Physical monotonicity guided morphing aircraft aerodynamic force intelligent prediction method and system

Abstract

The invention discloses a physical monotonicity guided morphing aircraft aerodynamic force intelligent prediction method and system, and the method comprises the steps: constructing a prediction model which is based on a deep lattice crossing network algorithm DLC frame, the frame comprises two branches, and the branch 1 predicts a resistance coefficient Cd based on a deep crossing network; the branch 2 predicts a lift coefficient Cl based on a deep lattice network DLN, and a lattice layer of an attack angle alpha is added in an algorithm of the DLN and is used for constraining a monotone increasing relation of Cl along with alpha; and a training set is constructed to train the prediction model, after training is completed, the attack angle alpha, the wing unfolding angle beta and the wing unfolding time t of the variant aircraft are input into the prediction model, and the lift coefficient Cl and the resistance coefficient Cd of the variant aircraft in the current state are output. The method can improve the external universal capability of the prediction model, enables the prediction result to meet the basic physical law, and improves the interpretability and universality of a machine learning model in the aerodynamic prediction field.

Description

Technical Field

[0001] The present invention belongs to the field of aircraft, and in particular relates to a method and system for intelligent prediction of aerodynamic force of a variant aircraft guided by physical monotonicity. Background Art

[0002] With the development of artificial intelligence technology and computing power hardware and software, artificial intelligence technology has been widely used in all walks of life. Artificial intelligence technology has significant advantages in rapid prediction and mining of laws. Aircraft design has entered the "intelligent era" and "digital twin era" driven by "data" + "model". Morphing aircraft are a type of aircraft that adapts to different flight environments and tasks by changing their own form, and they occupy an important position in the field of high-end aircraft. The rapid acquisition of the aerodynamic load of the aircraft can provide input for the intelligent flight control system and the ground digital twin system, but the acquisition of the aerodynamic load of the variant aircraft is more difficult. Due to the dynamic process of the variant, unsteady calculations often consume a lot of computing resources, and variant tests are even more difficult and costly. Artificial intelligence technology provides a new way for rapid and low-cost prediction of the aerodynamics of variant aircraft.

[0003] For example, a Chinese patent document with publication number CN115796021A discloses a method for setting adaptive PID control parameters of a variant aircraft based on a neural network. The method establishes a dynamic model of the variant aircraft, calculates the aircraft's moment of inertia and center of mass position changes in real time according to the deformation state, thereby ensuring the accuracy of the dynamic model. Then, a BP neural network is built. Based on the above-mentioned dynamic model of the variant aircraft, four groups of five state parameters are trained to correspond to one output neural network, and a large amount of simulated flight data is input for training to accurately predict the flight dynamics parameters.

[0004] A Chinese patent document with publication number CN112329140A discloses an aerodynamic optimization method for a variant aircraft based on an improved expected improvement of a position vector, wherein an initial reference airfoil and related shape coefficients are selected to determine a design condition; an optimization model considering structural consistency constraints is established to solve the shape coefficient of the deformed airfoil, and a high-precision aerodynamic analysis model is established to solve the lift coefficient and drag coefficient of the airfoil; a Kriging proxy model of the high-precision aerodynamic analysis model is established, the Kriging proxy model is optimized to obtain a pseudo-non-dominated solution, and it is determined whether the prediction variance of the Kriging proxy model is greater than a threshold; if the prediction variance is not less than a threshold, the position vector expected improvement criterion is used to obtain newly added sample points for updating the Kriging proxy model, otherwise, newly added sample points are selected from the pseudo-non-dominated solution to update the Kriging proxy model.

[0005] However, the application of artificial intelligence technology to aerodynamic prediction and aerodynamic design of variant aircraft will face the following challenges: the flow field of three-dimensional aircraft is complex and the unsteady effect is strong, the design parameter dimension of the variant configuration is high and the shape variation range is large. The above characteristics lead to poor physical interpretability and poor versatility of artificial intelligence models. Summary of the invention

[0006] In order to solve the problems of poor physical interpretability and poor versatility of existing machine learning models in aerodynamic prediction, the present invention provides a variant aircraft aerodynamic intelligent prediction method and system guided by physical monotonicity, which can improve the generalization ability of the prediction model, make the prediction results meet the basic physical laws, and enhance the interpretability and versatility of machine learning models in the field of aerodynamic prediction.

[0007] A physical monotonicity-guided aerodynamic intelligent prediction method for a variant aircraft, comprising:

[0008] Construct a prediction model based on the deep lattice cross network algorithm DLC framework, which contains two branches, among which branch 1 predicts the drag coefficient C based on the deep cross network algorithm DLC. d ; Branch 2 predicts lift coefficient C based on deep lattice network DLN l , and add the lattice layer of angle of attack α to the DLN algorithm to constrain C l A monotonically increasing relationship with α;

[0009] Construct a training set to train the prediction model. After the training is completed, the attack angle α, wing angle β and wing time t of the variant aircraft are input into the prediction model, and the aerodynamic load coefficient of the variant aircraft in the current state - lift coefficient C can be quickly output. l and the drag coefficient C d .

[0010] Furthermore, the lattice layer includes three modules, namely: a monotonic embedding layer, a calibration layer and a lattice collection layer; in order to ensure that the output result is monotonically related to the input parameters, the weight factor of the monotonic embedding layer network is non-negative; the calibration layer is a piecewise linear transformation relationship, which contains multiple piecewise control points. In order to ensure the monotonicity of input and output, linear inequality constraints are added; the lattice collection layer adopts a multilinear interpolation method to define the monotonic constraint relationship between each pair of adjacent parameters on the input lattice.

[0011] Furthermore, after the prediction model inputs the angle of attack α, the wing angle β, and the wing time t, it is first processed by data normalization. The processed wing angle β and the wing time t are passed to the parameter sharing layer, and are respectively input to the NCd layer of branch 1 and the NCl layer of branch 2; the NCl layer is combined with the lattice layer considering the physical monotonicity constraint of the angle of attack α to predict the lift coefficient C lThe NCd layer combines with the Nα layer at an angle of attack α to transmit data to the connection layer, which is further combined with the cross layer to finally predict the drag coefficient C d .

[0012] Furthermore, in branch 1, the cross layer contains product terms, linear terms, and bias terms, and the formula is as follows:

[0013]

[0014] Among them, x i 、x j is the connection vector between NCd layer and Nα layer, w L is the weight matrix of the linear part, w pij is the weight matrix of the dot product part; b Cd is the deviation term.

[0015] Furthermore, the NCl layer, NCd layer and Na layer are constructed using a multi-layer perceptron algorithm.

[0016] Furthermore, the data source of the training set is CFD simulation data. CFD simulates dynamic unsteady aerodynamic characteristics based on overlapping structured grids, thereby obtaining the aerodynamic load coefficients of the aircraft under different variant forms and working conditions - lift coefficient and drag coefficient; the angle of attack of the training set ranges from -4° to 12°, the wing angle ranges from 0° to 90°, and the wing time is 0 to 2s.

[0017] During the training process, the loss function of branch 1 training is the absolute value of the average relative error with the drag coefficient calculated by CFD. After completing the network training of branch 1, the parameters of the branch 1 network are kept unchanged, and the network of branch 2 is trained. The loss function of branch 2 training is the absolute value of the average error with the lift coefficient calculated by CFD.

[0018] The network weights and biases are trained through the back-propagation algorithm, and the optimizer used adopts the adaptive moment estimation method; after the prediction model is trained, the optimal hyperparameter combination value is obtained, thereby obtaining the prediction model of the aerodynamic lift and drag of the variant aircraft.

[0019] A variant aircraft aerodynamic force intelligent prediction system guided by physical monotonicity includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned variant aircraft aerodynamic force intelligent prediction method.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] The present invention is based on the deep lattice cross network algorithm (DLC), combines the advantages of the cross network algorithm, and can set different lattice constraints for different input parameters. In the present invention, the monotonically increasing relationship between the lift coefficient and the angle of attack is injected into the prediction network of the lift coefficient without affecting the prediction accuracy of the drag coefficient. The prediction can be performed in parallel and the prediction accuracy of each target value can be improved. Compared with the multi-layer perceptron algorithm and the multi-task cross algorithm (MTC), the prediction accuracy and efficiency of the present invention are higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a schematic diagram of a variant aircraft in an embodiment of the present invention;

[0023] Figure 2 A schematic diagram of the structure of a prediction model in an embodiment of the present invention;

[0024] Figure 3 The structured overlapping grid used in the CFD simulation in the embodiment of the present invention;

[0025] Figure 4 Comparison between CFD simulation data and wind tunnel test data at different angles of attack when the wing span angles are 0°, 30°, and 90° in an embodiment of the present invention;

[0026] Figure 5 The prediction results of the aerodynamic lift coefficient and drag coefficient of the prediction model of the present invention under two sets of extrapolated angles of attack are compared with the CFD test set;

[0027] Figure 6 When the angle of attack is extremely deviated from the training set range, the lift coefficient prediction results of the prediction model of the present invention and the comparison model MTC are compared under different variant forms. DETAILED DESCRIPTION

[0028] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0029] Figure 1 As shown, the variant aircraft involved in the present invention is composed of a fuselage 1, wings 2, and a tail 3, wherein the wings 2 can rotate around a vertical axis, thereby changing the sweep angle of the wings 2 to adapt to different flight missions. The method proposed in the present invention is used to predict the aerodynamic lift and aerodynamic drag of the variant aircraft during the sweep change process.

[0030] (1) Preparation of training set and test set data

[0031] The training set and test set data sets are obtained by CFD method. The CFD method uses structured overlapping grids to divide the flow field area (such as the attached Figure 3 The Reynolds-averaged Navier-Stokes equations are solved by the finite volume method to solve the flow field of the morphing vehicle, and the turbulence model is kw SST.

[0032] The structured overlapping grid used in CFD simulation mainly includes the mesh of variable swept wing components, fixed fuselage components and tail components. By changing the sweep angle of the wing components, the numerical solution of different variant forms can be achieved. Through CFD calculation, the lift coefficient and drag coefficient under different wing angles, wing time and angle of attack can be obtained.

[0033] The angle of attack of the training set ranges from -4° to 12°, the angle of the wingspan ranges from 0° to 90°, and the wingspan time is 0 to 2s. In order to verify the generalization ability of the model, the test set additionally calculated angles of attack outside the training set range, including -15°, -10°, -5°, and 14°. The number of training sample sets is 1,809, the number of test sets is 804, and the number of validation sets is 402. There are three groups of input parameters for the prediction model, namely: angle of attack α, time t, and wing sweep angle β. There are two groups of output parameters of the model, namely, the lift coefficient C during the aircraft deformation process l , drag coefficient C d .

[0034] (2) Network framework

[0035] like Figure 2 As shown in the figure, it is the network structure of the prediction model of the present invention. The prediction model is based on the deep lattice cross network algorithm (DLC) framework. The network framework contains two cross network branches, which are used to predict the lift coefficient and drag coefficient under the current variant form. The input of the network is the angle of attack (α), the wing angle (β), and the wing time (t). First, the data is normalized. The processed wing angle β and wing time t are passed to the parameter sharing layer and respectively input into the NCd layer of branch 1 and the NCl layer of branch 2; the NCl layer is combined with the lattice layer that considers the physical monotonicity constraint of the angle of attack α to predict the lift coefficient C l The NCd layer combines with the Nα layer at an angle of attack α to transmit data to the connection layer, which is further combined with the cross layer to finally predict the drag coefficient C d .

[0036] In branch 1, the cross network layer for predicting the drag coefficient can make full use of sparse input features and limited training data, and can effectively integrate the flow field characteristics of the aircraft. The cross layer consists of a product term, a linear term, and a bias term, and its formula is:

[0037]

[0038] Among them, x i 、xj is the connection vector between NCd and Nα, w L is the weight matrix of the linear part, w pij is the weight matrix of the dot product part, b Cd is the deviation term.

[0039] In branch 2, the lattice layer for predicting lift coefficients consists of three modules: a monotonic embedding layer, a calibration layer, and a lattice ensemble layer.

[0040] The monotone embedding layer is a simple neural network that contains an input layer and an output layer. In order to ensure that the output result is monotonically related to the input parameters, the weight factor of the network is non-negative, and the linear constraint relationship of the monotone embedding layer is:

[0041] y E =w m0 x E +b

[0042] Among them, w m0 is a non-negative weight factor, b is a bias term, x E is the input quantity, y E is the output. In the present invention, x E is the angle of attack α, y E is the lift coefficient.

[0043] The calibration layer is a piecewise linear transformation relationship that includes K p To ensure the monotonicity of input and output, linear inequality constraints are added.

[0044] The lattice collection layer uses multilinear interpolation, and its main function is to function, where the S-dimensional input parameter x L , converted to 2 s Dimensional features, the specific formula is as follows:

[0045]

[0046] d is x L The d-th order component, j is the output The j-th dimension parameter, v j is the vector coordinate of the S-dimensional space. To ensure that the input x L (d) Linear relationship with output, introducing interpolation function:

[0047]

[0048] For this model, the angle of attack and lift coefficient have a monotonically increasing relationship. Define each pair of adjacent θ on the input lattice as L (k')>θ L (k).

[0049] (3) Training of prediction models

[0050] The network framework contains two cross network branches, which are used to predict the lift coefficient and drag coefficient under the current variant form. The prediction of the drag coefficient (branch 1) adopts a deep cross network, and the prediction network of the lift coefficient (branch 2) adopts a deep lattice network (DLN) method. The input of the network is the angle of attack (α), the wing angle (β), and the wing time (t). The output of the network is the lift coefficient and drag coefficient of the variant aircraft in the current state. The mapping relationship of the network model is as follows:

[0051]

[0052] in are the training network parameters for predicting lift coefficient and drag coefficient, θ share is the training network parameter of the shared parameter layer Nshare. In the branch 1 network, the main training parameters are With θ share ; In branch network 2, the main training parameters With θ share The loss function of the training is the absolute value of the average relative error with the drag coefficient calculated by CFD.

[0053] After completing the network training of branch 1, With θ share In branch 2, the main training The loss function for training is the absolute value of the average error with the lift coefficient calculated by CFD.

[0054] The network weights and biases are trained using the back-propagation algorithm. The optimizer used uses the adaptive moment estimation method, a stochastic gradient descent method that calculates an adaptive learning rate for each parameter.

[0055] In the network, the hyperparameters mainly include the number of layers and nodes of the network, as well as the activation function and activation factor of each layer. In the present invention, the optimal hyperparameter combination value is obtained through multiple iterations of optimization. In order to balance the training efficiency and accuracy, the learning rate is set to 0.001. Each training is set to 1000 iterations until the loss function reaches convergence, thereby completing the training of the model and obtaining the prediction model of the aerodynamic load of the variant aircraft.

[0056] like Figure 4As shown in the figure, the comparison between CFD simulation data and wind tunnel test data at different angles of attack is shown when the wing span angle is 0°, 30°, and 90°. The solid line in the figure is CFD simulation data, and the symbols of different shapes are wind tunnel test data. It can be seen that the CFD simulation data is in good agreement with the wind tunnel test data, and the CFD simulation data can be used as training samples for high-precision prediction models.

[0057] like Figure 5 As shown, the prediction results of the aerodynamic lift coefficient and the drag coefficient of the prediction model (DLC) of the present invention under two sets of extrapolated angles of attack α = [-5°, 14°] are compared with the CFD test set. In the figure, (a) is the result at an angle of attack of -5°; (b) is the result at an angle of attack of 14°. By comparing with the CFD test set, it can be found that even at the extrapolated angle of attack, the prediction results are in perfect agreement with the CFD simulation results.

[0058] The present invention is based on the deep lattice cross network algorithm (DLC) and combines the advantages of the cross network algorithm. Different lattice constraints can be set for different input parameters. In the present invention, the monotonically increasing relationship between the lift coefficient and the angle of attack is injected into the prediction network of the lift coefficient without affecting the prediction accuracy of the drag coefficient. The prediction can be performed in parallel and the prediction accuracy of each target value can be improved. As shown in Table 1, the prediction errors of different algorithms in the test set are shown. Compared with the multi-layer perceptron algorithm and the multi-task cross algorithm (MTC), the prediction model (DLC) of the present invention has the highest prediction accuracy.

[0059] Table 1

[0060]

[0061] Taking physical constraints into account can improve the generalization ability of the model. The angle of attack of the training set ranges from -4° to 12°. In order to compare the generalization ability of the test model, the prediction accuracy at angles of attack of -10° and -15° was tested. The results are shown in the attached figure. Figure 6 Shown

[0062] like Figure 6 As shown, the lift coefficient prediction results based on the prediction model of the present invention (DLC) and the comparison model MTC under different variant forms are compared when the angle of attack is -10° and -15°, which is extremely deviated from the training set angle range. In the figure, (a) is the result when t=0.5s; (b) is the result when t=2s. It can be seen from the figure that the results predicted by the MTC model have obviously deviated from the CFD results, while the results of the prediction model of the present invention (DLC) are closer to the CFD results. The prediction results of the lift coefficient of the prediction model proposed by the present invention under the external working conditions are higher than the prediction accuracy of MTC, which proves that the prediction model proposed by the present invention has a stronger generalization ability and the prediction results are closer to the real physical characteristics.

[0063] Traditional aerodynamic prediction methods generally use numerical simulation methods such as CFD, but the CFD method is time-consuming. The prediction method proposed in the present invention can replace the inefficient CFD aerodynamic prediction model and greatly improve the prediction efficiency of aerodynamic force. Table 2 shows the time consumption of the prediction model of the present invention and the CFD calculation model in a variant aircraft during a variant process. CFD requires 30 hours, but the trained prediction model only requires 0.1 seconds, which has great advantages in efficiency compared to the traditional CFD model.

[0064] Table 2

[0065]

[0066] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for intelligent prediction of aerodynamic force of a variant aircraft guided by physical monotonicity, characterized in that: include: Construct a prediction model based on the deep lattice cross network algorithm DLC framework, which contains two branches, among which branch 1 predicts the drag coefficient C based on the deep cross network algorithm d ; Branch 2 predicts lift coefficient C based on deep lattice network DLN l , and add the lattice layer of angle of attack α to the DLN algorithm to constrain C l A monotonically increasing relationship with α; Construct a training set to train the prediction model. After the training is completed, the angle of attack α, wing angle β and wing time t of the variant aircraft are input into the prediction model to quickly output the lift coefficient C of the variant aircraft in the current state. l and the drag coefficient C d .

2. The method for intelligent prediction of aerodynamic force of a variant aircraft guided by physical monotonicity according to claim 1 is characterized in that: The lattice layer contains three modules, namely: monotone embedding layer, calibration layer and lattice collection layer; In order to ensure that the output result is monotonically related to the input parameters, the weight factor of the monotonic embedding layer network is non-negative; the calibration layer is a piecewise linear transformation relationship, which contains multiple piecewise control points. In order to ensure the monotonicity of input and output, linear inequality constraints are added; the lattice set layer uses multilinear interpolation method to define the monotonic constraint relationship of each pair of adjacent parameters on the input lattice.

3. The method for intelligent prediction of aerodynamic force of a variant aircraft guided by physical monotonicity according to claim 1 is characterized in that: After the prediction model inputs the angle of attack α, the wing angle β, and the wing time t, it is first processed by data normalization. The processed wing angle β and the wing time t are passed to the parameter sharing layer and respectively input to the NCd layer of branch 1 and the NCl layer of branch 2; the NCl layer is combined with the lattice layer considering the physical monotonicity constraint of the angle of attack α to predict the lift coefficient C l The NCd layer combines with the Nα layer at an angle of attack α to transmit data to the connection layer, which is further combined with the cross layer to finally predict the drag coefficient C d .

4. The method for intelligent prediction of aerodynamic force of a variant aircraft guided by physical monotonicity according to claim 3 is characterized in that: In branch 1, the cross layer contains product terms, linear terms, and bias terms, and the formula is as follows: Among them, x i 、x j is the connection vector between NCd layer and Nα layer, w L is the weight matrix of the linear part, w pij is the weight matrix of the dot product part; b Cd is the deviation term.

5. The method for intelligent prediction of aerodynamic force of a variant aircraft guided by physical monotonicity according to claim 3 is characterized in that: The NCl layer, NCd layer and Naα layer are constructed using a multi-layer perceptron algorithm.

6. The method for intelligent prediction of aerodynamic force of a variant aircraft guided by physical monotonicity according to claim 1, characterized in that: The data source of the training set is CFD simulation data. CFD simulates dynamic unsteady aerodynamic characteristics based on overlapping structured grids, thereby obtaining the aerodynamic load coefficients of the aircraft under different variant forms and working conditions - the lift coefficient and the drag coefficient; the angle of attack of the training set ranges from -4° to 12°, the wing angle ranges from 0° to 90°, and the wing time is 0 to 2s.

7. The method for intelligent prediction of aerodynamic force of a variant aircraft guided by physical monotonicity according to claim 1, characterized in that: During the training process, the loss function of branch 1 training is the absolute value of the average relative error with the drag coefficient calculated by CFD. After completing the network training of branch 1, the parameters of the branch 1 network are kept unchanged, and the network of branch 2 is trained. The loss function of branch 2 training is the absolute value of the average error with the lift coefficient calculated by CFD. The network weights and biases are trained through the back-propagation algorithm, and the optimizer used adopts the adaptive moment estimation method; after the prediction model is trained, the optimal hyperparameter combination value is obtained, thereby obtaining the prediction model of the aerodynamic lift and drag of the variant aircraft.

8. A physical monotonicity-guided aerodynamic intelligent prediction system for a variant aircraft, characterized in that: The invention comprises a memory and one or more processors, wherein the memory stores executable codes, and when the one or more processors execute the executable codes, the method for intelligent prediction of aerodynamic force of a variant aircraft according to any one of claims 1 to 7 is implemented.

Images