EVTOL pneumatic modeling method based on multi-task learning

Abstract

The invention relates to the technical field of aviation pneumatic modeling, in particular to an eVTOL pneumatic modeling method based on multi-task learning. The method comprises the following steps: analyzing the aerodynamic effect and mutual influence of each aerodynamic component of eVTOL, identifying key characteristic parameters influencing the aerodynamic performance of the aircraft, and obtaining an aerodynamic effect analysis result; constructing a multi-task learning model based on aerodynamic effect characteristics by utilizing an aerodynamic effect analysis result and combining a multi-layer perceptron neural network and a long-short-term memory neural network; acquiring original data of the aircraft, and constructing a sample data set according to a flight state and aerodynamic characteristics; and training the multi-task learning model by using the sample data set, verifying and evaluating, and obtaining and outputting the multi-task learning model of the eVTOL. According to the method, aerodynamic interference and coupling characteristics can be reflected more accurately, and rapid and accurate prediction of the complex aerodynamic characteristics of the eVTOL is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of aviation aerodynamic modeling, and more specifically, to an eVTOL (electric Vertical Takeoff and Landing) aerodynamic modeling method and system based on multi-task learning, in particular to an eVTOL high-efficiency and high-fidelity aerodynamic modeling method and system based on multi-task learning. Background Art

[0002] With the acceleration of global urbanization, the continuous growth of urban population has brought unprecedented challenges to the transportation system. How to effectively alleviate traffic congestion, improve safety performance and reduce carbon emissions while meeting the growing travel needs of passengers has become a key issue that needs to be urgently addressed in the global transportation field. Against this background, Urban Air Mobility (UAM) came into being, among which eVTOL has attracted much attention as an innovative solution. As the core carrier for describing the aerodynamic characteristics of aircraft, the aerodynamic model of eVTOL not only helps in the structural design of aircraft and the development and verification of flight control systems, but also provides model support for research such as flight dynamics simulation and driver training simulators. Therefore, the development of a high-fidelity aerodynamic model for eVTOL has become a vital need.

[0003] In the field of eVTOL aerodynamic modeling, the aerodynamic modeling methods of existing technologies can be mainly divided into four paradigms: experimental methods, theoretical analysis, numerical analysis, and data-driven methods. The experimental method mainly obtains basic data through wind tunnel tests or test flights to provide data support for other modeling methods; the theoretical analysis method is an inductive summary study based on calculus analysis. It analyzes the physical meaning of parameters through deductive reasoning, models typical flow characteristics, and provides theoretical prior knowledge for other methods. With the rapid development of computing technology, computational fluid dynamics (CFD) has become an important research method for the third paradigm. However, these methods often face problems such as high consumption of computing resources and long modeling cycles when dealing with the complex coupled aerodynamic characteristics modeling of eVTOL, especially in dealing with multi-source data fusion and accurate calculation of non-equilibrium turbulence. There are significant challenges.

[0004] In recent years, data-driven machine learning methods have been widely applied as the fourth paradigm in aerodynamic research. This method is based on the massive data generated by wind tunnel tests, flight tests, and CFD numerical simulations, combined with the prior knowledge provided by aerodynamic theory, and conducts modeling through data-driven neural networks. Research shows that the data-driven modeling method based on neural networks exhibits good prediction performance when dealing with nonlinear problems, and has significant advantages compared with traditional physical mechanism modeling methods and CFD simulation methods, including powerful nonlinear modeling capabilities, low computational costs, no need to explicitly express the system mathematical model, and good generalization capabilities, etc.

[0005] Current research in the eVTOL field mainly focuses on control algorithms and trajectory planning, etc., and there is relatively little research on aerodynamic modeling. For example, Chinese invention patent CN117010219A proposes an eVTOL aircraft obstacle avoidance simulation method to provide safety guarantees for actual flights through a simulation platform. In terms of aerodynamic modeling, existing research mainly adopts physical mechanism, CFD, and data-driven methods: Chinese invention patent CN104298805A discloses a CFD aerodynamic modeling method for hypersonic vehicles, covering a complete process such as numerical modeling, mesh generation, computational solution, and result analysis; Chinese invention patent CN117195763A discloses a meta-flight aerodynamic modeling method for fixed-wing aircraft considering wind interference, realizing the online prediction of aerodynamic forces of fixed-wing aircraft under unknown wind conditions; Chinese invention patent CN116484713A discloses an aerodynamic modeling method for the tail-retreat separation scenario based on neural networks, significantly reducing the workload of CFD numerical simulations using neural network methods; Chinese invention patent CN117910386A discloses a transfer learning aerodynamic modeling method based on dendritic neural networks, further reducing the computational cost while ensuring the prediction accuracy through the combination of dendritic neural networks and transfer learning.

[0006] Although the existing technologies have made significant progress in aspects such as eVTOL control algorithms and trajectory planning, there are still obvious deficiencies in eVTOL aerodynamic performance prediction. Traditional mechanism or CFD-based modeling methods are usually limited to single tasks such as predicting airfoil and fuselage aerodynamic coefficients, and only for a limited number of airfoil shapes or specific angle-of-attack and Mach number ranges. In addition, these methods require independent modeling of each aerodynamic component and simulate the aerodynamic coupling effect by quantifying parameters such as induced velocity and airflow angle. This method is difficult to apply to scenarios such as fine-tuning of flight control systems and human-in-the-loop flight simulations, and the simulation accuracy of the aerodynamic coupling phenomenon is limited by the accuracy of parameter calculations, and its fidelity is relatively low.

[0007] Existing data-driven pneumatic modeling methods mostly adopt a single deep neural network structure, making it difficult to adapt to the pneumatic effect characteristics and mutual interference of different modules such as eVTOL rotors and fuselages. Since eVTOL is essentially a much more complex multi-body dynamic system, it is often composed of multiple pneumatic components. Traditional neural network methods usually need to model and train individual components such as the airframe and rotors separately. Although the structure of a single network is relatively simple, the large number of networks seriously affects the training efficiency.

[0008] Therefore, there is an urgent need for a new eVTOL pneumatic modeling method to solve the above problems. Summary of the Invention

[0009] The purpose of the present invention is to provide a pneumatic modeling method for eVTOL based on multi-task learning, which solves the problem that the existing eVTOL pneumatic modeling methods are difficult to accurately simulate the pneumatic effect characteristics and mutual interference between multiple pneumatic components.

[0010] To achieve the above purpose, the present invention provides a pneumatic modeling method for eVTOL based on multi-task learning, including the following steps:

[0011] Step S1, analyze the pneumatic effects and mutual influences of each pneumatic component of the eVTOL, identify the key characteristic parameters affecting the aerodynamic performance of the aircraft, and obtain the pneumatic effect analysis results;

[0012] Step S2, use the pneumatic effect analysis results, combine a multi-layer perceptron neural network and a long short-term memory neural network to construct a multi-task learning model based on pneumatic effect characteristics;

[0013] Step S3, obtain the original data of the aircraft, and construct a sample data set according to the flight state and aerodynamic characteristics;

[0014] Step S4, use the sample data set in Step S3 to train the multi-task learning model in Step S2, and conduct verification and evaluation to obtain and output the multi-task learning model of the eVTOL.

[0015] In some embodiments, in Step S1, the pneumatic effects and mutual influences between the pneumatic components of the eVTOL further include:

[0016] The pneumatic effect of an isolated rotor, the mutual influence between rotors, and the pneumatic effect of the same components as a fixed-wing aircraft.

[0017] In some embodiments, in Step S1, the key characteristic parameters affecting the aerodynamic performance of the aircraft further include: state and control parameters, intermediate parameters, special aerodynamic characteristic parameters, state parameters affected by pneumatic effects, and aerodynamic coefficients;

[0018] Starting from the state and control parameters, through the transmission of intermediate parameters, special aerodynamic characteristic parameters, and state parameters affected by aerodynamic effects, until the aerodynamic coefficients, a coupling and cyclic relationship among the factors affecting aerodynamic effects is formed.

[0019] In some embodiments, the state and control parameters further include: airspeed, angular velocity, barometric altitude, rotor speed, tilt angle, and control surface deflection angle.

[0020] The intermediate parameters further include: rotor force coefficient, induced velocity, and center of gravity position.

[0021] The special aerodynamic characteristic parameters further include: rotor interaction, rotor wake, rotor downwash, wing wake, wing downwash, and rotor wake - fuselage - ground effect.

[0022] The state variables affected by aerodynamic effects further include: rotor wake area, rotor wake radius, angle of attack, sideslip angle, and dynamic pressure.

[0023] The aerodynamic coefficients further include: fuselage aerodynamic force and moment, wing aerodynamic force and moment, rotor aerodynamic force and moment, and control surface aerodynamic force and moment.

[0024] In some embodiments, step S2 further includes selecting a multi - layer perceptron neural network and a long short - term memory neural network as the basic neural networks of the multi - task learning model:

[0025] The multi - layer perceptron neural network is used for modeling the aerodynamic effects of the wing, fuselage, and control surface.

[0026] The long short - term memory neural network is used for modeling the aerodynamic effects of the rotor.

[0027] In some embodiments, step S2 further includes implementing neuron connections between network layers using sparse connections and skip connections:

[0028] For the sparse connection, the proportion of the number of neurons that each neuron sparsely connects to the next layer in the total number of neurons in the next layer is used as a hyperparameter of the network, and the connection of the neuron to the input of the previous layer is disconnected.

[0029] For the skip connection, a residual connection is adopted, and the output of a certain layer is directly passed to the input of a deeper layer through matrix addition without adding any parameters.

[0030] In some embodiments, for the multi-task learning model in step S2, for different sub-tasks corresponding to each aerodynamic component of the eVTOL, according to different input situations, the weights of the contributions of different aerodynamic coefficients in each sub-task to the total aerodynamic effect are dynamically adjusted, and the losses of all aerodynamic coefficients are weighted according to the weights to obtain the total loss function.

[0031] In some embodiments, step S2 further includes:

[0032] Using the weighted average method to adjust the weights of different aerodynamic coefficients in each sub-task.

[0033] In some embodiments, step S2 further includes:

[0034] Using the Z-score method to normalize the weights of different aerodynamic coefficients in each sub-task.

[0035] In some embodiments, the multi-task learning model based on aerodynamic effect characteristics in step S2 further includes an input layer, several shared layers, task-specific layers, and an output layer:

[0036] The input layer receives the state and control variables from the aerodynamic effect analysis as inputs;

[0037] The several shared layers are used to extract aerodynamic effect characteristics and share parameters of different sub-tasks;

[0038] The task-specific layers receive the outputs of the last shared layer required for each sub-task, complete non-linear transformation through their respective activation functions, and obtain the aerodynamic coefficients corresponding to each sub-task by weighting with different weights;

[0039] The output layer is used to output the aerodynamic force and moment coefficients.

[0040] In some embodiments, the several shared layers include long short-term memory network neurons and multi-layer perceptron neurons, which are used to extract the cyclic characteristics and non-linear characteristics of aerodynamic characteristics respectively;

[0041] The several shared layers include a first shared layer, a second shared layer, and a third shared layer:

[0042] The first shared layer is sparsely connected to the input layer and is used to represent the intermediate parameters affecting aerodynamic characteristics;

[0043] The second shared layer is sparsely connected to the first shared layer, and after combining the output processed by the long short-term memory network neurons and the output of the multi-layer perceptron neurons by weighting, it is transmitted to the third shared layer;

[0044] The third shared layer is sparsely connected to the second shared layer. Combining the output of the second shared layer, after weighted combination with the outputs of the long short-term memory network neurons and the multi-layer perceptron neurons, it is transmitted to the task-specific layer.

[0045] In some embodiments, step S3 further includes:

[0046] Analyze the sample data requirements and the characteristics of the original data, and divide different sample data into continuous variables and discrete variables;

[0047] For continuous variables, perform non-uniform sampling based on aerodynamic characteristics and uniform sampling based on flight envelope constraints;

[0048] For discrete variables, use Latin hypercube sampling.

[0049] In some embodiments, the non-uniform sampling includes variable granularity sampling, and the uniform sampling includes rejection sampling.

[0050] To achieve the above object, the present invention provides an eVTOL aerodynamic modeling system based on multi-task learning, including a memory and a processor:

[0051] The memory is used to store instructions executable by the processor;

[0052] The processor is used to execute the instructions to implement the method as described above.

[0053] To achieve the above object, the present invention provides a computer storage medium, on which computer instructions are stored, wherein when the computer instructions are executed by a processor, the method as described above is executed.

[0054] The present invention proposes an eVTOL aerodynamic modeling method and system based on multi-task learning. By constructing an efficient and high-fidelity multi-task learning model (ABMTL) based on aerodynamic effect characteristics, combining the basic structures of a multi-layer perceptron (MLP) and a long short-term memory network (LSTM), and an innovative data sampling method, compared with traditional physical mechanism modeling and fluid mechanics simulation, it significantly improves the non-linear modeling ability and generalization ability, reduces the computational cost and time cost, and at the same time improves the sample data utilization rate and training efficiency, and can more accurately reflect the aerodynamic interference and coupling characteristics, realizing the rapid and accurate prediction of the complex aerodynamic characteristics of eVTOL. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The above and other features, properties, and advantages of the present invention will become more apparent from the following description in conjunction with the drawings and embodiments, where the same reference numerals in the drawings always represent the same features, wherein:

[0056] Figure 1Reveals the step diagram of the eVTOL aerodynamic modeling method based on multi-task learning according to an embodiment of the present invention;

[0057] Figure 2 Reveals the logic diagram of the aerodynamic effect generation of the eVTOL according to an embodiment of the present invention;

[0058] Figure 3 Reveals the residual block structure diagram according to an embodiment of the present invention;

[0059] Figure 4 Reveals the schematic diagram of the ABMTL neural network model structure according to an embodiment of the present invention;

[0060] Figure 5 Reveals the schematic diagram of the sample data set construction process according to an embodiment of the present invention;

[0061] Figure 6 Reveals the multi-task learning aerodynamic modeling flowchart of a certain eVTOL according to an embodiment of the present invention;

[0062] Figure 7a Reveals the curve diagram of the hover power varying with the thrust coefficient under the hover condition according to an embodiment of the present invention;

[0063] Figure 7b Reveals the curve diagram of the thrust coefficient varying with the pitch angle under the hover condition according to an embodiment of the present invention;

[0064] Figure 7c Reveals the curve diagram of the torque coefficient varying with the thrust coefficient under the hover condition according to an embodiment of the present invention;

[0065] Figure 8a Reveals the curve diagram of the thrust coefficient varying with the pitch angle at different advance ratios under the vertical takeoff and landing condition according to an embodiment of the present invention;

[0066] Figure 8b Reveals the curve diagram of the torque coefficient varying with the thrust coefficient at different rotor angles of attack under the vertical takeoff and landing condition according to an embodiment of the present invention;

[0067] Figure 9a Reveals the curve diagram of the lift coefficient varying with the angle of attack in different flap / aileron modes under the vertical takeoff and landing condition according to an embodiment of the present invention;

[0068] Figure 9b Reveals the curve diagram of the drag coefficient varying with the angle of attack in different flap / aileron modes under the vertical takeoff and landing condition according to an embodiment of the present invention;

[0069] Figure 9c Reveals the curve diagram of the pitching moment coefficient varying with the angle of attack in different flap / aileron modes under the vertical takeoff and landing condition according to an embodiment of the present invention;

[0070] Figure 10a Reveals the curve of the lateral force coefficient varying with the yaw angle under different flap / aileron modes in the vertical takeoff and landing condition according to an embodiment of the present invention;

[0071] Figure 10b Reveals the curve of the rolling moment coefficient varying with the yaw angle under different flap / aileron modes in the vertical takeoff and landing condition according to an embodiment of the present invention;

[0072] Figure 10c Reveals the curve of the yaw moment coefficient varying with the angle of attack under different flap / aileron modes in the vertical takeoff and landing condition according to an embodiment of the present invention;

[0073] Figure 11a Reveals the curve of the wing lift coefficient varying with the angle of attack in the vertical takeoff and landing condition according to an embodiment of the present invention;

[0074] Figure 11b Reveals the curve of the wing drag coefficient varying with the angle of attack in the vertical takeoff and landing condition according to an embodiment of the present invention;

[0075] Figure 12a Reveals the curve of the horizontal tail lift coefficient varying with the angle of attack in the vertical takeoff and landing condition according to an embodiment of the present invention;

[0076] Figure 12b Reveals the curve of the horizontal tail drag coefficient varying with the angle of attack in the vertical takeoff and landing condition according to an embodiment of the present invention;

[0077] Figure 13a Reveals the curve of the vertical tail lateral force coefficient varying with the sideslip angle in the vertical takeoff and landing condition according to an embodiment of the present invention;

[0078] Figure 13b Reveals the curve of the vertical tail drag coefficient varying with the angle of attack in the vertical takeoff and landing condition according to an embodiment of the present invention;

[0079] Figure 14a Reveals the curve of the lift coefficient varying with the angle of attack under different flap / aileron modes in the transition condition according to an embodiment of the present invention;

[0080] Figure 14b Reveals the curve of the drag coefficient varying with the angle of attack under different flap / aileron modes in the transition condition according to an embodiment of the present invention;

[0081] Figure 14c Reveals the diagram of the pitching moment coefficient varying with the angle of attack under different flap / aileron modes in the transition condition according to an embodiment of the present invention;

[0082] Figure 15a Reveals the curve of the wing lift coefficient varying with the angle of attack in the transition condition according to an embodiment of the present invention;

[0083] Figure 15b Reveals the curve of the wing lift coefficient varying with the angle of attack under transitional conditions according to an embodiment of the present invention. Detailed implementation manners

[0084] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the invention and are not used to limit the invention.

[0085] Aiming at the problems in the prior art that in the eVTOL (electric vertical takeoff and landing vehicle) configuration, the aerodynamic components such as rotors, fuselages, wings, and various control surfaces are coupled and interfere with each other, and the traditional aerodynamic modeling methods are difficult to accurately simulate the unsteady aerodynamic characteristics of eVTOL, with high time costs and requirements for computing resources, the present invention proposes an eVTOL aerodynamic modeling method based on multi-task learning. By introducing a multi-task learning aerodynamic modeling framework based on aerodynamic characteristics, it can effectively improve the prediction accuracy of the complex aerodynamic characteristics of eVTOL, while improving the utilization rate of sample data and the training speed, and reducing the model training cost.

[0086] Figure 1 Reveals the step diagram of the eVTO aerodynamic modeling method based on multi-task learning according to an embodiment of the present invention, as Figure 1 shown, an eVTOL aerodynamic modeling method based on multi-task learning proposed by the present invention includes the following steps:

[0087] Step S1, analyze the aerodynamic effects and mutual influences of each aerodynamic component of the eVTOL, identify the key characteristic parameters affecting the aerodynamic performance of the aircraft, and obtain the aerodynamic effect analysis result;

[0088] Step S2, use the aerodynamic effect analysis result, combine a multi-layer perceptron neural network and a long short-term memory neural network to construct a multi-task learning model based on aerodynamic effect characteristics;

[0089] Step S3, obtain the original data of the aircraft, and construct a sample data set according to the flight state and aerodynamic characteristics;

[0090] Step S4, use the sample data set in Step S3 to train the multi-task learning model in Step S2, and perform verification and evaluation to obtain and output the multi-task learning model of the eVTOL.

[0091] An eVTOL aerodynamic modeling method proposed by the present invention realizes the rapid and accurate prediction of the aerodynamic coefficients of multiple aerodynamic components of the eVTOL by constructing an efficient and high-fidelity multi-task learning model (Aero-based Multi-task Learning, ABMTL) based on the aerodynamic effect characteristics of the eVTOL.

[0092] More specifically, first, analyze the aerodynamic effect characteristics of each component of the eVTOL, and respectively select the multi-layer perceptron (MLP) applicable to non-linearity and the long short-term memory network (LSTM) for cyclic characteristic learning as the basic structure of multi-task learning; second, combine the incomplete connection effect and cross-layer influence relationship between the input variables and output variables, select the neuron connection methods of skip connection and cross-layer connection, and construct a network structure by combining the loss function of aerodynamic effect characteristics, and establish a multi-task learning model structure based on the characteristics of aerodynamic effects; third, according to the formal characteristics and change characteristics of multi-source raw data, adopt innovative data sampling methods (including non-uniform sampling method based on aerodynamic characteristics, uniform sampling method based on flight envelope constraints, and discrete state parameter sampling method based on Latin hypercube method) to quickly construct a high-quality sample data set, and finally, achieve the effects of improving the prediction accuracy of the complex aerodynamic characteristics of eVTOL, increasing the utilization rate of sample data, accelerating the training speed, and reducing the model training cost.

[0093] These steps will be described in detail below. It should be understood that within the scope of the present invention, the above technical features of the present invention and the technical features specifically described below (such as in the embodiments) can be combined with each other and are interrelated to form a preferred technical solution.

[0094] Step S1: Analyze the aerodynamic effects and mutual influences between the aerodynamic components of the eVTOL, identify the key characteristic parameters affecting the aerodynamic performance of the aircraft, and obtain the aerodynamic effect analysis results.

[0095] In this embodiment, the aerodynamic effects and mutual influences between the aerodynamic components of the eVTOL further include:

[0096] The aerodynamic effects of isolated rotors, the mutual influences between rotors, and the aerodynamic effects of the same components as those of fixed-wing aircraft.

[0097] Furthermore, the analysis process and results of the aerodynamic effects of isolated rotors are as follows:

[0098] For eVTOLs of various structures, there are usually aerodynamic effects generated by isolated rotors. When the rotational motion of the rotors drives the air to flow through the rotors, an induced velocity opposite to the direction of the lift force will be generated.

[0099] The induced velocity, together with control variables such as the rotor speed and the rotor pitch angle, affects the flow state of the rotor, that is, the inflow ratio and the advance ratio of the rotor, and further affects the lift coefficient and torque coefficient of the rotor. And various studies on rotors have shown that the induced velocity is also a function of the lift force, the forward flight speed, and the inflow velocity, resulting in an implicit cycle of the lift force and the induced velocity.

[0100] Flapping of the rotor blades refers to a phenomenon caused by the different flow conditions at the blade due to the forward and backward movement of the rotor relative to the oncoming flow. The flapping of the rotor blades will change the magnitude of the thrust component on the rotor disk, thereby affecting the inflow dynamics of the rotor.

[0101] Furthermore, the analysis process and results of the mutual influence between rotors are as follows:

[0102] Based on the aerodynamic effects of isolated rotors, the mutual influence between rotors is manifested in that state variables such as the induced velocity and effective angle of attack of the rotor are affected by the interaction of the wake and blade vortices generated by other rotors. At the same time, the rotor is also affected by the wake and aerodynamic blockage effects of the wing, nacelle, and fuselage.

[0103] Furthermore, the analysis process and results of the aerodynamic effects of the same components as those of fixed-wing aircraft are as follows:

[0104] For the parts of the eVTOL structure that are the same as those of fixed-wing aircraft, such as the wing, fuselage, and control surfaces, it can be learned from the research on the development of eVTOL aircraft models by analogy with computational fluid dynamics (CFD) and wind tunnel tests that the aerodynamic force and moment coefficients of these components can be expressed as functions of state variables and some control variables.

[0105] Different from the aerodynamic effects of fixed-wing aircraft, there will be strong aerodynamic interference effects between these aerodynamic components and the rotor, especially during the vertical takeoff and transition phases of low-speed flight.

[0106] Taking the wing, which has the most complex aerodynamic effects among all components, as an example, the rotor will generate rotor wake to impact a part of the wing area during flight.

[0107] Therefore, in the relevant research on tilt-rotor aircraft, the aerodynamic model of the wing module is usually divided into the slipstream region aerodynamic model caused by the rotor wake and the free-stream region aerodynamic model not affected by the rotor wake.

[0108] The aerodynamic effects of the fuselage, horizontal tail, and other control surfaces affected by the rotor are similar to those of the wing, mainly reflected by the induced velocities acting on each component generated by the rotor wake. These induced velocities change the overall airflow velocity of each component while affecting the magnitude of the airflow angle and dynamic pressure, thereby affecting the aerodynamic force and moment coefficients.

[0109] In addition, as mentioned above, the wing, fuselage, and control surfaces will also generate wake and affect the aerodynamic effects of the rotor.

[0110] By analyzing the generation process of the aerodynamic effects of the above components, the key characteristic parameters affecting the aerodynamic performance of the aircraft and the relationships between them can be obtained.

[0111] In this embodiment, the key characteristic parameters affecting the aerodynamic performance of the aircraft further include state and control parameters, intermediate parameters, special aerodynamic characteristic parameters, state parameters affected by aerodynamic effects, and aerodynamic coefficients.

[0112] Figure 2 Discloses a logic diagram of the generation of aerodynamic effects of an eVTOL according to an embodiment of the present invention, as Figure 2 shown, the state and control parameters further include: airspeed, angular velocity, barometric altitude, rotor speed, tilt angle, control surface deflection angle;

[0113] The intermediate parameters further include: rotor force coefficient, induced velocity, center of gravity position;

[0114] The special aerodynamic characteristic parameters further include: rotor interaction, rotor wake, rotor downwash, wing wake, wing downwash, rotor wake - fuselage - ground effect;

[0115] The state parameters affected by aerodynamic effects further include: rotor wake area, rotor wake radius, angle of attack, sideslip angle, dynamic pressure;

[0116] The aerodynamic coefficients are the aerodynamic forces and moments of different aerodynamic components, and further include: fuselage aerodynamic forces and moments, wing aerodynamic forces and moments, rotor aerodynamic forces and moments, control surface aerodynamic forces and moments.

[0117] The interactions and dependencies between these parameters constitute a complex aerodynamic model of the eVTOL.

[0118] Starting from the state and control parameters of the eVTOL, through complex non - linear relationships, the influence of the parameters is transmitted to the intermediate parameters, special aerodynamic characteristic parameters, state parameters affected by aerodynamic effects, until the aerodynamic coefficients of the eVTOL, including the aerodynamic force and moment coefficients.

[0119] From Figure 2 it can be seen that for the influencing factors in each layer, such as variables like the airspeed and rotor speed of the eVTOL, not all variables in the next layer are affected. Certain input parameters, such as the nacelle tilt angle and rudder surface angle, can affect aerodynamic effect parameters such as the rotor wake area across layers, while state variables such as airspeed and airflow angle can directly act on the aerodynamic coefficients. At the same time, the inflow dynamics of the rotor, that is, the cycle of induced velocity and thrust coefficient, as well as the interactions between rotors and between rotors and wings, result in complex coupling and cyclic relationships between the influencing factors of aerodynamic effects.

[0120] The aerodynamic models of each component of the eVTOL are exactly non - linear relationship networks intertwined by these incomplete associations and cross - layer associations between parameters.

[0121] Step S2: Using the results of aerodynamic effect analysis, a multi-task learning model is constructed by combining a multi-layer perceptron neural network and a long short-term memory neural network.

[0122] In this embodiment, by analyzing the aerodynamic characteristics of eVTOL multi-modules and the relationship between influencing factors, a multi-task learning method is selected. This method can effectively utilize the internal relationship between tasks and has a high model training efficiency, and is used for data-driven aerodynamic modeling of eVTOL.

[0123] On this basis, considering the incomplete connection effect and cross-layer influence relationship between input and output variables, a multi-task learning model framework based on the characteristics of aerodynamic characteristics is constructed. The input of the model is part of the state variables and control quantities of eVTOL, and the output is the aerodynamic force and moment coefficients. The model constructed based on the characteristics of aerodynamic characteristics and the relationship between influencing factors adopts a jumping connection and sparse connection ABMTL neural network, which can more accurately fit the relationship between non-linear aerodynamic characteristics and influencing factors and predict aerodynamic coefficients.

[0124] More specifically, step S2 further includes: Selecting MLP (multi-layer perceptron neural network) and LSTM (long short-term memory neural network) as the basic neural networks of the multi-task learning model:

[0125] The MLP is used for aerodynamic effect modeling of control surfaces such as wings, fuselages, horizontal tails, and vertical tails. There are complex non-linear relationships between the aerodynamic effects of these components;

[0126] The LSTM is used for aerodynamic effect modeling of rotors. The aerodynamic effects of rotors have temporal and cyclic characteristics, and LSTM can effectively handle such non-linear temporal coupling problems.

[0127] As the basic neural network types in the multi-task learning model, MLP and LSTM are used to construct neurons in the shared layer and task-specific layers.

[0128] Step S2 further includes: Implementing neuron connections between layers using sparse connections and jumping connections.

[0129] Based on the analysis of the characteristics of eVTOL aerodynamic effects, the non-linear relationship network of the eVTOL aerodynamic model has two characteristics, namely, there are incomplete associations and cross-layer associations between parameters in the generation process of aerodynamic effects. In this embodiment, sparse connections and jumping connections are respectively used to achieve incomplete associations and cross-layer associations between parameters.

[0130] Taking the wing wake of eVTOL aerodynamic effects as an example, the wing wake usually affects state variables such as the airspeed and airflow angle of the fuselage, horizontal tail, and vertical tail, but does not include all state variables. For such aerodynamic characteristics, sparse connections are used for expression.

[0131] For the sparse connection, the proportion of the number of neurons in each neuron's sparse connection to the next layer to the number of neurons in the next layer is used as a hyperparameter of the network, and based on this, the connection of the neuron to the input of the previous layer is disconnected, that is, the weight of this neuron relative to the input is set to zero to represent the incomplete correlation between aerodynamic effect parameters;

[0132] By restricting the number of connections between neurons, the sparse connection avoids the influence of irrelevant factors on intermediate variables, makes the network structure closer to the real aerodynamic model, and at the same time reduces the number of parameters to be calculated, thereby reducing the computational cost and memory occupancy.

[0133] Taking the nacelle tilt angle and control surface angle of eVTOL aerodynamic effects as an example, the nacelle tilt angle and control surface angle can affect aerodynamic characteristics such as rotor wake across layers, and state variables such as airspeed and airflow angle can directly affect aerodynamic coefficients. Such characteristics of aerodynamic effects are preferably represented by skip connections, that is, matrix addition is used as the information transmission method.

[0134] For the skip connection, a residual connection is adopted, and the output of a certain layer is directly transmitted to the input of a deeper layer through matrix addition without adding any parameters.

[0135] Furthermore, in this embodiment, for the residual connection relationship, the number of neurons with residual connection relationships in each layer and the number of neurons receiving information in the next two layers are used as hyperparameters, and the neuron selection method uses a random method.

[0136] The residual connection directly transmits the output of a certain layer to the input of a deeper layer through matrix addition by constructing a residual block without adding any parameters. Figure 3 Reveals the structural diagram of a residual block according to an embodiment of the present invention, as Figure 3 shown, a single residual block includes two weight layers, and its information transmission method is as follows:

[0137]

[0138] Among them, x is the input of the residual block, F(x) is the output of the residual block, is the residual function, W1,

[0139] W2 are weights, b1, b2 are bias terms, and σ is the activation function between layers;

[0140] In this embodiment, the ReLU function is used as the activation function.

[0141] From Figure 3 It can be seen that the output F(x) of the residual block not only contains the information of the original input, but also incorporates the input information after being processed by the residual function transformation. This is very similar to the characteristic in aerodynamic characteristics where influencing factors affect other parameters through intermediate variables. In addition, the skip connection enables information to flow freely in the network, which helps to more completely transmit the input information to the deeper layers of the network, such as the input parameters directly affecting the aerodynamic coefficients, thereby improving the model's expressive ability and enhancing the model's prediction accuracy.

[0142] The existing multi-task learning methods have limitations in eVTOL aerodynamic modeling, mainly reflected in the inability to dynamically reflect the change in the contribution of the aerodynamic coefficients of each sub-task to the total aerodynamic effect under different input conditions. To address this problem, the present invention proposes a loss function combining aerodynamic effect characteristics. For different sub-tasks corresponding to each aerodynamic component of the eVTOL, it is necessary to dynamically adjust the weights of different aerodynamic coefficients in each sub-task for their contributions to the total aerodynamic effect according to different input situations. Finally, the losses of all aerodynamic coefficients are weighted according to the weights to obtain the total loss function.

[0143] Due to the complex non-linear relationship between the input state, control variables, and output aerodynamic coefficients, when the input samples change, the weights of the outputs of each sub-task in the total task will change accordingly. Therefore, the present invention uses a data-driven method to determine the weights of the outputs in the MTL (multi-task learning) neural network. The specific steps for the weights of the aerodynamic coefficients of all sub-tasks are as follows:

[0144] Define a function for calculating the weights of the aerodynamic coefficients of all sub-tasks, and this weight represents the relative contribution of a certain aerodynamic coefficient of each aerodynamic component.

[0145] For the j-th aerodynamic coefficient output C in the i-th sub-task of multi-task learning i,j , then its corresponding weight output α i,j is calculated by the formula (3).

[0146]

[0147] When constructing the ABMTL network structure of the present invention, using this weight as the output and the training sample data as the input, a corresponding non-linear mapping function is constructed and embedded in the forward propagation process.

[0148] The step S2 further includes:

[0149] Adopt the weighted average method to adjust the weights of different aerodynamic coefficients of each sub-task.

[0150] To ensure that when dynamically adjusting the weights of aerodynamic coefficients, the stability and training efficiency of the model are not affected due to excessive weight fluctuations, in this embodiment, the weighted average method is used to smoothly adjust the weights to help the weights maintain stability during the dynamic change process. The corresponding expression of the weighted average method is as follows:

[0151]

[0152] Among them, is the weight obtained from the previous n calculations in the t-th cycle of training, and w n is the corresponding weight factor.

[0153] In this embodiment, the results of the recent four weight calculations are selected and saved, and the values of the weight factors are set to 0.4, 0.3, 0.2, and 0.1 respectively, so that the final result depends more on the recent weights.

[0154] The step S2 further includes:

[0155] The Z-score method is used to normalize the weights of different aerodynamic coefficients for each subtask.

[0156] After determining the weights by the above method, the weights need to be normalized for comparison on the same scale; in this embodiment, the Z-score method is used to normalize the weights, and the corresponding expression is as follows:

[0157]

[0158] Among them, μ is the mean of the weights, σ is the standard deviation of the weights, and α i,j is the weight of the j-th aerodynamic coefficient in the i-th subtask of multi-task learning, is the normalized weight.

[0159] In the loss function of the ABMTL model, these dynamically adjusted weights are used to adjust the loss of each aerodynamic coefficient and calculate the total loss. The corresponding expression is as follows:

[0160]

[0161] Among them, L i,j is the loss of the j-th aerodynamic coefficient in the i-th subtask, is the dynamically adjusted weight. This loss function replaces the MSE loss function in traditional multi-task learning.

[0162] The calculation time of the loss function in the training process of the ABMTL network proposed by the present invention based on the improvement of eVTOL aerodynamic characteristics is slightly longer. However, compared with the traditional MTL neural network directly applying the MSE loss function, the ABMTL model has higher accuracy. Although the ABMTL structure becomes more complex due to the characteristics of combining eVTOL aerodynamic effects, compared with the method of training multiple single-task neural networks separately, ABMTL occupies less total training time and total storage space.

[0163] In addition, in view of the coupling effect of mutual influence between eVTOL aerodynamic components, when training a single-task learning network model, there will be a situation of reusing the same sample data, while the ABMTL model can effectively reduce this information reuse phenomenon.

[0164] Furthermore, the multi-task learning model based on aerodynamic effect features in step S2 further includes an input layer, several shared layers, task-specific layers, and an output layer:

[0165] The input layer receives the state and control variables from the aerodynamic effect analysis as inputs;

[0166] The several shared layers are used to extract aerodynamic effect features and share the parameters of different subtasks;

[0167] The task-specific layers receive the outputs of the last shared layer required by each subtask, complete non-linear transformation through their respective activation functions, and obtain the aerodynamic coefficients corresponding to each subtask by weighting with different weights;

[0168] The output layer is used to output aerodynamic force and moment coefficients.

[0169] Figure 4 Discloses a schematic diagram of the ABMTL neural network model structure according to an embodiment of the present invention, as Figure 4 The specific structure of the ABMTL model shown includes an input layer, a first shared layer, a second shared layer, a third shared layer, a task-specific layer, and an output layer:

[0170] The input layer receives the state and control variables from the aerodynamic effect analysis as inputs;

[0171] The first shared layer, the second shared layer, the third shared layer, and the task-specific layer are hidden layers;

[0172] The first shared layer is sparsely connected to the input layer and is used to represent intermediate parameters affecting aerodynamic characteristics;

[0173] The second shared layer and the third shared layer both contain LSTM neurons and MLP neurons and are used to extract cyclic characteristics and non-linear characteristics;

[0174] The second shared layer is sparsely connected to the first shared layer. After combining the output processed by the long short-term memory network neurons and the output of the MLP neurons through weighting, it is transmitted to the third shared layer;

[0175] The third shared layer is sparsely connected to the second shared layer. Combining the output of the second shared layer, and after weighting and combining with the outputs of the long short-term memory network neurons and the multi-layer perceptron neurons, it is transmitted to the task-specific layer.

[0176] During the forward propagation process, the corresponding output data of the second shared layer extracts the output of the last time step after being processed by the LSTM neurons, and is weighted and output to the next layer, the third shared layer, together with the MLP neurons of this layer. These shared layers are jointly used by each sub-task, so their parameters need to be trained and verified using the sample data of all sub-tasks.

[0177] In addition, the connection method between each shared layer and the next layer is a sparse connection, and there are skip connections between the neurons of each layer and the next two layers. These two connection methods are used to reflect the characteristics that the influencing factors do not fully act on the aerodynamic characteristics and have cross-layer effects.

[0178] The neurons of different colors in the task-specific layer represent the fully connected layers specific to each sub-task. The task-specific layer receives the output of the last shared layer required by each sub-task, completes the non-linear transformation through its respective activation function, and obtains the aerodynamic coefficients corresponding to each sub-task by weighting with different weights.

[0179] The output layer is used to output the aerodynamic force and moment coefficients.

[0180] The present invention constructs a multi-task learning model (ABMTL) based on aerodynamic effect characteristics by combining MLP and LSTM neural networks and adopting sparse connection and skip connection methods. This model can dynamically adjust the weight of the loss function, accurately reflect the aerodynamic effects of each aerodynamic component of the eVTOL, and significantly improve the prediction accuracy and stability of the model.

[0181] Step S3: Obtain the original data of the aircraft, and construct a sample data set according to the flight state and aerodynamic characteristics;

[0182] To ensure the quality and construction efficiency of the sample data set, according to the sample data requirements and data characteristics of the ABMTL network model, it is first necessary to clarify the constraint relationship between each input and output parameter, and use appropriate methods to divide and expand the sample data. Through the analysis of the sample data requirements and in-depth study of the characteristics of the original data, different sample data are reasonably divided, and the following sampling methods are respectively adopted according to their characteristics: non-uniform sampling based on aerodynamic characteristics, uniform sampling based on flight envelope constraints, and discrete state parameter sampling based on the Latin hypercube method.

[0183] Figure 5 Reveals a schematic diagram of the sample dataset construction process according to an embodiment of the present invention, as Figure 5 shown. Based on the analysis of the requirements and data characteristics of the sample data, the data can be divided into continuous variables and discrete variables. For continuous variables, considering their non-linear relationship, non-uniform sampling can be performed, such as variable-granularity sampling; when based on the flight envelope constraint, uniform sampling can be used, such as rejection sampling. For discrete variables, Latin hypercube sampling can be used.

[0184] More specifically, the non-uniform sampling can employ variable-granularity sampling;

[0185] For example, both the lift coefficient and the pitching moment coefficient are approximately linearly varying within a certain angle of attack range, while the non-linear characteristics of the drag coefficient are more significant in the middle section of the interval.

[0186] To ensure the rationality of the selection of sample data, after comprehensively considering the resolution of the original data and the value range of the input variables, the union processing is performed on the non-linear regions, and then the part with stronger non-linearity is selected. The sampling granularities of the input variable intervals with non-linear and linear variations in the mapping relationship are set to 2° and 1° respectively.

[0187] More specifically, the uniform sampling can adopt the rejection sampling method;

[0188] For the two-dimensional plane constraint, the rejection sampling method is used. Specifically, a rectangle containing the entire irregular region is selected as the proposal distribution, and the process of randomly sampling and checking the constraint conditions is repeated continuously to screen out the points falling within the flight envelope, thereby obtaining a certain number of samples.

[0189] More specifically, the Latin hypercube sampling further includes:

[0190] The intervals in each dimension are equally divided, and then a point is randomly selected within each sub-interval to ensure that the samples in each dimension do not concentrate in a certain interval, reducing the aggregation phenomenon of the sample points and ensuring the uniformity of the sample point distribution in the multi-dimensional space.

[0191] Step S3 of the present invention can significantly improve the construction efficiency of the sample dataset while retaining the distribution characteristics of the original data. Finally, the ABMTL model has relatively high prediction accuracy and relatively low data requirements, verifying the effectiveness of the above sampling method.

[0192] Step S4, use the sample dataset of step S3 to train the multi-task learning model of step S2, and perform verification and evaluation, and finally output the multi-task learning model for eVTOL.

[0193] The training of a neural network model generally involves determining a set of hyperparameter values, and using a given dataset and loss function to adjust the weights and biases of the neural network through the backpropagation algorithm to minimize the loss function.

[0194] Hyperparameters not only determine the scale of the model, but also relate to the calculation methods from input parameters to output parameters. Given a certain sample dataset, they will have an important impact on the training efficiency and quality of the model, as well as the model accuracy.

[0195] Therefore, a scientific method is needed to determine the appropriate hyperparameter values.

[0196] In this embodiment, the hyperparameters include the number of inputs and outputs, the number of training epochs, the batch size, the number of neurons, the number of hidden layers, the memory sequence length, the comprehensive error, etc. These parameters determine the scale of the model, the calculation method, as well as the training efficiency and quality.

[0197] In addition, the loss function of traditional deep learning methods usually uses the mean squared error function. In actual aerodynamic modeling, the weights of the output of each aerodynamic coefficient and aerodynamic component in the total task will change continuously. The present invention adopts a loss function combined with aerodynamic effect characteristics, and dynamically adjusts the weights of the aerodynamic coefficients of each subtask through a data-driven method according to the input-output relationship of the samples, so that the loss function can accurately reflect the contribution of each aerodynamic coefficient to the total task.

[0198] The generalization ability of the model is evaluated through the cross-validation method to avoid overfitting. It can be evaluated by comparing with single-task MLP and LSTM neural network aerodynamic models, and the accuracy, stability and computational efficiency of the model in predicting eVTOL aerodynamic coefficients can be evaluated.

[0199] Through in-depth analysis of eVTOL aerodynamic effect characteristics, careful design of the neural network structure, reasonable generation of the training dataset, and a complete training and verification process, through the introduction of a multi-task learning model and an innovative design combining aerodynamic effect characteristics, the complex coupling relationship between various components of eVTOL can be effectively processed, thereby improving the prediction accuracy of aerodynamic performance and optimizing the aircraft design and control strategy.

[0200] Figure 6 Discloses a multi-task learning aerodynamic modeling flowchart of a certain eVTOL according to an embodiment of the present invention, as Figure 6 shown. Taking the aerodynamic modeling of a certain configuration of eVTOL as an example, the eVTOL aerodynamic modeling method based on multi-task learning proposed by the present invention is described. Steps S1 to S4 correspond to four stages: analysis of aerodynamic effect characteristics, construction of the neural network model structure, generation and division of the dataset, and training and verification of the model.

[0201] Step S1, Analysis of Aerodynamic Effect Characteristics

[0202] Taking the tilt-rotor configuration eVTOL as an example, based on the aerodynamic modeling method and physical mechanism, sort out the aerodynamic characteristics of each module of the tilt-rotor aircraft and the complex coupling relationship between parameters.

[0203] The configuration characteristics of the tilt-rotor aircraft determine that it exhibits extremely complex aerodynamic characteristics during both the vertical takeoff phase and the transition phase.

[0204] During the operation of the rotor, once the velocity and direction of the inflow change, the blade flapping dynamics will also change accordingly, and these two will act together to affect the thrust and induced velocity generated by the rotor.

[0205] At the same time, the rotor will generate aerodynamic interference on other rotors and the airframe, including but not limited to the rotor wake acting on the wing, horizontal tail, vertical tail and other parts. The complex coupling relationship existing between these aerodynamic characteristics is the main reason for the difficulty of traditional physical models to achieve high fidelity.

[0206] By analyzing the generation mechanism and process of aerodynamic characteristics in Step S1 and obtaining the analysis results, it provides a theoretical basis for subsequent model selection, construction and data requirements.

[0207] Step S2, Construction of Neural Network Model Structure

[0208] Based on the aerodynamic characteristic mechanism analyzed in Step S1, select a suitable data-driven method to construct the neural network model structure.

[0209] In this embodiment, a multi-task learning model is adopted in Step S2.

[0210] For eVTOLs with various complex configurations, aerodynamic modeling involves multiple related tasks, that is, the separate modeling of aerodynamic components such as eVTOL rotors, fuselages, wings, etc. This makes multi-task learning more suitable for eVTOL aerodynamic modeling work compared with single-task learning. In addition, compared with training separate models for each task, multi-task learning has the advantages of low model training cost, high data utilization rate and fast training speed.

[0211] On this basis, the multi-task learning model adopted in Step S2, through the analysis of aerodynamic characteristics, introduces sparse connections and skip connections, and constructs a neural network model that conforms to the eVTOL aerodynamic effect characteristics on the basis of the classical neural network structure.

[0212] Step S3, Generation and Division of Data Set

[0213] The sample dataset is mainly used for the optimal selection, training, and validation of the hyperparameters of the network model. Its quantity and quality play a decisive role in the accuracy of the network model.

[0214] To ensure the quality and construction efficiency of the sample dataset, the sample dataset is constructed and divided according to the characteristics of the original data and the specific requirements of the network model. Usually, the dataset is divided into a training set, a validation set, and a test set in a ratio of 7:3:1.

[0215] Step S4, training and validation of the model

[0216] To verify the prediction accuracy of the proposed ABMTL model of the present invention, a high-precision dataset is constructed using the data of the XV-15 tiltrotor configuration aircraft publicly available from NASA, and single-task MLP and LSTM neural network aerodynamic models of each aerodynamic component are established corresponding to the characteristics of the fuselage, wing, control surface, and rotor. The mean square error function MSE is used as the evaluation index for model training, and the network hyperparameter combination with the minimum comprehensive error is determined, as shown in Table 1 specifically.

[0217] Table 1 Network hyperparameter combination of the neural network model

[0218]

[0219] It can be seen from Table 1 that the ABMTL model with the weight adaptive loss function constructed by the present invention, although its structure is relatively complex, its comprehensive error is reduced compared with the MTL model using the traditional MSE loss function. In addition, the error of this model is significantly lower than the error of the single-task network models of each aerodynamic component, and can improve the accuracy of the model to a certain extent.

[0220] On this basis, the same test samples of unsteady aerodynamic characteristics are selected for comparative analysis of different models.

[0221] Figures 7a - 15b They are the prediction results of the aerodynamic coefficients of each aerodynamic component under the hover condition, the vertical takeoff and landing condition, and the two transition flight conditions respectively. The Wind tunnel data, LSTM / MLP network prediction, and MTL network prediction results in the figure refer to the original dataset of the wind tunnel test, the single-task LSTM / MLP models of each aerodynamic component, and the prediction results of the ABMTL model of the present invention respectively.

[0222] For a tiltrotor, in the hover condition, the thrust coefficient of the rotor increases nearly linearly with the increase of the rotor pitch angle. The torque coefficient of the rotor has a non-linear relationship with the thrust coefficient, and the hover power of the rotor fluctuates up and down after increasing continuously with the thrust to a certain level. The aerodynamic characteristics of the rotor in the vertical takeoff and transition conditions are similar, but at this time, the lift of the tiltrotor mainly comes from the rotor, and the aerodynamic coefficients of the rotor only include the thrust coefficient and the torque coefficient of the rotor.

[0223] Figures 7a through 7c is the change curve of the aerodynamic coefficients of the rotor in the hover condition. Specifically, Figure 7a is the curve of the hover power varying with the thrust coefficient in the hover condition, Figure 7b is the curve of the thrust coefficient varying with the pitch angle in the hover condition, Figure 7c is the curve of the torque coefficient varying with the thrust coefficient in the hover condition; Figures 8a through 8b is the change curve of the aerodynamic coefficients of the rotor in the vertical takeoff and landing condition. Specifically, Figure 8a is the curve of the thrust coefficient varying with the pitch angle under different advance ratios in the vertical takeoff and landing condition, Figure 8b is the curve of the torque coefficient varying with the thrust coefficient under different rotor angles of attack in the vertical takeoff and landing condition.

[0224] As Figures 7a - 8b shown, the aerodynamic coefficients of the rotor are predicted by the LSTM neural network and the ABMTL neural network respectively, and the prediction results of both neural networks can better reflect the change trend of the data.

[0225] In the vertical takeoff and transition stages, with the changes of the flow angle and state variables, the fuselage module also provides a part of the aerodynamic force and moment. Figures 9a through 15b The curves in

[0226] Figures 9a through 9c represent the original data sets of the wind tunnel test, the prediction results of the MLP neural network and the multi-task learning neural network in the vertical takeoff and landing condition and the transition condition respectively. Figure 9a is the change curve of the longitudinal aerodynamic coefficients of the fuselage in the vertical takeoff and landing condition. Specifically, Figure 9b is the curve of the lift coefficient varying with the angle of attack under different flap / aileron modes in the vertical takeoff and landing condition, Figure 9c is the curve of the drag coefficient varying with the angle of attack under different flap / aileron modes in the vertical takeoff and landing condition,

[0227] Figures 10a through 10c is the change curve of the lateral aerodynamic coefficients of the fuselage in the vertical takeoff and landing condition. Specifically, Figure 10a is the curve of the side force coefficient varying with the yaw angle under different flap / aileron modes in the vertical takeoff and landing condition, Figure 10bIt is a curve graph of the roll moment coefficient varying with the yaw angle under different flap / aileron modes in the vertical takeoff and landing condition. Figure 10c It is a curve graph of the yaw moment coefficient varying with the angle of attack under different flap / aileron modes in the vertical takeoff and landing condition.

[0228] Figures 11a through 11b It is the curve of the wing aerodynamic coefficient in the vertical takeoff and landing condition. Specifically, Figure 11a It is a curve graph of the wing lift coefficient varying with the angle of attack in the vertical takeoff and landing condition. Figure 11b It is a curve graph of the wing drag coefficient varying with the angle of attack in the vertical takeoff and landing condition.

[0229] Figures 12a through 12b It is the curve of the horizontal tail aerodynamic coefficient in the vertical takeoff and landing condition. Specifically, Figure 12a It is a curve graph of the horizontal tail lift coefficient varying with the angle of attack in the vertical takeoff and landing condition. Figure 12b It is a curve graph of the horizontal tail drag coefficient varying with the angle of attack in the vertical takeoff and landing condition.

[0230] Figures 13a through 13b It is the curve of the vertical tail aerodynamic coefficient in the vertical takeoff and landing condition. Specifically, Figure 13a It is a curve graph of the vertical tail side force coefficient varying with the sideslip angle in the vertical takeoff and landing condition. Figure 13b It is a curve graph of the vertical tail drag coefficient varying with the angle of attack in the vertical takeoff and landing condition.

[0231] Figures 14a through 14c It is the curve of the fuselage longitudinal aerodynamic coefficient in the transition condition. Specifically, Figure 14a It is a curve graph of the lift coefficient varying with the angle of attack under different flap / aileron modes in the transition condition. Figure 14b It is a curve graph of the drag coefficient varying with the angle of attack under different flap / aileron modes in the transition condition. Figure 14c It is a graph of the pitch moment coefficient varying with the angle of attack under different flap / aileron modes in the transition condition.

[0232] Figures 15a through 15b It is the curve of the wing aerodynamic coefficient in the transition condition. Specifically, Figure 15a It is a curve graph of the wing lift coefficient varying with the angle of attack in the transition condition. Figure 15b It is a curve graph of the wing lift coefficient varying with the angle of attack in the transition condition.

[0233] Table 2 shows the maximum relative error and average relative error of the two networks for each aerodynamic coefficient. As can be seen from Table 2, the maximum relative error of the single-task learning network model with the optimal hyperparameter combination cannot fully meet the accuracy requirement of 5% of the simulation model, while the ABMTL network model of the present invention can more accurately predict the non-linear relationship between the influencing factors and the aerodynamic coefficients. Its maximum relative error is controlled within 5%, and the average relative error accuracy is maintained at about 1%. In the simulation or actual system of eVTOL, these errors usually accumulate over time or with the increase in the number of iterations. Therefore, even if the difference in the predicted relative errors is very small, after multiple iterations or a long time of simulation, these errors will still accumulate to a level sufficient to cause a significant change in the generated aerodynamic force, resulting in an overall decrease in the flight simulation fidelity.

[0234] Table 2 Maximum relative error / average relative error of aerodynamic coefficients

[0235]

[0236] On this basis, various indicators of the two modeling schemes other than accuracy are comprehensively considered, namely the training time of the network, the storage space of the network model (saved in the.pth format), and the number of training and validation sample data points used to achieve similar prediction accuracy. Table 3 shows that MLP1, MLP2, MLP3, and LSTM respectively represent the single-task learning network models of the fuselage, wing, control rudder surface, and rotor modules, while the ABMTL model represents the multi-task learning network model based on aerodynamic effects proposed in the present invention.

[0237] Table 3 Evaluation indicators of neural network models

[0238]

[0239]

[0240] As can be seen from Table 3, compared with the MLP neural network, the ABMTL and LSTM neural networks are more complex in structure, resulting in their average training time and network model size being several times that of the MLP neural network.

[0241] However, the single-task model only considers a certain aerodynamic component of eVTOL. From the perspective of the overall aerodynamic model of eVTOL, that is, calculating the training time and occupied space of all single-task learning network models, their sum is 1265 sec and 5026 KB respectively, which is 1.32 times the training time consumed by ABMTL and 1.85 times the storage space occupied by the model. In addition, the total number of training and validation sample points used for the single-task learning to reach the current training result is 1.369×10 6The number is 1.65 times the number of sample points used by ABMTL. Nevertheless, the ABMTL network model has higher accuracy, while there is a large amount of duplicate information in the sample data of single-task learning.

[0242] In summary, single-task learning is generally applicable to the case where only the aerodynamic characteristics of some components of the eVTOL need to be modeled. Its model structure is relatively simple, and it can obtain the aerodynamic model of the corresponding degrees of freedom relatively quickly and accurately. However, the ABMTL network model has a high training efficiency. Its advantage lies in that it can construct only one neural network structure and the dataset for network training and verification, with less data requirements and high data utilization rate. Compared with single-task learning, it can further reduce the time cost, computational cost, and storage cost of aerodynamic modeling. On this basis, through a detailed analysis of the aerodynamic characteristics, the structure of the ABMTL neural network can be optimized and improved, making the feature-sharing part of different subtasks clearer, thereby enhancing the generalization ability of the model.

[0243] The present invention also provides an aerodynamic modeling system for eVTOL based on multi-task learning. The aerodynamic modeling system for eVTOL based on multi-task learning includes a memory and a processor. The memory can store programs, and the processor can execute these programs. When the processor executes the program, it implements the aerodynamic modeling method for eVTOL based on multi-task learning provided by the present invention.

[0244] The present invention also provides a computer-readable storage medium. The computer-readable storage medium is a non-volatile storage medium or a non-transient storage medium, on which computer instructions are stored. When the computer instructions run, they execute the steps corresponding to any of the above methods, which will not be elaborated here.

[0245] When the implementation process file of the aerodynamic modeling method for eVTOL based on multi-task learning is a computer program, it can also be stored in a computer-readable storage medium as an article of manufacture. For example, the computer-readable storage medium can include, but is not limited to, magnetic storage devices (such as hard disks, floppy disks, magnetic strips), optical discs (such as compact discs (CDs), digital versatile discs (DVDs)), smart cards, and flash memory devices (such as electrically erasable programmable read-only memories (EPROMs), cards, sticks, key drives). In addition, the various storage media described herein can represent one or more devices and / or other machine-readable media for storing information. The term "machine-readable medium" can include, but is not limited to, wireless channels and various other media (and / or storage media) that can store, contain, and / or carry code and / or instructions and / or data.

[0246] An aerodynamic modeling method for eVTOL based on multi-task learning provided by the present invention specifically has the following beneficial effects:

[0247] 1) Compared with the traditional physical mechanism modeling method, the present invention has stronger non - linear modeling ability, and does not require explicit expression of the mathematical model of the system, and has better generalization ability;

[0248] 2) Compared with the computational simulation method of fluid mechanics software, the present invention significantly reduces the computational cost and improves the modeling efficiency at the same time;

[0249] 3) Compared with using a single neural network model or single - task aerodynamic modeling method, the present invention has higher training efficiency based on multi - task learning. Its advantage lies in that only one neural network structure needs to be constructed, and a unified data set is used for network training and verification. The data requirements are less and the utilization rate is higher. Compared with single - task learning, it can further reduce the time cost, computational cost and storage cost of aerodynamic modeling;

[0250] 4) Compared with the neural network model of classical multi - task learning, through detailed analysis of aerodynamic characteristics, the ABMTL neural network established by the present invention can more accurately reflect the aerodynamic interference and coupled aerodynamic characteristics existing in the original data, so as to accurately predict the non - linear relationship between aerodynamic coefficients and influencing factors, and significantly improve the generalization ability of the model.

[0251] Although the above - mentioned methods are illustrated and described as a series of actions for simplicity of explanation, it should be understood and appreciated that these methods are not limited by the order of actions, because according to one or more embodiments, some actions may occur in a different order and / or occur concurrently with other actions that are illustrated and described herein or not illustrated and described herein but are understandable to those skilled in the art.

[0252] As shown in this application and the claims, unless the context clearly indicates an exception, words such as "a", "an", "one" and / or "the" are not specifically singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0253] Those skilled in the art will understand that information, signals and data can be represented using any of a variety of different technologies and techniques. For example, the data, instructions, commands, information, signals, bits, symbols, and chips described throughout the above description can be represented by voltage, current, electromagnetic waves, magnetic fields or magnetic particles, optical fields or optical particles, or any combination thereof.

[0254] Those skilled in the art will further appreciate that the various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, the various illustrative components, blocks, modules, circuits, and steps are described above in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.

[0255] The various illustrative logical modules and circuits described in connection with the embodiments disclosed herein can be implemented using a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.

[0256] The steps of a method or algorithm described in connection with the embodiments disclosed herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of both. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor such that the processor can read from, and write to, the storage medium. In the alternative, the storage medium may be integral to the processor. The processor and the storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In the alternative, the processor and the storage medium may reside as discrete components in a user terminal.

[0257] The above embodiments are provided for those skilled in the art to implement or use the present invention. Those skilled in the art can make various modifications or variations to the above embodiments without departing from the inventive concept of the present invention. Therefore, the protection scope of the present invention is not limited by the above embodiments, but should be the maximum scope that conforms to the innovative features mentioned in the claims.

Claims

1. An eVTOL aerodynamic modeling method based on multi-task learning, characterized in that: The following steps are involved: Step S1, analyzing the aerodynamic effects and mutual influences of various aerodynamic components of the eVTOL, identifying key characteristic parameters that affect the aerodynamic performance of the aircraft, and obtaining aerodynamic effect analysis results; Step S2, using the aerodynamic effect analysis results, combined with a multi-layer perceptron neural network and a long short-term memory neural network, to construct a multi-task learning model based on aerodynamic effect characteristics; Step S3, obtaining the original data of the aircraft, and constructing a sample data set according to the flight state and aerodynamic characteristics; Step S4, using the sample data set of step S3 to train the multi-task learning model of step S2, and perform verification and evaluation to obtain and output the multi-task learning model of eVTOL.

2. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 1, characterized in that: In step S1, the aerodynamic effects and mutual influences between the aerodynamic components of the eVTOL further include: The aerodynamic effects of an isolated rotor, the interaction between rotors, and the aerodynamic effects of the same components as fixed-wing aircraft.

3. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 1, characterized in that: In step S1, the key characteristic parameters affecting the aerodynamic performance of the aircraft further include: state and control parameters, intermediate parameters, special aerodynamic characteristic parameters, state parameters affected by aerodynamic effects, and aerodynamic coefficients; Starting from the state and control parameters, through the intermediate parameters, special aerodynamic characteristic parameters, the state parameters affected by the aerodynamic effect, until the aerodynamic coefficient, the coupling and circulation relationship between the factors affecting the aerodynamic effect is formed.

4. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 3 is characterized in that: The state and control parameters further include: airspeed, angular velocity, pressure altitude, rotor speed, tilt angle, and control surface deflection angle; The intermediate parameters further include: rotor force coefficient, induced speed, and center of gravity position; The special aerodynamic characteristic parameters further include: rotor interaction, rotor wake, rotor downwash, wing wake, wing downwash, rotor wake-fuselage-ground effect; The state variables affected by aerodynamic effects further include: rotor wake area, rotor wake radius, angle of attack, sideslip angle, and dynamic pressure; The aerodynamic coefficients further include: fuselage aerodynamic force and moment, wing aerodynamic force and moment, rotor aerodynamic force and moment, and control surface aerodynamic force and moment.

5. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 1, characterized in that: The step S2 further includes selecting a multi-layer perceptron neural network and a long short-term memory neural network as the basic neural networks of the multi-task learning model: The multi-layer perceptron neural network is used to model the aerodynamic effects of the wing, fuselage, and control surfaces; The long short-term memory neural network is used for aerodynamic effect modeling of the rotor.

6. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 1, characterized in that: The step S2 further includes using sparse connections and skip connections to realize neuron connections between network layers: The sparse connection uses the ratio of the number of neurons that each neuron sparsely connects to the next layer to the total number of neurons in the next layer as a hyperparameter of the network, and disconnects the connection of the neuron corresponding to the input of the previous layer; The skip connection uses a residual connection to directly transfer the output of a certain layer to the input of a deeper layer through matrix addition without adding any parameters.

7. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 1, characterized in that: The multi-task learning model of step S2 dynamically adjusts the weights of the contributions of different aerodynamic coefficients in each subtask to the total aerodynamic effect according to different input conditions for different subtasks corresponding to each aerodynamic component of the eVTOL, and weights the losses of all aerodynamic coefficients according to the weights to obtain a total loss function.

8. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 7, characterized in that: The step S2 further comprises: The weighted average method is used to adjust the weights of different aerodynamic coefficients of each subtask.

9. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 7, characterized in that: The step S2 further comprises: The Z-score method is used to normalize the weights of different aerodynamic coefficients of each subtask.

10. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 1, characterized in that: The multi-task learning model based on aerodynamic effect features in step S2 further includes an input layer, a plurality of shared layers, a task-specific layer and an output layer: The input layer receives the state and control variables from the aerodynamic effect analysis as input; The several shared layers are used to extract aerodynamic effect features and share parameters of different subtasks; The task-specific layer receives the output of the last shared layer required by each subtask, completes nonlinear transformation through respective activation functions, and obtains the aerodynamic coefficients corresponding to each subtask with different weights; The output layer is used to output aerodynamic force and moment coefficients.

11. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 10, characterized in that: The plurality of shared layers include long short-term memory network neurons and multilayer perceptron neurons, which are used to extract the cyclic characteristics and nonlinear characteristics of the aerodynamic characteristics respectively; The several shared layers include a first shared layer, a second shared layer, and a third shared layer: The first shared layer is sparsely connected to the input layer to represent intermediate parameters that affect aerodynamic characteristics; The second shared layer is sparsely connected to the first shared layer, and the output processed by the long short-term memory network neurons is weightedly combined with the output of the multilayer perceptron neurons and then transmitted to the third shared layer; The third shared layer is sparsely connected with the second shared layer, and is combined with the output of the second shared layer, the output of the long short-term memory network neurons, and the output of the multilayer perceptron neurons, and then transmitted to the task-specific layer.

12. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 1, characterized in that: The step S3 further comprises: Analyze the sample data requirements and the characteristics of the original data, and divide different sample data into continuous variables and discrete variables; For continuous variables, non-uniform sampling is performed based on aerodynamic characteristics, and uniform sampling is performed based on flight envelope constraints; For discrete variables, Latin hypercube sampling is used.

13. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 12, characterized in that: The non-uniform sampling includes variable granularity sampling, and the uniform sampling includes rejection sampling.

14. An eVTOL aerodynamic modeling system based on multi-task learning, characterized in that: Including memory and processor: The memory is used to store instructions executable by the processor; The processor is used to execute the instructions to implement the method according to any one of claims 1-13.

15. A computer storage medium having computer instructions stored thereon, wherein when the computer instructions are executed by a processor, the method according to any one of claims 1 to 13 is executed.

Images