A physical monotonicity-guided method and system for intelligent prediction of aerodynamic forces of a variable aircraft

Abstract

The application discloses a physical monotonicity guided variant aircraft aerodynamic force intelligent prediction method and system, and comprises the following steps: constructing a prediction model, wherein the prediction model is based on a deep lattice cross network algorithm DLC framework, the framework comprises two branches, branch 1 is based on a deep cross network to predict a drag coefficient C d ; branch 2 is based on a deep lattice network DLN to predict a lift coefficient C l , and a lattice layer of an angle of attack alpha is added in the algorithm of the DLN and used for constraining the monotonic increasing relationship between C l and the angle of attack alpha; a training set is constructed to train the prediction model, after the training is completed, the angle of attack alpha, a wing spread angle beta and a wing spread time t of the variant aircraft are input into the prediction model, and a lift coefficient C l and a drag coefficient C d of the variant aircraft in the current state are output. The application can improve the out-of-sample ability of the prediction model, so that the predicted result meets the basic physical law, and the explainability and universality of the machine learning model in the aerodynamic prediction field are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of aircraft, and particularly relates to a physical monotonicity guided variable aircraft aerodynamic force intelligent prediction method and system. BACKGROUND

[0002] With the development of artificial intelligence technology and computing power software and hardware, artificial intelligence technology has been widely applied in various industries. Artificial intelligence technology has significant advantages in rapid prediction and rule mining. Aircraft design has entered the "intelligent era" and "digital twin era" driven by "data" and "models". Variable aircraft is a kind of aircraft that adapts to different flight environments and tasks by changing its own shape, and occupies an important position in the field of high-end aircraft. Rapid acquisition of aerodynamic load of aircraft can provide input for intelligent flight control system and ground digital twin system, but the acquisition of aerodynamic load of variable aircraft is difficult, as it involves variable dynamic process, and non-steady-state calculation often consumes a large amount of computing resources, while variable test is more difficult and extremely costly. Artificial intelligence technology provides a new way for rapid and low-cost prediction of variable aircraft aerodynamics.

[0003] A variable aircraft adaptive PID control parameter setting method based on neural network is disclosed in Chinese patent document with publication number CN115796021A, a dynamic model of the variable aircraft is established, the moment of inertia and the change of the center of mass of the aircraft are calculated in real time according to the deformation state, so as to ensure the accuracy of the dynamic model; then a BP neural network is built, based on the above-mentioned dynamic model of the deformed aircraft, a method of training four groups of five state parameters corresponding to one output neural network is adopted, a large amount of simulation flight data is input for training, and flight dynamics parameters are accurately predicted.

[0004] A variable aircraft aerodynamic optimization method based on improved position vector expected improvement degree is disclosed in Chinese patent document with publication number CN112329140A, an initial reference airfoil and related shape coefficients are selected, and a design operating condition is determined; an optimization model considering structural consistency constraints is established to solve the shape coefficients of the deformed airfoil, a high-precision aerodynamic analysis model is established to solve the lift coefficient and drag coefficient of the airfoil; a Kriging surrogate model of the high-precision aerodynamic analysis model is established, pseudo-non-dominated solutions are obtained by optimizing the Kriging surrogate model, and it is judged whether the prediction variance of the Kriging surrogate model is greater than a threshold value; if the prediction variance is not less than the threshold value, the position vector expected improvement degree criterion is used to obtain new sample points for updating the Kriging surrogate model, otherwise, new sample points are selected from the pseudo-non-dominated solutions to update the Kriging surrogate model.

[0005] However, the application of artificial intelligence technology in the aerodynamic prediction and aerodynamic design of a variable aircraft will face the following challenges: the flow field of a three-dimensional aircraft is complex, the unsteady effect is strong, the design parameters of a variable configuration have high dimensions, and the shape change range is large. The above characteristics result in poor physical interpretability and poor universality of the artificial intelligence model. SUMMARY

[0006] To solve the problem of poor physical interpretability and poor universality of existing machine learning models in aerodynamic prediction, the present application provides a physical monotonicity-guided intelligent prediction method and system for aerodynamic force of a variable aircraft, which can improve the generalization ability of the prediction model, so that the predicted results meet the basic physical laws, and improve the interpretability and universality of the machine learning model in the field of aerodynamic prediction.

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

[0008] A prediction model is constructed, which is based on a deep lattice cross network algorithm DLC framework containing two branches. Branch 1 is based on a deep cross network to predict the drag coefficient C d ; and branch 2 is based on a deep lattice network DLN to predict the lift coefficient C l , and a lattice layer of the angle of attack a is added in the algorithm of the DLN, which is used to constrain the monotonic increasing relationship of C l with a;

[0009] A training set is constructed to train the prediction model. After the training is completed, the angle of attack a, the wing spread angle b and the wing spread time t of the variable aircraft are input into the prediction model, and the aerodynamic load coefficients of the variable aircraft in the current state, i.e. the lift coefficient C l and the drag coefficient C d , can be quickly output.

[0010] Further, the lattice layer contains three modules, specifically: a monotonic embedding layer, a calibration layer and a lattice set layer. In order to ensure that the output result has a monotonic relationship with the input parameters, the weight factor of the monotonic embedding layer network is non-negative. The calibration layer is a kind of piecewise linear transformation relationship, which contains a plurality of piecewise control points. In order to ensure the monotonicity of the input and output, a linear inequality constraint is added. The lattice set layer adopts a multi-linear interpolation method, and defines the monotonic constraint relationship of each pair of adjacent parameters on the input lattice.

[0011] Further, after the prediction model inputs the angle of attack a, the wing spread angle b and the wing spread time t, the processed wing spread angle b and wing spread time t are transmitted to the parameter sharing layer, and are respectively input into the NCd layer of branch 1 and the NCl layer of branch 2. The NCl layer predicts the lift coefficient C l; the NCd layer transmits data to the connection layer by combining with the Nalpha layer of the angle of attack alpha, further combines with the cross layer, and finally predicts the drag coefficient C d .

[0012] Further, in the branch 1, the cross layer comprises a product term, a linear term and a bias term, and the formula is as follows:

[0013]

[0014] Wherein, x i , x j is the connection vector of the NCd layer and the Nalpha 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 bias term.

[0015] Further, the construction of the NCl layer, the NCd layer and the Nalpha layer adopts a multilayer perception algorithm.

[0016] Further, the data source of the training set is CFD simulation data, CFD simulates dynamic unsteady aerodynamic characteristics based on an overlapped structured grid, so as to obtain the aerodynamic load coefficients of the aircraft under different variant morphologies and working conditions, i.e., the lift coefficient and the drag coefficient; the range of the angle of attack of the training set is-4°-12°, the range of the angle of the spread wing is 0°-90°, and the spread wing time is 0-2s.

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

[0018] The network weight and bias are trained by the back propagation algorithm, and the optimizer used adopts an adaptive moment estimation method; after the prediction model is trained, the optimal super parameter combination value is obtained, so that the prediction model of the aerodynamic lift and drag of the variant aircraft is obtained.

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

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

[0021] The present application is based on a deep lattice cross network algorithm (DLC), which combines the advantages of the cross network algorithm, can set different lattice constraints for different input parameters, and injects the monotonic increasing relationship between the lift coefficient and the angle of attack into the prediction network of the lift coefficient, without affecting the prediction accuracy of the drag coefficient, that is, it can be predicted in parallel, and the prediction accuracy of each target value can be improved. Compared with the multilayer perception algorithm and the multi-task cross algorithm (MTC), the prediction accuracy and efficiency of the present application are higher. BRIEF DESCRIPTION OF DRAWINGS

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

[0023] Figure 2 is a structural schematic diagram of a prediction model in the embodiment of the present application;

[0024] Figure 3 is a structured overlapping grid used in CFD simulation in the embodiment of the present application;

[0025] Figure 4 is a comparison of CFD simulation data and wind tunnel test data at different angles of attack under the condition that the wing extension angle is 0°, 30° and 90° in the embodiment of the present application;

[0026] Figure 5 is a comparison of the prediction results of the lift coefficient and the drag coefficient of the prediction model of the present application at two groups of extrapolated angles of attack with the CFD test set;

[0027] Figure 6 is a comparison of the lift coefficient prediction results of the prediction model of the present application and the comparison model MTC at different variant morphologies when the angle of attack deviates greatly from the training set. DETAILED DESCRIPTION

[0028] The present application will be further described in detail below in combination with the drawings and embodiments, and it should be pointed out that the following embodiments are intended to facilitate the understanding of the present application and do not limit the present application in any way.

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

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

[0031] The data set of the training set and the test set is obtained by CFD method. The CFD method adopts a structured subcomponent overlapping grid to divide the flow field region (as shown in the attachedFigure 3 The flow field of the morphing aircraft is solved by solving the Reynolds-averaged Navier-Stokes equations by the finite volume method, and the k-w SST turbulence model is adopted.

[0032] The structured overlapping grid used in the CFD simulation mainly includes a variable sweep wing component grid, a fixed fuselage component grid and a tail component grid. By changing the sweep angle of the wing component, numerical solutions of different morphing modes are realized. Through CFD calculation, the lift coefficient and the drag coefficient under different wing extension angles, wing extension times and angles of attack can be obtained.

[0033] The angle of attack of the training set ranges from -4° to 12°, the angle of the wing extension ranges from 0° to 90°, and the wing extension time ranges from 0 to 2s. In order to verify the generalization ability of the model, the angle of attack outside the training set range is additionally calculated in the test set, including -15°, -10°, -5°, 14°. The number of training sample set is 1809, the number of test set is 804, and the number of verification set is 402. The input parameters of the prediction model are three groups, which are: angle of attack α, time t and sweep angle β of the wing. The output parameters of the model are two groups, which are the lift coefficient C l and the drag coefficient C d .

[0034] (2) Network framework

[0035] As shown in Figure 2 , the network structure of the prediction model of the application, the prediction model is based on the deep lattice cross network algorithm (DLC) framework, and the network framework includes two cross network branches for predicting the lift coefficient and the drag coefficient under the current morphing mode. The input of the network is the angle of attack (α), the wing extension angle (β) and the wing extension time (t). First, the data is normalized, and the processed wing extension angle β and wing extension time t are transmitted to the parameter sharing layer and input into the NCd layer of branch 1 and the NCl layer of branch 2 respectively; the NCl layer is combined with the lattice layer considering the physical monotonicity constraint of the angle of attack α, and the lift coefficient C l is predicted; the NCd layer is combined with the Nα layer of the angle of attack α, and the data is transmitted to the connection layer, and further combined with the cross layer, and finally the drag coefficient C d is predicted.

[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 fuse the flow field characteristics of the aircraft. The cross layer is composed of a product term, a linear term and a bias term, and its formula is:

[0037]

[0038] wherein, x i , xj Let w be the vector connecting NCd and Nα. L Let w be the weight matrix for the linear part. pij Let b be the weight matrix of the dot product part. Cd This is the deviation term.

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

[0040] A monotonic embedding layer is a simple neural network consisting of an input layer and an output layer. To ensure that the output is monotonic with the input parameters, the network's weights are non-negative. The linear constraint relationship of a monotonic embedding layer is as follows:

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

[0042] Among them, w m0 Let b be a non-negative weighting factor, and x be a bias term. E For input quantity, y E This is the output quantity. In this invention, x E Let α be the angle of attack, and y be the angle of attack. E This 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 were added to the control points of each segment.

[0044] The lattice assembly layer employs multilinear interpolation, and its main function is to... The function takes the S-dimensional input parameter x as its input. L , converted to 2 s 3D 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 Let x be vector coordinates in S-dimensional space. To ensure the input x... L (d) The linear relationship with the output is explained by introducing an interpolation function:

[0047]

[0048] For the monotonically increasing relationship between the angle of attack and the lift coefficient in this model, we define each pair of adjacent θ values ​​on the input lattice as... L (k')>θ L (k).

[0049] (3) Training of the prediction model

[0050] The two cross-network branches in the network framework are used to predict the lift coefficient and the drag coefficient under the current variant configuration, respectively. The prediction of the drag coefficient (branch 1) uses a deep cross-network, and the prediction network of the lift coefficient (branch 2) uses a deep lattice network (DLN) method. The inputs of the network are the angle of attack (a), the wing sweep angle (b), and the wing sweep time (t). The outputs of the network are the lift coefficient and the drag coefficient of the variant aircraft under the current state. The mapping relationship of the network model is as follows:

[0051]

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

[0053] After the network training of branch 1 is completed, and share remain unchanged. In branch 2, the main training and the loss function for training is the absolute value of the average error of the lift coefficient calculated by the CFD.

[0054] The network weights and biases are trained by the backpropagation algorithm. The optimizer used is the adaptive moment estimation method, which is a stochastic gradient descent method that can calculate the adaptive learning rate of 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 application, 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 converges, thereby completing the training of the model and obtaining the prediction model of the aerodynamic load of the variant aircraft.

[0056] As Figure 4As shown, this figure compares CFD simulation data with wind tunnel test data at different angles of attack, with wingspans of 0°, 30°, and 90°. The solid lines in the figure represent CFD simulation data, and the different shaped markers represent wind tunnel test data. It can be seen that the CFD simulation data and wind tunnel test data agree well, indicating that 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 predictive model (DLC) of this invention for aerodynamic lift coefficient and drag coefficient are compared with the CFD test set under two sets of extrapolated angles of attack α = [-5°, 14°]. Figure (a) shows the result at an angle of attack of -5°; (b) shows the result at an angle of attack of 14°. By comparing with the CFD test set, it can be found that even at extrapolated angles of attack, the prediction results are in good agreement with the CFD simulation results.

[0058] This invention combines the advantages of deep lattice cross-network algorithms (DLC) with the advantages of cross-network algorithms, allowing for different lattice constraint settings for different input parameters. In this invention, the monotonically increasing relationship of the lift coefficient with the angle of attack is injected into the lift coefficient prediction network without affecting the prediction accuracy of the drag coefficient. This enables parallel prediction and improves the prediction accuracy of each target value. Table 1 shows the prediction errors of different algorithms on the test set. Compared with the multilayer perceptron algorithm and the multi-task cross-network algorithm (MTC), the prediction model (DLC) of this invention has the highest prediction accuracy.

[0059] Table 1

[0060]

[0061] Considering physical constraints can improve the model's generalization ability. The training set has an angle of attack range of -4° to 12°. To compare and test the model's outward generalization ability, the prediction accuracy was tested at angles of attack of -10° and -15°. The results are attached. Figure 6 As shown

[0062] like Figure 6 As shown, the figure compares the lift coefficient prediction results of the proposed prediction model (DLC) and the comparative model MTC under different variant configurations at angles of attack of -10° and -15°, when the angle of attack deviates significantly from the training set range. In the figure, (a) shows the result at t = 0.5s; (b) shows the result at t = 2s. It can be seen from the figure that the MTC model's prediction result deviates significantly from the CFD result, while the result of the proposed prediction model (DLC) is closer to the CFD result. The proposed prediction model has higher accuracy in predicting the lift coefficient under generalized operating conditions compared to MTC, demonstrating that the proposed prediction model has stronger generalization ability and the prediction result is closer to the physical reality.

[0063] The traditional aerodynamic force prediction method is generally a numerical simulation method such as CFD, but the CFD method is time-consuming. The prediction method provided in the present application can replace the inefficient CFD aerodynamic prediction model, and greatly improve the prediction efficiency of the aerodynamic force. Table 2 shows the time consumption of the variable aircraft in a variable process, respectively using the prediction model of the present application and the CFD calculation model. The CFD needs 30h, but the trained prediction model only needs 0.1s, which has great advantages in efficiency compared with the traditional CFD model.

[0064] Table 2

[0065]

[0066] The above embodiments have described the technical solutions and beneficial effects of the present application in detail. It should be understood that the above description is only a specific embodiment of the present application and is not used to limit the present application. Any modification, supplement and equivalent replacement made within the principle range of the present application shall be included in the protection scope of the present application.

Claims

1. A physically monotonicity-guided variant aircraft aerodynamic force intelligent prediction method, characterized in that, Comprise: Constructing a prediction model based on a deep lattice cross network algorithm DLC framework, which contains two branches, wherein branch 1 predicts the drag coefficient based on a deep cross network ; branch 2 predicts the lift coefficient based on a deep lattice network DLN , and adding a lattice layer of angle of attack in the algorithm of DLN for constraining the monotonic increasing relationship of with The lattice layer contains three modules, specifically: a monotone embedding layer, a calibration layer, and a lattice set layer; the weight factor of the monotone embedding layer network is non-negative; the calibration layer is a piecewise linear transformation relationship, contains multiple piecewise control points, and adds linear inequality constraints; the lattice set layer uses a multi-linear interpolation method to define the monotonicity constraint relationship of each pair of adjacent parameters on the input lattice; The training set is constructed to train the prediction model, and after the training is completed, the angle of attack , the wing spread angle and the wing spread time of the variant aircraft are input into the prediction model, and the lift coefficient and the drag coefficient of the variant aircraft in the current state are output.

2. The physically monotonicity-guided variant aircraft aerodynamic force intelligent prediction method according to claim 1, characterized in that, Predictive model input angle of attack wing angle Wing spread time Then, the data is first normalized to obtain the wingspan angle. With Wings Spread Time The data is passed to the parameter sharing layer and then input to the NCd layer of branch 1 and the NCL layer of branch 2, respectively; the NCL layer is connected to the parameter sharing layer considering the angle of attack. The physical monotonicity-constrained lattice layers are interlinked to predict the lift coefficient. ;NCd layer through angle of attack The Nα layer is combined to transmit data to the connection layer, and further combined with the cross-layer to finally predict the drag coefficient. .

3. The physically monotonicity-guided variant aircraft aerodynamic force intelligent prediction method according to claim 2, characterized in that, In branch 1, the cross layer contains product terms, linear terms, and bias terms, and the formula is as follows: ； wherein, , is the connection vector of the NCd layer and the Nα layer, is the weight matrix of the linear part, is the weight matrix of the dot product part; is the bias term.

4. The physically monotonicity-guided variant aircraft aerodynamic force intelligent prediction method according to claim 2, characterized in that, The construction of the NCl layer, the NCd layer, and the Nα layer adopts a multi-layer perception algorithm.

5. The physically monotonicity-guided variant aircraft aerodynamic force intelligent prediction method according to claim 1, characterized in that, The data source of the training set is CFD simulation data. CFD simulates the dynamic unsteady aerodynamic characteristics based on overlapping structured grids, thereby obtaining the aerodynamic load coefficients of the aircraft under different variant morphologies and working conditions, i.e., the lift coefficient and the drag coefficient; the angle of attack range of the training set is-4°~12°, the angle range of the wings is 0°~90°, and the wingspan time is 0~2s.

6. The physically monotonicity-guided variant aircraft aerodynamic force intelligent prediction method 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 of the drag coefficient calculated by CFD; after completing the network training of branch 1, the parameters of branch 1 network are kept unchanged, and the network of branch 2 is trained again, and the loss function of branch 2 training is the absolute value of the average error of the lift coefficient calculated by CFD; The network weights and biases are trained through the back propagation algorithm, and the optimizer uses 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.

7. A physically monotonicity-guided variant aircraft aerodynamic force intelligent prediction system, characterized in that, The device comprises a memory and one or more processors, the memory stores executable code, and the one or more processors execute the executable code to implement the variant aircraft aerodynamic force intelligent prediction method in any one of claims 1-6.

Images