scaling and translation, they are input into the deep convolutional neural network for training.

[0094] Considering that the deep convolutional neural network needs to have the ability of input coordinate translation invariance and precise spatial capture, this embodiment integrates full connection technology and convolution technology to form the deep convolutional neural network to realize the construction of the rudder displacement-load solver.

[0095] In this embodiment, a structure example of a deep convolutional neural network is as follows: Figure 11As shown in the figure, "InPut" represents the input layer, "Conv" represents the convolutional layer, ReLU represents the activation function layer, "Pooling" represents the maximum pooling layer, "FC" represents the fully connected layer, "CONCAT" represents the fusion layer, and "OutPut" represents the output layer. This deep convolutional neural network has three branch paths, which are described as follows:

[0096] Branch path 1 receives external input from the input layer, including: convolution layer 1_1 → activation function layer 1_1 → max pooling layer 1_1 → convolution layer 1_2 → activation function layer 1_2 → fully connected layer 1_2;

[0097] The input of branch path 2 is the output of activation function layer 1_1 in branch path 1, including: convolution layer 2_1 → activation function layer 2_1 → max pooling layer 2_1 → convolution layer 2_2 → activation function layer 2_2 → fully connected layer 2_2;

[0098] The input of branch path 3 is the output of activation function layer 2_1 in branch path 2, including: convolution layer 3_1 → activation function layer 3_1 → fully connected layer 3_1.

[0099] Subsequently, the fusion layer fuses the outputs of fully connected layer 1_2, fully connected layer 2_2, and fully connected layer 3_1, and after passing through activation function layer 4 and fully connected layer 4 in sequence, it is output to the outside by the output layer.

[0100] Based on the deep convolutional neural network, the rudder displacement-load solver constructed in this embodiment is expressed as the following formula:

[0101]

[0102] Where, is the parameter set of the deep convolutional neural network; is the displacement cloud map of the rudder surface, which serves as the input of the solver; The rudder surface load cloud map is the output of the solver and is used to represent the distributed load on the rudder surface.

[0103] In a deep convolutional neural network, the convolution layer and activation function layer of each branch path make the input features undergo convolution operation and nonlinear transformation, and the output is as follows:

[0104]

[0105] Where, is the input feature map of the convolutional layer, is the convolution operator, is a trainable convolution kernel, is the bias parameter, is a nonlinear activation function, is the output feature map of the activation function layer; the nonlinear activation function in the activation function layer The ReLU activation function is used as follows:

[0106]

[0107] Where, Represents the input value of the activation function layer; Represents the input value and 0, select the larger value.

[0108] Loss Functions for Deep Convolutional Neural Networks As follows:

[0109]

[0110] Where, represents the distributed load on the rudder surface obtained by CFD calculation, represents the rudder surface distributed load predicted by the deep convolutional neural network, Represents the total number of feature points; , represents the trainable convolution kernel and its corresponding weight A collection of L2 regularization coefficient to prevent overfitting of deep convolutional neural networks; Represents the system of differential equations that govern the displacement and stress of the control surface.

[0111] Finally, in step S4, this embodiment uses a high-definition camera to photograph and monitor the aircraft rudder surface, and measures the deformation of the rudder surface after being loaded in real time through binocular vision to form a rudder surface displacement cloud map; the rudder surface displacement-load solver outputs the corresponding rudder surface load cloud map based on the measured rudder surface displacement cloud map, thereby realizing the prediction of the rudder surface distributed load.

Claims

1. A method for predicting the distributed load on a control surface in a deformation-based flow field, characterized in that: The steps include: S1. Based on the wing-rudder fusion model, calculate the aerodynamic load field of the aircraft's rudder surface under different flight conditions and establish a typical rudder surface aerodynamic load database; S2. Using the thin plate governing equations or the finite element method, determine the correspondence between the deformation displacement field of the control surface and the aerodynamic load field, establish a control surface load-displacement database, and obtain a training data set from it; S3. Obtain a deep convolutional neural network, complete training of the deep convolutional neural network using the training data set, and construct a rudder displacement-load solver; S4. In engineering applications, the measured displacement cloud map of the rudder surface is input into the rudder surface displacement-load solver, and the rudder surface displacement-load solver outputs the rudder surface load cloud map to realize the prediction of the rudder surface distributed load; In step S2, thin plate control equations or finite element method are used to process the two forms of the control surface; The first form is that the upper and lower surfaces of the rudder are both symmetrical planes, and the thickness change rate between the symmetrical planes is less than the preset thickness threshold. This form is processed using the thin plate control equation; The second form is that the thickness variation rate between the upper and lower surfaces of the rudder surface is greater than or equal to a preset thickness threshold, or the curvature of the upper / lower surface of the rudder surface is greater than or equal to a preset curvature threshold. This form is processed using the finite element method; For the second type of rudder, the finite element method is directly used to solve it. The deformation displacement field of the rudder and the aerodynamic load field are related through a set of partial differential equations that control the elastic deformation of the solid, forming the Lame-Navier equations shown below: In the equation, is the elastic modulus of the rudder surface, is the gradient operator, is Poisson's ratio; 、 、 are the corresponding coordinates in the three-dimensional coordinate system; 、 、 Respectively represent the 、 、 Displacement in the direction; 、 、 Respectively represent 、 、 Directional body forces; is the volume strain of the rudder surface, and its value is equal to the first invariant of the strain tensor, as shown below: The above partial differential equations are a linear system. For any distributed load on the rudder surface By giving boundary conditions and load conditions, the deformation displacement field of the linear system can be obtained, thereby determining the corresponding relationship between the deformation displacement field of the rudder surface and the aerodynamic load field; In step S3, the training data set includes a rudder displacement cloud map as a training input, and a rudder load cloud map corresponding to the rudder displacement cloud map obtained based on a rudder load-displacement database; the deep convolutional neural network includes multiple branch paths, each branch path includes multiple convolution layers, multiple activation function layers, and a fully connected layer; the input of some convolution layers in a preset number of branch paths comes from the output of the activation function layer in any other branch path; the output of each branch path is used as the output of the deep convolutional neural network after an output fusion operation; The constructed rudder displacement-load solver is expressed as follows: Where, is the parameter set of the deep convolutional neural network; is the displacement cloud map of the rudder surface, which serves as the input of the solver; The rudder surface load cloud map is the output of the solver and is used to represent the distributed load on the rudder surface. In a deep convolutional neural network, the convolution layer and activation function layer of each branch path make the input features undergo convolution operation and nonlinear transformation, and the output is as follows: Where, is the input feature map of the convolutional layer, is the convolution operator, is a trainable convolution kernel, is the bias parameter, is a nonlinear activation function, is the output feature map of the activation function layer; Non-linear activation function in activation function layer The ReLU activation function is used, and its operation is as follows: Where, Represents the input value of the activation function layer, Represents the input value and 0, select the larger value; Loss Functions for Deep Convolutional Neural Networks As follows: Where, represents the distributed load on the rudder surface obtained by CFD calculation, represents the rudder surface distributed load predicted by the deep convolutional neural network, Represents the total number of feature points; , represents the trainable convolution kernel and its corresponding weight A collection of L2 regularization coefficient to prevent overfitting of deep convolutional neural networks; Represents the system of differential equations that control the displacement and stress of the control surface, namely the Lame-Navier equations.

2. The method for predicting the distributed load on the control surface in the flow field according to claim 1, characterized in that: In step S1, the wing-rudder fusion model consists of a wing with a preset area, shape and size, and a rudder hinged to the wing. When calculating the rudder aerodynamic load field, the corresponding calculation conditions are designed based on the flight envelope parameters of the aircraft, and CFD calculations are performed to obtain the aerodynamic load distribution of the rudder within the flight envelope, which serves as the basis for establishing a typical rudder aerodynamic load database.

3. The method for predicting the distributed load on the control surface in the flow field according to claim 2, characterized in that: During CFD calculations, a second-order method based on finite volume discretization was used to solve the Reynolds-averaged Navier-Stokes equations under the calculation conditions, and the k-ω SST turbulence model was used to close the equations. The mesh type used in the calculations was a polyhedral unstructured mesh, and free far-field conditions were used all around the calculation domain, while adiabatic wall conditions were used for the wing and rudder surfaces.

4. The method for predicting the distributed load on the control surface in the flow field according to claim 2, characterized in that: In the wing-rudder fusion model, the aerodynamic forces acting on the rudder surface include the pressure perpendicular to any surface of the rudder surface and the friction parallel to any surface of the rudder surface. After performing CFD calculations for different flight conditions of the aircraft, the CFD calculation results are extracted to obtain the upper surface pressure distribution acting on the rudder surface. and the lower surface pressure distribution , forming a rudder pressure matrix corresponding to each flight condition; by traversing the rudder pressure matrix corresponding to each flight condition, a typical rudder aerodynamic load database can be formed.

5. The method for predicting the distributed load on the control surface in the flow field according to claim 1, characterized in that: For the first type of rudder surface, based on Kirchhoff thin plate theory, with deflection as the unknown quantity, the thin plate control equation is established as follows: Where, is the bending stiffness of the rudder surface, is the distributed load on the rudder surface, is the deflection distribution of the rudder surface, coordinate Indicates the horizontal and vertical coordinates in the plane coordinate system established based on the plane where the upper / lower surface of the rudder is located. is the gradient operator; The expression is as follows: Where, is the elastic modulus of the rudder surface, is the thickness of the rudder surface, is Poisson's ratio; when the above thin plate control equation is applied to the rudder, when there is a distributed load in the rudder , then the rudder surface will have a deflection distribution at the same time , so for each flight condition, there is ; By distributing the load After expressing the net load acting on the rudder surface, the control equation of the thin plate is solved to obtain the deflection distribution corresponding to the rudder surface , forming the deformation displacement field of the rudder surface, that is, determining the corresponding relationship between the deformation displacement field of the rudder surface and the aerodynamic load field.

6. The method for predicting the distributed load on a control surface in a flow field according to claim 1, characterized in that: In step S4, a high-definition camera is used to photograph and monitor the aircraft's control surfaces, and the deformation of the control surfaces after being loaded is measured in real time through binocular vision to form a control surface displacement cloud map; the control surface displacement-load solver outputs the corresponding control surface load cloud map based on the measured control surface displacement cloud map, thereby realizing the prediction of the control surface distributed load.

Images