Aircraft modeling method based on combination of mechanism and PSO feedforward neural network optimization

Abstract

An aircraft modeling method based on mechanism and PSO feedforward neural network optimization comprises the following steps that an aircraft dynamics mechanism model is constructed by establishing an aircraft rigid body model and a projectile body model; embedding the aircraft dynamics mechanism model into a PSO feedforward neural network by using the PSO feedforward neural network, adopting a mechanism embedding neural network method and utilizing rich information contained in flight data, and continuously training a difference value between a mechanism value and an actual value under different conditions according to input characteristics to obtain a neural network residual compensation model; and adding the neural network residual compensation model into the aircraft dynamics mechanism model to obtain an optimized variable configuration aircraft model. According to the method, accurate modeling of different configurations under different environments and targets can be realized, and particularly modeling of variable-configuration aircrafts can be realized.

Description

Technical Field

[0001] The present invention relates to a configuration-variable aircraft modeling method, and in particular to an aircraft modeling method combining mechanism with PSO feedforward neural network optimization, belonging to the technical field of aircraft. Background Art

[0002] Variable configuration aircraft modeling refers to a modeling method that allows the configuration of an aircraft to be dynamically adjusted or changed according to mission requirements and environmental changes during the design and operation of the aircraft. As space missions become increasingly complex and diverse, variable configuration aircraft can adjust their own appearance, structure or function according to mission objectives and real-time conditions to optimize performance and efficiency. Therefore, establishing an accurate mathematical model of a variable configuration aircraft can effectively describe the dynamic adaptability and changeability of the aircraft during the mission, ensuring that the aircraft can perform at its best in different stages and environments.

[0003] However, the current modeling technology of variable configuration aircraft has certain limitations. Most of the existing mechanism modeling methods are based on static or simplified assumptions, and fail to fully consider the complex interaction effects and nonlinear behaviors during the aircraft configuration change process. This results in the inability of existing models to accurately predict the dynamic response of variable configuration aircraft when switching between different configurations, and cannot guarantee the accuracy and reliability of the model. As a result, in the actual application of the aircraft, its variable configuration advantages cannot be fully utilized, resulting in performance degradation or instability.

[0004] In summary, the current variable configuration aircraft modeling technology has large deviations when dealing with issues such as aircraft configuration adjustment, mission switching and environmental adaptability, resulting in problems such as insufficient accuracy and poor robustness, making it difficult to meet the increasingly complex aircraft mission requirements. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the present invention proposes an aircraft modeling method based on a mechanism-combined PSO feedforward neural network optimization, which can achieve accurate modeling of different configurations under different environments and targets.

[0006] A method for aircraft modeling based on mechanism-joint PSO feedforward neural network optimization includes the following steps:

[0007] S1. Construct the aircraft dynamics mechanism model by establishing the aircraft rigid body model and the missile body model;

[0008] S2. Using a PSO feedforward neural network and a method of embedding a mechanism into a neural network, the aircraft dynamics mechanism model is embedded into the PSO feedforward neural network using the rich information contained in the flight data, and the difference between the mechanism value and the actual value under different conditions is continuously trained according to the input features to obtain a residual compensation model;

[0009] S3. Add the residual compensation model to the aircraft dynamics mechanism model to obtain the optimized variable configuration aircraft model, analyze the effectiveness of the optimized model under different configurations, and specifically compare and analyze the output values ​​of the aircraft mechanism model and the output values ​​of the mechanism combined neural network optimization model with the actual flight data to study their error changes.

[0010] Compared with the prior art, the present invention has the following beneficial effects:

[0011] The most significant advantages of the variable configuration aircraft modeling method for the mechanism combined with PSO feedforward neural network optimization in this application are three aspects. First, for the switching of different configurations, the complex nonlinear and coupling effects that are difficult to describe in the traditional mechanism model are compensated, so that the model is more adaptable in the variable configuration state and the modeling accuracy is significantly improved. Secondly, the method can maintain strong robustness under a variety of external interferences (such as airflow changes, structural deformation, etc.), ensuring the stability and accuracy of the model in different speed domains, different flight states and other environments. Finally, the high-precision dynamic model provided by the method provides a solid foundation for the design of the aircraft control system, ensuring the stable flight and efficient control of the aircraft in a variety of configurations and complex environments.

[0012] The modeling accuracy of the method of this application is high, which is significantly improved compared with the mechanism model. Therefore, the aircraft modeling method of mechanism combined with PSO feedforward neural network optimization has important theoretical and engineering significance.

[0013] The present application will be further described below in conjunction with the accompanying drawings and implementation methods: BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is the flow chart of modeling of a variable configuration aircraft by combining the mechanism of the present application with PSO feedforward neural network optimization;

[0015] Figure 2 This is the overall flow chart of the neural network training flight data;

[0016] Figure 3 It is a comparison curve of lift changes of aircraft under different configurations;

[0017] Figure 4 It is a comparison curve of the drag change of the aircraft under different configurations;

[0018] Figure 5 It is a comparison curve of the torque change of the aircraft under different configurations. DETAILED DESCRIPTION

[0019] The embodiments of the technical solution of the present invention will be described in detail below in conjunction with the accompanying drawings. Unless otherwise specified, the technical terms or scientific terms used in this application are generally understood by those skilled in the art.

[0020] This implementation method first constructs a dynamic mechanism model of a variable configuration aircraft, then embeds the mechanism model into a neural network, uses real flight data to train the residual compensation term between the mechanism and the real value, and finally evaluates the effectiveness of the method. Figure 1 As shown, the specific implementation steps are as follows:

[0021] Step 1: Build a model of the aircraft dynamics mechanism

[0022] Step 1.1 Establishment of aircraft rigid body model

[0023] The forces acting on the rigid body of the aircraft are mainly aerodynamic force, thrust, and gravity. This step uses the oblique shock wave and Prandtl-Meyer expansion wave theory to calculate the forces and moments on each surface of the aircraft, obtain the resultant force acting on the aircraft, and estimate the lift, drag, and pitch moment.

[0024] For the upper surface of the front body of the aircraft, the upper surface area is S and the upper surface pressure is P up , the pressures of different surfaces in the aircraft are obtained from the Mach number through the relationship between the oblique shock wave and the Prandtl-Meyer expansion wave, and the fuselage angle on the forebody is τ1. up ,z up ) is located at the center of the upper surface, and the coordinates of the force point are based on the center of mass. Then the force F on the x-axis and z-axis of the body is x,up 、F z,up , and the pitch moment M of the upper surface of the forebody on the x-axis and z-axis up The formula is as follows:

[0025] F x,up =-P up Stanτ1 (1)

[0026] F z,up =P up S (2)

[0027] M up =z up F x,up -x up F z,up (3)

[0028] For the lower surface of the front body of the aircraft, the lower surface area is S1 and the lower surface pressure is P d1 The pressures of different surfaces in the aircraft are obtained from the Mach number through the relationship between the oblique shock wave and the Prandtl-Meyer expansion wave. The angle of the forebody under the fuselage is τ2. d1 ,z d1 ) is located at the center of the lower surface of the front body, and the coordinates of the force point are based on the center of mass. The force F on the x-axis and z-axis of the body x,d1 、F z,d1, and the pitch moment M of the lower surface of the forebody on the x-axis and z-axis d1 The formula is as follows:

[0029] F x,d1 =-P d1 S1tanτ2 (4)

[0030] F z,d1 =-P d1 S1 (5)

[0031] M d1 =z d1 F x,d1 -x d1 F z,d1 (6)

[0032] For the lower surface of the rear body of the aircraft, the lower surface area is S2, and the pressure on the lower surface of the rear body is P d2 , the free flow pressure is P, and the rear body angle is τ3. The rear body force point (x d2 ,z d2 ) is located at the center of the lower surface of the rear body, and the coordinates of the force point are based on the center of mass. x,d2 、F z,d2 And the pitch moment M of the lower surface of the rear body on the x-axis and z-axis d2 The formula is as follows:

[0033]

[0034]

[0035] M d2 =z d2 F x,d2 -x d2 F z,d2 . (9)

[0036] For the aircraft elevator surface, the elevator area is S3, and the pressures on the upper and lower surfaces of the elevator surface are P e_up ,P e_d , the rudder angle is δ e The force acting point (x e , z e ) is located at the center of the elevator, and the coordinates of the force point are based on the center of mass. The force F on the x-axis and z-axis of the aircraft c,e 、F z,e , and the pitch moment M of the elevator surface on the x-axis and z-axis e The formula is as follows:

[0037] F x,e =P e_up S3 sinδ e -P e_dS3 sinδ e (10)

[0038] F z,e =P e_up S3 cosδ e -P e_d S3 cosδ e (11)

[0039] M e =z e F x,e -x e F z,e (12)

[0040] For the aircraft wing surface, the wing span length L b The sweep angle ε and the wing area are variable, and the wing area is the span length multiplied by the average chord length C. avg The air density is ρ and the wing pressure is P b , the flight speed is V. The force acting point (x b ,z b ) is located at the center of the wing, and the coordinates of the force point are based on the center of mass. The force F on the x-axis and z-axis of the fuselage x,b 、F z,b , and the pitch moment M of the wing surface on the x-axis and z-axis b The formula is as follows:

[0041] L b1 =L b cosε (13)

[0042]

[0043] F z,b =P b L b1 c avg (15)

[0044]

[0045] Among them, L b1 is the wingspan length, and its random wing span length L b and changes with the sweep angle ε.

[0046] Step 1.2 Establishment of aircraft body model

[0047] The fuselage body will undergo elastic deformation when flying at high speed, resulting in the deformation angle ΔT of the front and rear bodies. f , ΔT a , at this time, the center of mass of the aircraft will change, which will affect the forces and moments acting on the fuselage.

[0048] Front body deformation angle Δτ f for

[0049]

[0050] Rear body deformation angle Δτ a for

[0051]

[0052] In the above formula represents the vibration mode, η i (t) represents the generalized coordinates.

[0053] Therefore, the rigid body model can be elastically corrected. The corrected upper and lower body angles τ'1 and τ'2 of the front body of the aircraft are

[0054] τ'1=τ1-Δτ f (19)

[0055] τ'2=τ2+Δτ f (20)

[0056] The corrected rear body angles τ'3 are

[0057] τ'3=τ3-Δτ a (twenty one)

[0058] Substitute the corrected front body upper and lower body angles τ'1, τ'2 and rear body angle τ'3 into the series of formulas in step 1.1, and the elastic correction force F on the x-axis of the body can be obtained: x,up 、F x,d1 、F x,d2 、F x,e 、F x,b becomes F' x,up 、F' x,d1 、F' x,d2 、F' x,a 、F' x,b , the force F on the z-axis of the body z,up 、F z,d1 、F z,d2 、F z,e 、F z,b becomes F' z,up 、F' z,d1 、F' z,d2 、F' z,e 、f' z,b , the pitching moment is given by M up 、M d1 、M d2 、M e 、M b becomes M' up , M' d1, M' d2 , M' e , M' b .

[0059] Step 1.3 Establishment of aircraft dynamics mechanism model

[0060] Combining the above rigid body and projectile models, the forces and moments on each surface of the aircraft are added to obtain the force F acting on the x-axis and x 、F z They are

[0061] F x =F' x,up +F' x,d1 +F' x,d2 +F' x,e +F' x,b (twenty two)

[0062] F z =F' z,up +F' z,d1 +F' z,d2 +F' z,e +F' z,b (twenty three)

[0063] According to the coordinate transformation formula and the angle of attack α, the lift L, drag D and pitch moment M in the aircraft velocity coordinate system can be obtained as follows:

[0064] L=F x sinα-F z cosα (24)

[0065] D=-F x cosα-F z sinα (25)

[0066] M=M' up +M' d1 +M' d2 +M' e +M' b (26)

[0067] Therefore, through the above-mentioned aircraft lift L, drag D and pitch moment M, the aircraft dynamics equation can be obtained as follows:

[0068]

[0069] in, θ, h, T, I yy They respectively represent the change in aircraft speed, change in angle of attack, change in pitch angle, change in attitude angle, altitude, thrust and the aircraft's moment of inertia around the horizontal axis.

[0070] Step 2: Construct a neural network residual compensation model

[0071] Step 2.1 Dataset processing and division

[0072] The data set used in this embodiment is 1.5MB in size, with a total of 8020 data items, including the lift, drag, and pitch moment of the aircraft at different sweep angles, wingspan lengths, rudder angles, Mach numbers, angles of attack, and altitudes.

[0073] In terms of flight data processing, the actual flight data is subjected to denoising, normalization, missing value filling, outlier detection and processing operations, thereby effectively improving data quality and training accuracy, and providing a solid data foundation for the joint neural network modeling of the aircraft mechanism.

[0074] In terms of data set division, the built-in divideint() function of MATLAB is used to divide the data into training set, test set and validation set in a ratio of 6:2:2. The divideint() function is an index-based division function that randomly and evenly divides the data into different subsets according to the specified ratio, covering as much feature space as possible, thereby ensuring the diversity and representativeness of the samples in each subset. The divided training set is used to train the neural network model, and the test set is used to determine whether the neural network is overfitting, and then adjust the training strategy at any time. Finally, the validation set is used to evaluate the effectiveness of the neural network training.

[0075] Step 2.2 Feedforward Neural Network Design

[0076] FFNN feedforward neural network is a multi-layer neural network with nonlinear mapping from input to output. This neural network can construct complex feature space mapping relationships through multiple hidden layers and nonlinear activation functions, and has strong fitting ability and flexibility.

[0077] In this experiment, the FFNN feedforward neural network has four layers, namely the input layer, two hidden layers and the output layer. The input layer has six input values, namely the sweep angle, wingspan length, rudder angle, Mach number, angle of attack and altitude of the aircraft. The number of hidden layer nodes is determined by the subsequent PSO optimization algorithm.

[0078] The purpose of this experiment is to learn the error between the mechanism value and the actual value under different input conditions through a feedforward neural network, so as to construct a residual compensation model to compensate for the mechanism model error caused by factors such as variable configuration of the aircraft.

[0079] The specific design of the feedforward neural network is as follows:

[0080] (1) Hidden layer structure design

[0081] The number of hidden layers of the neural network is set to two, and the number of nodes is dynamically adjusted by the PSO algorithm. The specific adjustment method is shown in step 2.3.

[0082] (2) Loss function design

[0083] The mean square error (MSE) is used as the loss function. The neural network weights and biases are optimized by minimizing the loss function, and then the error of the model on the training set and the validation set is evaluated. The loss function formula for the training set and the validation set is J train , J val as follows:

[0084]

[0085] Among them, n is the number of samples in the data set, y i_train ,y i_val is the true value, is the calculated value of the mechanism.

[0086] (3) Training method design

[0087] Bayesian regularized back propagation is used as the training method. This method adds a regularization term to the loss function to penalize excessive or unreasonable network parameters, thereby preventing the neural network from overfitting. The training formula after Bayesian optimization is J' train as follows:

[0088] J' train =MSE train +λ*||w|| 2 (29)

[0089] Among them, MSE train is the mean square error of the training set, ||w|| 2 is the regularization term, w is the network weight parameter, and λ is the regularization coefficient, which is used to control the regularization weight.

[0090] (4) Neural network training

[0091] The error convergence target is set to 0.001, and the training set is used as input for training. The validation set is used to evaluate the network performance, and the network structure is adjusted in real time according to the validation set error. The network is iterated repeatedly until the optimal network is trained.

[0092] Step 2.3 PSO algorithm optimizes network parameters

[0093] PSO particle swarm algorithm is a probabilistic global optimization search algorithm. Its core idea is to seek the optimal solution through cooperation and information sharing between individuals in the population, so as to achieve fast and global optimization in a complex search space. This experiment uses the PSO algorithm to optimize the number of hidden layer nodes of the FFNN feedforward neural network, thereby improving the training efficiency and fitting generalization ability of the neural network.

[0094] The specific optimization design of the PSO algorithm is as follows:

[0095] (1) Particle swarm initialization

[0096] Set the particle swarm size to 50, the number of iterations to 40, the initial value of the speed to 0, the initial value of the global optimal error to inf, the initial value of the global optimal parameter to the initial value of the first particle, and the initial value of the particle position to be randomly generated in the range of [10,100]. Its dimension is 2, which represents the number of nodes in the two hidden layers of the feedforward neural network.

[0097] (2) Fitness function calculation

[0098] The current position value of each particle is used as the number of nodes in the hidden layer of the neural network. The minimum target value J, that is, the mean square error of the validation set, is obtained by training the neural network with data, and then the fitness function can be constructed as follows:

[0099]

[0100] (3) Position and velocity update

[0101] Use the PSO update formula to update the particle position and velocity. The specific velocity and position update formula is v i (t+1) and x i (t+1) is as follows:

[0102] v i (t+1)=wv i (t)+c1r1(p i -x i (t))+c2r2(gx i (t)) (31)

[0103] x i (t+1)=x i (t)+v i (t+1) (32)

[0104] Among them, x i (t) and v i (t) represent the current position and velocity of the particle, p iis the historical optimal position of particle i, g is the global optimal position, w is the inertia weight, c1 and c2 are acceleration factors, the default values ​​are used here, and r1 and r2 are random numbers in [0,1].

[0105] (4) Iterative Optimization

[0106] The above steps are repeated to update the historical optimal position and global optimal position of the particle according to the fitness until the maximum number of iterations is reached or the fitness meets the preset threshold.

[0107] (5) Output optimal parameters

[0108] When the termination condition is reached, the global optimal hidden layer structure and the corresponding neural network model are returned.

[0109] The output layer is the residual between the mechanism value and the actual value, which is expressed as follows:

[0110] e i =y 实际i -y 机理i

[0111] The above formula is the residual value corresponding to the i-th group of data.

[0112] Through the above PSO global optimization, the tedious process of manually adjusting the number of hidden layer nodes is avoided, making the exploration of the number of hidden layer nodes more efficient. And through the appropriate hidden layer node settings, the network performance is improved to prevent overfitting. The optimized network structure can better balance the model complexity and training error, thereby improving the performance of the validation set and test set, and achieving higher training efficiency and prediction accuracy.

[0113] Step 2.4 Establishment of residual compensation model based on PSO-feedforward neural network

[0114] The overall process of the residual compensation model trained by PSO-feedforward neural network is as follows Figure 2 As shown in the figure, the six variables of aircraft sweep angle, wingspan length, rudder deflection angle, Mach number, angle of attack and altitude are used as the input of PSO-feedforward neural network. The mechanism value of the aircraft is calculated through the embedded mechanism model of PSO-feedforward neural network. The actual value is subtracted from the mechanism value so that the neural network outputs the residual compensation model Model 残差补偿 .

[0115] Step 3: Construct a variable configuration aircraft model optimized by combining the mechanism with PSO-feedforward neural network

[0116] The residual compensation model obtained by the neural network is used to establish the variable configuration aircraft model optimized by the mechanism combined with PSO-feedforward neural network as follows:

[0117] Model 混合 =Model机理 +Model 残差补偿 (34)

[0118] Among them, Model 残差补偿 Represents the residual compensation model obtained by PSO-feedforward neural network, Model 机理 Refers to the aircraft dynamics mechanism model.

[0119] The above hybrid model can be used to predict the corresponding residual compensation value according to the input parameters of different conditions in actual tests, thereby compensating for the mechanism model error caused by the aircraft's variable configuration, missile body coupling, etc., and obtaining a more accurate aircraft model.

[0120] Example

[0121] The flight data set is obtained. The data contains the force and moment values ​​of the aircraft under different input conditions. The input parameter ranges are as follows: the sweep angle ranges from 0 to 60 degrees, the wingspan ranges from 20 to 40 feet, the rudder angle ranges from -5 to 5 degrees, the Mach number ranges from 1.2 to 5.6, the angle of attack ranges from 0 to 10 degrees, and the altitude ranges from 9000 to 11000 meters.

[0122] For the calculation of the variable configuration aircraft mechanism model, the aircraft related parameters are as follows:

[0123] Table 1 Parameters of variable configuration aircraft

[0124]

[0125] According to the aircraft-related parameters in Table 1, the mechanism values ​​of the aircraft, namely lift, drag, and pitch moment, are calculated through step 1. According to the obtained flight data set, the residual model is trained through the PSO-feedforward neural network in step 2, and the variable configuration aircraft model optimized by the mechanism combined with the PSO-feedforward neural network is obtained in step 3, so as to calculate the force and moment values ​​after data optimization.

[0126] In terms of the overall optimization effect, through the comparison of the error results of the test set, for lift, the error of the mechanism model before neural network correction was 107.85%, and the error after correction was 0.03%; for drag, the error before correction was 121.91%, and the error after correction was 0.06%; for torque, the error before correction was 96.47%, and the error after correction was 0.03%; the above error values ​​are all less than 0.1%, which shows that the variable configuration aircraft model optimized by the mechanism combined with PSO-feedforward neural network has a good effect.

[0127] Further analysis of the optimization effects of individuals under different conditions is carried out by Figure 3 : Comparison curve of lift changes of aircraft under different configurations, Figure 4 : Comparison curve of the drag change of the aircraft under different configurations, Figure 5 :Comparison curves of the torque changes of the aircraft under different configurations are shown. It shows the comparison curves of the real value, mechanism value and mechanism-neural network optimization value of force and torque when the altitude is 10000m and the Mach number is 2. It can be seen from the figure that with the changes in the angle of attack, sweep angle and wingspan line length, the mechanism-neural network optimization value curve of force and torque is consistent with the actual value curve. It can be seen that the variable configuration aircraft model optimized by the mechanism combined with PSO-feedforward neural network still has a good effect under changing conditions.

[0128] The present application has been disclosed as above with preferred implementation cases, but it is not intended to limit the present application. Any technician familiar with the profession can make slight changes or modifications to the above-disclosed structures and technical contents without departing from the scope of the technical solution of the present application, and the equivalent implementation cases with equivalent changes are still within the scope of the technical solution of the present application.

Claims

1. A method for aircraft modeling based on mechanism-joint PSO feedforward neural network optimization, characterized in that: The following steps are involved: S1. Construct the aircraft dynamics mechanism model by establishing the aircraft rigid body model and the missile body model; S2, using a PSO feedforward neural network, embedding the aircraft dynamics mechanism model into the PSO feedforward neural network, continuously training the difference between the mechanism value and the actual value in different situations according to the input features, and constructing a residual compensation model; S3. Add the residual compensation model to the aircraft dynamics mechanism model to obtain an optimized variable configuration aircraft model.

2. The aircraft modeling method of the mechanism combined with PSO feedforward neural network optimization according to claim 1 is characterized by: Step S1 uses oblique shock wave and Prandtl-Meyer expansion wave theory to calculate the forces and moments on each surface of the aircraft, obtain the resultant force acting on the aircraft, estimate the lift, drag and pitch moment, and obtain the aircraft dynamics mechanism model.

3. The aircraft modeling method of the mechanism combined with PSO feedforward neural network optimization according to claim 1 is characterized by: The process of establishing the aircraft rigid body model in step S1 is as follows: For the upper surface of the front body of the aircraft, set the upper surface area to S and the upper surface pressure to P up , the body angle on the front body is τ1, at this time the force point on the front body (x up , Z up ) is located at the center of the upper surface, and the force F on the x-axis and z-axis of the body x，up 、F z,up , and the pitch moment M of the upper surface of the forebody on the x-axis and z-axis up The formula is as follows: F x,up =-P up Stanτ1 (1) F z，up =P up S (2) M up =z up F x,up -x up F z,up (3) For the lower surface of the front body of the aircraft, the lower surface area is set to S1 and the lower surface pressure is set to P d1 , the lower body angle of the front body is τ2, at this time, the force point of the front body (x d1 , z d1 ) is located at the center of the lower surface of the front body, and the force F on the x-axis and z-axis of the body x,d1 、F z，d1 , and the pitch moment M of the lower surface of the forebody on the x-axis and z-axis d1 The formula is as follows: F x,d1 =-P d1 S1tanτ2 (4) F z，d1 =-P d1 S1 (5) M d1 =z d1 F x，d1 -x d1 F z，d1 (6) For the lower surface of the rear body of the aircraft, the lower surface area is set to S2 and the pressure on the lower surface of the rear body is set to P d2 , the free flow pressure is P, the rear body angle is τ3, and the rear body force point (x d2 , z d2 ) is located at the center of the lower surface of the rear body, and the force F on the x-axis and z-axis of the body x，d2 、F z,d2 And the pitch moment M of the lower surface of the rear body on the x-axis and z-axis d2 The formula is as follows: M d2 =z d2 F x，d2 -x d2 F z，d2 (9) For the aircraft elevator surface, the elevator area is set to S3, and the pressures on the upper and lower surfaces of the elevator surface are P e_up , P e_d , the rudder angle is δ e , upgrade the rudder force point (x e , z e ) is located at the center of the elevator, and the force F on the x-axis and z-axis of the fuselage x，e 、F z，e , and the pitch moment M of the elevator surface on the x-axis and z-axis e The formula is as follows: F x，e =P e_up S3sinδ e -P e_d S3sinδ e (10) F z，e =P e_up S3cosδ e -P e_d S3cosδ e (11) M e =z e F x，e -x e F z，e (12) For the aircraft wing surface, the wing span length L b The sweep angle ε and the wing area are variable, and the wing area is the span length multiplied by the average chord length c avg , the air density is ρ, and the wing pressure is P b , the flight speed is V, the force acting point on the wing surface (x b , z b ) is located at the center of the wing, and the force F on the x-axis and z-axis of the fuselage x，b 、F z，b , and the pitch moment M of the wing surface on the x-axis and z-axis b The formula is as follows: L b1 =L b cosε(13) F z，b =P b L b1 c avg (15) Among them, L b1 is the wingspan length, and its random wing span length L b and changes with the sweep angle ε.

4. The aircraft modeling method of the mechanism combined with PSO feedforward neural network optimization according to claim 1 or 3 is characterized in that: The process of establishing the aircraft body model in step S1 is as follows: The fuselage body will undergo elastic deformation when flying at high speed, resulting in the deformation angle ΔT of the front and rear bodies. f , ΔT a , at this time, the center of mass of the aircraft will change, which will affect the forces and moments on the fuselage; Definition: Forebody deformation angle Δτ f for Definition: Rear body deformation angle Δτ a for in, represents the vibration mode, η i (t) represents the generalized coordinates. The rigid body model is elastically corrected, and the corrected upper and lower body angles τ′1 and τ′2 of the aircraft front body are: τ′1=τ1-Δτ f (19) τ′2=τ2+Δτ f (20) The corrected rear body angle τ′3 is: τ′3=τ3-Δτ a (21) Substitute the corrected front body upper and lower body angles τ′1, τ′2 and rear body angle τ′3 into the aircraft rigid body model formula to obtain the elastically corrected body x-axis force F′ x，up , F′ x，d1 , F′ x，d2 , F′ x，e , F′ x，b , the corrected body z-axis force F′ z，up , F′ z，d1 , F′ z，d2 , F′ z，e , F′ z，b , the corrected body pitch moment M′ up , M′ d1 , M′ d2 , M′ e , M′ b .

5. The aircraft modeling method of the mechanism combined with PSO feedforward neural network optimization according to claim 4 is characterized by: The process of establishing the aircraft dynamics mechanism model in step S1 is as follows: According to the elastically corrected rigid body and projectile model, the forces and moments on each surface of the aircraft are added to obtain the force F acting on the x-axis and z-axis of the aircraft. x 、F z and pitch moment M are: F x =F′ x，up +F′ x，d1 +F′ x，d2 +F′ x，e +F′ x，b (22) F z =F′ z，up +F′ z，d1 +F′ z，d2 +F′ z，e +F′ z，b (23) According to the coordinate transformation formula and the angle of attack α, the lift L, drag D and pitch moment M in the aircraft velocity coordinate system can be obtained as follows: L=F x sinα-F z cosα (24) D=-F x cosα-F z sinα (25) M=M′ up +M′ d1 +M′ d2 +M′ e +M′ b (26) Based on the above-mentioned lift L, drag D and pitching moment M of the aircraft, the aircraft dynamic equation can be obtained as follows: in, θ, h, T, I yy They respectively represent the change in aircraft speed, change in angle of attack, change in pitch angle, change in attitude angle, altitude, thrust and the aircraft's moment of inertia around the horizontal axis.

6. The aircraft modeling method of the mechanism combined with PSO feedforward neural network optimization according to claim 5 is characterized by: The process of establishing the neural network residual compensation model in step S2 is: S21. Dataset processing and division: S22, Feedforward neural network design and PSO optimization of feedforward neural network; S23. Establish a residual compensation model based on PSO-feedforward neural network.

7. The aircraft modeling method of claim 6, characterized in that: The data set processing in step S21 includes denoising, normalization, missing value filling, and outlier detection and processing operations on the actual flight data; The data set is divided into training set, test set and validation set according to 6:2:2 using the built-in divideint() function of Matlab.

8. The aircraft modeling method of the mechanism combined with PSO feedforward neural network optimization according to claim 6 is characterized by: The feedforward neural network design process of step S22 is: The feedforward neural network is set to have four layers, namely the input layer, two hidden layers and the output layer. The input layer has six input values, namely the sweep angle, wingspan length, rudder angle, Mach number, angle of attack and altitude of the aircraft. (1) Hidden layer structure design The hidden layer of the neural network is set to two layers, and the number of nodes is dynamically adjusted by the PSO algorithm to perform the PSO optimization feedforward neural network process; (2) Loss function design Use mean square error as the loss function, optimize the neural network weights and bias by minimizing the loss function, and then evaluate the error of the model on the training set and validation set. The loss function J of the training set and validation set train , J val They are as follows: Among them, n is the number of samples in the data set, y i_train ,y i_val is the true value, is the calculated value of the mechanism. (3) Training method design Using Bayesian regularized back propagation as the training method, the loss function J' of the training set after Bayesian optimization train as follows: J' train =MSE train +λ*||w|| 2 (29) Among them, MSE train is the mean square error of the training set, ||w|| 2 is the regularization term, w is the network weight parameter, and λ is the regularization coefficient, which is used to control the regularization weight. (4) Neural network training Set the error convergence target, use the training set as input for training, use the validation set to evaluate network performance, and adjust the network structure in real time based on the validation set error, iterating repeatedly until the optimal network is trained; The PSO optimization feedforward neural network process in step S22 is: (1) Particle swarm initialization Set the particle swarm size, number of iterations and initial speed value. The global optimal error initial value is inf. The global optimal parameter initial value is the initial value of the first particle. The particle position initial value is randomly generated in the range of [10,100]. Its dimension is 2, which represents the number of nodes in the two hidden layers of the feedforward neural network. (2) Fitness function calculation The current position value of each particle is used as the number of nodes in the hidden layer of the neural network. The minimum target value J, that is, the mean square error of the validation set, is obtained by training the neural network with data, and then the fitness function can be constructed as follows: (3) Position and velocity update Use the PSO update formula to update the particle position and velocity. The specific velocity and position update formula is v i (t+1) and x i (t+1) is as follows: v i (t+1)=wv i (t)+c1r1(p i -x i (t))+c2r2(g-x i (t)) (31) x i (t+1)=x i (t)+v i (t+1) (32) According to the formula, x i (t) and v i (t) represent the current position and velocity of the particle, p i is the historical optimal position of particle i, g is the global optimal position, w is the inertia weight, c1 and c2 are acceleration factors, the default values ​​are used here, r1 and r2 are random numbers in [0,1]; (4) Iterative Optimization The above steps are repeated to update the historical optimal position and global optimal position of the particle according to the fitness until the maximum number of iterations is reached or the fitness meets the preset threshold. (5) Output optimal parameters The output layer is the residual between the actual value and the mechanism value, which is expressed as follows: e i =y 实际i -y 机理i (33)The above formula is the residual value corresponding to the i-th group of data.

9. The aircraft modeling method of the mechanism combined with PSO feedforward neural network optimization according to claim 8 is characterized by: The process of establishing the residual compensation model based on the PSO-feedforward neural network in step S23 is as follows: The six variables of aircraft sweep angle, wingspan length, rudder angle, Mach number, angle of attack and altitude are used as the input of PSO-feedforward neural network. The mechanism value of the aircraft is calculated through the embedded mechanism model of PSO-feedforward neural network. The actual value is subtracted from the mechanism value to make the PSO-feedforward neural network output a residual compensation model.

10. The aircraft modeling method of the mechanism combined with PSO feedforward neural network optimization according to claim 1 or 9, characterized in that: The variable configuration aircraft model constructed in step S23 is: Model 混合 =Model 机理 +Model 残差补偿 (34) Among them, Model 残差补偿 Represents the residual compensation model obtained by neural network, Model 机理 Refers to the aircraft dynamics mechanism model.

Images