An eVTOL aerodynamic modeling method based on multi-task learning

Abstract

The present application relates to the technical field of aviation aerodynamic modeling, more specifically, to an eVTOL aerodynamic modeling method based on multi-task learning. The method comprises: analyzing the aerodynamic effects and mutual influences of each aerodynamic component of the eVTOL, identifying key characteristic parameters affecting the aerodynamic performance of the aircraft, and obtaining aerodynamic effect analysis results; using the aerodynamic effect analysis results, combining a multilayer perceptron neural network and a long short-term memory neural network, and constructing a multi-task learning model based on aerodynamic effect characteristics; obtaining aircraft original data, constructing a sample data set according to the flight state and aerodynamic characteristics; training the multi-task learning model using the sample data set, and performing verification and evaluation to obtain and output the multi-task learning model of the eVTOL. The present application can more accurately reflect the aerodynamic interference and coupling characteristics, and realize rapid and accurate prediction of the complex aerodynamic characteristics of the eVTOL.

Description

Technical Field

[0001] This invention relates to the field of aerodynamic modeling technology, and more specifically, to an aerodynamic modeling method and system for eVTOL (electric vertical takeoff and landing) based on multi-task learning, particularly an efficient and high-fidelity aerodynamic modeling method and system for eVTOL based on multi-task learning. Background Technology

[0002] With the acceleration of global urbanization, the continuous growth of urban populations has brought unprecedented challenges to transportation systems. How to effectively alleviate traffic congestion, improve safety performance, and reduce carbon emissions while meeting the ever-increasing travel demands of passengers has become a critical issue that urgently needs to be addressed in the global transportation sector. Against this backdrop, Urban Air Mobility (UAM) has emerged, with eVTOL (eVTOL) attracting significant attention as an innovative solution. As a core carrier describing the aerodynamic characteristics of aircraft, the eVTOL aerodynamic model not only aids in aircraft structural design and flight control system development and verification but also provides model support for research such as flight dynamics simulation and pilot training simulators. Therefore, developing high-fidelity aerodynamic models for eVTOL has become a crucial requirement.

[0003] In the field of eVTOL aerodynamic modeling, existing aerodynamic modeling methods can be mainly divided into four paradigms: experimental methods, theoretical analysis, numerical analysis, and data-driven methods. Experimental methods primarily acquire basic data through wind tunnel tests or flight tests, providing data support for other modeling methods. Theoretical analysis methods are inductive summarization studies based on computational analysis, using deductive reasoning to analyze the physical meaning of parameters and model typical flow characteristics, providing theoretical prior knowledge for other methods. With the rapid development of computing technology, computational fluid dynamics (CFD) has become an important research tool in the third paradigm. However, these methods often face problems such as high computational resource consumption and long modeling cycles when dealing with the complex coupled aerodynamic characteristics of eVTOL, especially in handling multi-source data fusion and accurate calculation of non-equilibrium turbulence.

[0004] In recent years, data-driven machine learning methods, as the fourth paradigm, have been widely applied in aerodynamics research. This method leverages massive amounts of data generated from wind tunnel tests, flight tests, and CFD numerical simulations, combined with prior knowledge provided by aerodynamic theory, to perform data-driven modeling through neural networks. Research shows that neural network-based data-driven modeling methods exhibit better predictive performance when dealing with nonlinear problems, demonstrating significant advantages over traditional physical mechanism modeling methods and CFD simulation methods. These advantages include powerful nonlinear modeling capabilities, lower computational costs, no need to explicitly express the system's mathematical model, and good generalization ability.

[0005] Current research in the eVTOL field mainly focuses on control algorithms and trajectory planning, with relatively little research on aerodynamic modeling. For example, Chinese invention patent CN117010219A proposes an eVTOL aircraft obstacle avoidance simulation method, providing safety assurance for actual flight through a simulation platform. In terms of aerodynamic modeling, existing research mainly employs physical mechanisms, CFD, and data-driven methods: Chinese invention patent CN104298805A discloses a CFD aerodynamic modeling method for hypersonic vehicles, covering the complete process of numerical modeling, mesh generation, computational solution, and result analysis; Chinese invention patent...

[0006] CN117195763A discloses a method for modeling the aerodynamics of a fixed-wing aircraft considering wind interference, enabling online prediction of aerodynamic forces of a fixed-wing aircraft under unknown wind conditions; Chinese Invention Patent

[0007] CN116484713A discloses an aerodynamic modeling method for tail-retraction separation scenarios based on neural networks, which significantly reduces the workload of CFD numerical simulation using neural network methods; Chinese Invention Patent

[0008] CN117910386A discloses a transfer learning aerodynamic modeling method based on dendritic neural networks. By combining dendritic neural networks with transfer learning, the computational cost is further reduced while ensuring prediction accuracy.

[0009] Despite significant advancements in eVTOL control algorithms and trajectory planning, existing technologies still have considerable limitations in predicting eVTOL aerodynamic performance. Traditional mechanism-based or CFD-based modeling methods are typically limited to single tasks such as predicting airfoil and fuselage aerodynamic coefficients, and are only applicable to a limited number of airfoil shapes or specific angles of attack and Mach number ranges. Furthermore, these methods require independent modeling of each aerodynamic component and simulation of aerodynamic coupling effects by quantifying parameters such as induced velocity and airflow angle. This approach is difficult to apply to scenarios such as flight control system fine-tuning and human-in-the-loop flight simulation, and the simulation accuracy of aerodynamic coupling phenomena is limited by the accuracy of parameter calculations, resulting in relatively low fidelity.

[0010] Existing data-driven aerodynamic modeling methods mostly employ a single deep neural network structure, which struggles to adapt to the aerodynamic characteristics and mutual interference of different modules in an eVTOL, such as the rotor and fuselage. Since an eVTOL is essentially a complex multibody dynamic system, it typically consists of multiple aerodynamic components. Traditional neural network methods usually require separate modeling and training for individual components such as the fuselage and rotor. While individual network structures are relatively simple, the sheer number of networks severely impacts training efficiency.

[0011] Therefore, a new eVTOL aerodynamic modeling method is urgently needed to solve the above problems. Summary of the Invention

[0012] The purpose of this invention is to provide an aerodynamic modeling method for eVTOL based on multi-task learning, which solves the problem that existing eVTOL aerodynamic modeling methods are difficult to accurately simulate the aerodynamic effects and mutual interference between multiple aerodynamic components.

[0013] To achieve the above objectives, this invention provides an eVTOL aerodynamic modeling method based on multi-task learning, comprising the following steps:

[0014] Step S1: Analyze the aerodynamic effects and interactions of each aerodynamic component of the eVTOL, identify key characteristic parameters affecting the aerodynamic performance of the aircraft, and obtain the aerodynamic effect analysis results.

[0015] Step S2: Using the aerodynamic effect analysis results, combined with multilayer perceptron neural network and long short-term memory neural network, a multi-task learning model based on aerodynamic effect characteristics is constructed.

[0016] Step S3: Obtain the raw data of the aircraft and construct a sample dataset based on the flight status and aerodynamic characteristics;

[0017] Step S4: Use the sample dataset from step S3 to train the multi-task learning model from step S2, and perform validation and evaluation to obtain and output the eVTOL multi-task learning model.

[0018] In some embodiments, step S1, the aerodynamic effects and interactions between the various pneumatic components of the eVTOL, further includes:

[0019] The aerodynamic effects of isolated rotors, the mutual influence between rotors, and the aerodynamic effects of components identical to those in fixed-wing aircraft.

[0020] 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 aerodynamic effects, and aerodynamic coefficients.

[0021] Starting from the state and control parameters, the process passes through intermediate parameters, special aerodynamic characteristic parameters, and state parameters affected by aerodynamic effects, until the aerodynamic coefficient, thus forming a coupling and cyclical relationship among the factors influencing aerodynamic effects.

[0022] In some embodiments, the state and control parameters further include: airspeed, angular velocity, barometric altitude, rotor speed, tilt angle, and control surface deflection angle;

[0023] The intermediate parameters further include: rotor force coefficient, induced velocity, and center of gravity position;

[0024] The specific aerodynamic characteristic parameters further include: rotor interaction, rotor wake, rotor downwash, wing wake, wing downwash, and rotor wake-fuselage-ground effect;

[0025] The state variables affected by aerodynamic effects further include: rotor wake area, rotor wake radius, angle of attack, sideslip angle, and dynamic pressure;

[0026] The aerodynamic coefficients further include: fuselage aerodynamic forces and moments, wing aerodynamic forces and moments, rotor aerodynamic forces and moments, and control surface aerodynamic forces and moments.

[0027] In some embodiments, step S2 further includes selecting a multilayer perceptron neural network and a long short-term memory neural network as the base neural networks of the multi-task learning model:

[0028] The multilayer perceptron neural network is used for aerodynamic effect modeling of wings, fuselage, and control surfaces;

[0029] The long short-term memory neural network is used for aerodynamic effect modeling of the rotor.

[0030] In some embodiments, step S2 further includes implementing neuron connections between network layers using sparse connections and skip connections:

[0031] The sparse connection uses the proportion of the number of neurons in the next layer that are sparsely connected to each neuron to the next layer to the total number of neurons in the next layer as a hyperparameter of the network, and disconnects the connections of neurons to the previous layer's input.

[0032] The skip connection uses a residual connection, which directly passes the output of a certain layer to the input of a deeper layer through matrix addition, without adding any parameters.

[0033] In some embodiments, the multi-task learning model in step S2 dynamically adjusts the weights of different aerodynamic coefficients contributing to the total aerodynamic effect in each sub-task according to different input conditions for each aerodynamic component of the eVTOL, and weights the losses of all aerodynamic coefficients to obtain the total loss function.

[0034] In some embodiments, step S2 further includes:

[0035] The weighted average method is used to adjust the weights of different aerodynamic coefficients for each sub-task.

[0036] In some embodiments, step S2 further includes:

[0037] The Z-score method is used to normalize the weights of different aerodynamic coefficients for each subtask.

[0038] In some embodiments, the multi-task learning model based on aerodynamic effect features in step S2 further includes an input layer, several shared layers, a task-specific layer, and an output layer:

[0039] The input layer receives state and control variables from aerodynamic effect analysis as input;

[0040] The shared layers are used to extract aerodynamic effect features and share parameters for different subtasks;

[0041] The task-specific layer receives the output of the last shared layer required by each subtask, completes the nonlinear transformation through its respective activation function, and obtains the aerodynamic coefficients corresponding to each subtask by weighting with different weights.

[0042] The output layer is used to output aerodynamic forces and torque coefficients.

[0043] In some embodiments, the plurality of shared layers include long short-term memory network neurons and multilayer perceptron neurons, used to extract the cyclic characteristics and nonlinear features of aerodynamic properties, respectively.

[0044] The plurality of shared layers includes a first shared layer, a second shared layer, and a third shared layer:

[0045] The first shared layer, which is sparsely connected to the input layer, is used to represent intermediate parameters that affect aerodynamic characteristics;

[0046] The second shared layer, which is sparsely connected to the first shared layer, transmits the output processed by the long short-term memory network neurons and the output of the multilayer perceptron neurons in a weighted combination to the third shared layer.

[0047] The third shared layer is sparsely connected to the second shared layer. The output of the second shared layer is combined with the outputs of the long short-term memory network neurons and the multilayer perceptron neurons in a weighted manner and then transmitted to the task-specific layer.

[0048] In some embodiments, step S3 further includes:

[0049] The sample data requirements and characteristics of the original data were analyzed, and the different sample data were divided into continuous variables and discrete variables.

[0050] For continuous variables, non-uniform sampling is performed based on aerodynamic characteristics, while uniform sampling is performed based on flight envelope constraints.

[0051] For discrete variables, Latin hypercube sampling is used.

[0052] In some embodiments, the non-uniform sampling includes variable granularity sampling, and the uniform sampling includes rejection sampling.

[0053] To achieve the above objectives, the present invention provides an eVTOL aerodynamic modeling system based on multi-task learning, comprising a memory and a processor:

[0054] The memory is used to store instructions that can be executed by a processor;

[0055] The processor is configured to execute the instructions to implement the method as described above.

[0056] To achieve the above objectives, the present invention provides a computer storage medium storing computer instructions thereon, wherein when the computer instructions are executed by a processor, the method described above is performed.

[0057] This 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, combined with the basic structure of multilayer perceptron (MLP) and long short-term memory network (LSTM), and an innovative data sampling method, compared with traditional physical mechanism modeling and fluid dynamics simulation, it significantly improves nonlinear modeling ability and generalization ability, reduces computational and time costs, and improves sample data utilization and training efficiency. It can more accurately reflect aerodynamic disturbances and coupling characteristics, and realize rapid and accurate prediction of complex aerodynamic characteristics of eVTOL. Attached Figure Description

[0058] The above and other features, properties and advantages of the present invention will become more apparent from the following description taken in conjunction with the accompanying drawings and embodiments, in which the same reference numerals always denote the same features, wherein:

[0059] Figure 1A step diagram of an eVTOL aerodynamic modeling method based on multi-task learning according to an embodiment of the present invention is disclosed;

[0060] Figure 2 A logic diagram for generating the aerodynamic effect of an eVTOL according to an embodiment of the present invention is disclosed;

[0061] Figure 3 A residual block structure diagram according to an embodiment of the present invention is disclosed;

[0062] Figure 4 A schematic diagram of the ABMTL neural network model structure according to an embodiment of the present invention is disclosed;

[0063] Figure 5 A schematic diagram of the sample dataset construction process according to an embodiment of the present invention is disclosed;

[0064] Figure 6 A flowchart of a multi-task learning aerodynamic modeling process for a certain eVTOL according to an embodiment of the present invention is disclosed;

[0065] Figure 7a A graph showing the variation of hovering power with tensile force coefficient under hovering conditions according to an embodiment of the present invention is disclosed.

[0066] Figure 7b A graph showing the variation of the hovering thrust coefficient with the pitch angle according to an embodiment of the present invention is disclosed.

[0067] Figure 7c A graph showing the variation of the torque coefficient with the tension coefficient under hovering conditions according to an embodiment of the present invention is disclosed.

[0068] Figure 8a A graph showing the variation of the thrust coefficient with pitch angle under different advance ratios in vertical takeoff and landing conditions according to an embodiment of the present invention is disclosed.

[0069] Figure 8b The graphs showing the variation of torque coefficient with thrust coefficient under different rotor angles of attack in vertical takeoff and landing conditions according to an embodiment of the present invention are disclosed.

[0070] Figure 9a The figure shows the lift coefficient as a function of angle of attack under different flap / aileron modes in vertical takeoff and landing according to an embodiment of the present invention.

[0071] Figure 9b The graphs showing the drag coefficient versus angle of attack under different flap / aileron modes in vertical takeoff and landing operation according to an embodiment of the present invention are disclosed.

[0072] Figure 9c The graphs showing the pitching moment coefficient versus angle of attack under different flap / aileron modes in vertical takeoff and landing operation according to an embodiment of the present invention are disclosed.

[0073] Figure 10a The graphs showing the variation of lateral force coefficient with yaw angle under different flap / aileron modes in vertical takeoff and landing according to an embodiment of the present invention are disclosed.

[0074] Figure 10b The figure shows the curves of the roll moment coefficient as a function of yaw angle under different flap / aileron modes in vertical takeoff and landing according to an embodiment of the present invention;

[0075] Figure 10c The graphs showing the variation of yaw moment coefficient with angle of attack under different flap / aileron modes in vertical takeoff and landing according to an embodiment of the present invention are disclosed.

[0076] Figure 11a A graph showing the variation of wing lift coefficient with angle of attack under vertical takeoff and landing conditions according to an embodiment of the present invention is disclosed.

[0077] Figure 11b A graph showing the variation of wing drag coefficient with angle of attack under vertical takeoff and landing conditions according to an embodiment of the present invention is disclosed.

[0078] Figure 12a A graph showing the variation of the horizontal stabilizer lift coefficient with angle of attack under vertical takeoff and landing conditions according to an embodiment of the present invention is disclosed.

[0079] Figure 12b A graph showing the variation of the horizontal stabilizer drag coefficient with angle of attack under vertical takeoff and landing conditions according to an embodiment of the present invention is disclosed.

[0080] Figure 13a A graph showing the variation of the lateral force coefficient of the vertical tail under vertical takeoff and landing conditions with the sideslip angle according to an embodiment of the present invention is disclosed.

[0081] Figure 13b A graph showing the variation of the vertical tail drag coefficient with angle of attack under vertical takeoff and landing conditions according to an embodiment of the present invention is disclosed.

[0082] Figure 14a The graphs showing the lift coefficient versus angle of attack under different flap / aileron modes during the transition condition according to an embodiment of the present invention are disclosed.

[0083] Figure 14b The graphs showing the drag coefficient as a function of angle of attack under different flap / aileron modes during the transition condition according to an embodiment of the present invention are disclosed.

[0084] Figure 14c The diagram showing the pitching moment coefficient as a function of angle of attack under different flap / aileron modes during the transition condition according to an embodiment of the present invention is disclosed.

[0085] Figure 15a A graph showing the change of wing lift coefficient with angle of attack under transition conditions according to an embodiment of the present invention is disclosed.

[0086] Figure 15b A graph showing the change of wing lift coefficient with angle of attack under transition conditions according to an embodiment of the present invention is disclosed. Detailed Implementation

[0087] To make the objectives, technical solutions, and advantages of this invention clearer, the 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 for illustrative purposes only and are not intended to limit the invention.

[0088] To address the problems of mutual coupling and interference among aerodynamic components such as rotors, fuselages, wings, and various control surfaces in existing eVTOL (electric vertical takeoff and landing) configurations, and the difficulty of accurately simulating the unsteady aerodynamic characteristics of eVTOLs using traditional aerodynamic modeling methods, which also incur high time costs and computational resource requirements, this 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, this method can effectively improve the prediction accuracy of complex aerodynamic characteristics of eVTOLs, while also increasing sample data utilization and training speed, and reducing model training costs.

[0089] Figure 1 The following diagram illustrates the steps of an eVTOL aerodynamic modeling method based on multi-task learning according to an embodiment of the present invention: Figure 1 As shown, the present invention proposes an eVTOL aerodynamic modeling method based on multi-task learning, which includes the following steps:

[0090] Step S1: Analyze the aerodynamic effects and interactions of each aerodynamic component of the eVTOL, identify key characteristic parameters affecting the aerodynamic performance of the aircraft, and obtain the aerodynamic effect analysis results.

[0091] Step S2: Using the aerodynamic effect analysis results, combined with multilayer perceptron neural network and long short-term memory neural network, a multi-task learning model based on aerodynamic effect characteristics is constructed.

[0092] Step S3: Obtain the raw data of the aircraft and construct a sample dataset based on the flight status and aerodynamic characteristics;

[0093] Step S4: Use the sample dataset from step S3 to train the multi-task learning model from step S2, and perform validation and evaluation to obtain and output the eVTOL multi-task learning model.

[0094] This invention proposes an eVTOL aerodynamic modeling method based on multi-task learning. By constructing an efficient and high-fidelity multi-task learning (ABMTL) model based on the aerodynamic effect characteristics of eVTOL, it can achieve rapid and accurate prediction of the aerodynamic coefficients of multiple aerodynamic components of eVTOL.

[0095] More specifically, firstly, the aerodynamic characteristics of each component of the eVTOL are analyzed, and two network architectures suitable for nonlinearity (MLP) and recurrent characteristic learning (LSTM) are selected as the basic structures for multi-task learning. Secondly, considering the incomplete connectivity and cross-layer influence between input and output variables, skip connections and cross-layer connections are selected as neuron connection methods, along with a loss function based on aerodynamic characteristics, to construct the network structure and establish a multi-task learning model structure based on aerodynamic characteristics. Thirdly, based on the form and variation characteristics of multi-source raw data, innovative data sampling methods (including non-uniform sampling based on aerodynamic characteristics, uniform sampling based on flight envelope constraints, and discrete state parameter sampling based on the Latin hypersolution method) are adopted to rapidly construct a high-quality sample dataset. Ultimately, this achieves the effects of improving the prediction accuracy of complex aerodynamic characteristics of eVTOL, increasing the utilization rate of sample data, accelerating training speed, and reducing model training costs.

[0096] These steps will be described in detail below. It should be understood that, within the scope of this invention, the above-described technical features of this invention and the technical features specifically described below (such as in the embodiments) can be combined and related to each other to form preferred technical solutions.

[0097] Step S1: Analyze the aerodynamic effects and interactions between the various aerodynamic components of the eVTOL, identify the key characteristic parameters affecting the aerodynamic performance of the aircraft, and obtain the aerodynamic effect analysis results.

[0098] In this embodiment, the aerodynamic effects and mutual influences between the various pneumatic components of the eVTOL further include:

[0099] The aerodynamic effects of isolated rotors, the mutual influence between rotors, and the aerodynamic effects of components identical to those in fixed-wing aircraft.

[0100] Furthermore, the analysis process and results for the aerodynamic effects of an isolated rotor are as follows:

[0101] For eVTOLs of various structures, there is usually an aerodynamic effect generated by an isolated rotor. When the rotational motion of the rotor drives the airflow through the rotor, an induced velocity is generated in the opposite direction to the thrust.

[0102] Induced velocity, along with control variables such as rotor speed and rotor pitch angle, affects the rotor's flow state, namely the rotor's inflow ratio and forward ratio, and further influences the rotor's thrust coefficient and torque coefficient. Various rotor studies have shown that induced velocity is also a function of thrust, forward velocity, and inflow velocity, thus creating an implicit cycle between thrust and induced velocity.

[0103] Blade flapping refers to a phenomenon caused by the different flow conditions at the blades during the forward and backward motion of the rotor relative to the incoming flow. Blade flapping alters the magnitude of the thrust component on the rotor disk, thus affecting the rotor's inflow dynamics.

[0104] Furthermore, the analysis process and results regarding the interaction between rotors are as follows:

[0105] Building upon the aerodynamic effects of an isolated rotor, the interaction between rotors manifests as follows: the rotor's induced velocity, effective angle of attack, and other state variables are affected by the wake and vortex interactions generated by other rotors. Simultaneously, the rotor is also affected by the wake and aerodynamic blocking effects of the wings, nacelles, and fuselage.

[0106] Furthermore, the analysis process and results for the aerodynamic effects of components identical to those in fixed-wing aircraft are as follows:

[0107] For parts of the eVTOL structure that are the same as those in a fixed-wing aircraft, such as the wings, fuselage, and control surfaces, it can be seen from the development research of eVTOL aircraft models by analogy with computational fluid dynamics (CFD) and wind tunnel testing that the aerodynamic and moment coefficients of these components can be expressed as functions of state variables and some control variables.

[0108] Unlike the aerodynamic effects of fixed-wing aircraft, these aerodynamic components generate strong aerodynamic interference effects with the rotor, especially during the vertical takeoff and transition phases of low-speed flight.

[0109] Taking the wing, which has the most complex aerodynamic effects among all components, as an example, the rotor will generate a rotor wake that impacts part of the wing area during flight.

[0110] Therefore, in research on tiltrotor aircraft, the aerodynamic models of the wing module are usually divided into aerodynamic models of the slipstream region caused by the rotor wake and aerodynamic models of the free flow region unaffected by the rotor wake.

[0111] The aerodynamic effects of the rotor on the control surfaces such as the fuselage and horizontal stabilizer are similar to those on the wings. They are mainly manifested through the induced velocities generated by the rotor wake on each component. These induced velocities change the overall airflow velocity of each component, while affecting the airflow angle and dynamic pressure, and thus the aerodynamic force and moment coefficient.

[0112] In addition, as mentioned earlier, the wings, fuselage, and control surfaces also generate wakes and affect the aerodynamic effects of the rotor.

[0113] By analyzing the aerodynamic effects of the above components, we can obtain the key characteristic parameters that affect the aerodynamic performance of the aircraft and the relationships between them.

[0114] 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.

[0115] Figure 2 A logic diagram for generating the aerodynamic effect of an eVTOL according to an embodiment of the present invention is disclosed, such as... Figure 2 As shown, the state and control parameters further include: airspeed, angular velocity, barometric altitude, rotor speed, tilt angle, and control surface deflection angle;

[0116] The intermediate parameters further include: rotor force coefficient, induced velocity, and center of gravity position;

[0117] The specific aerodynamic characteristic parameters further include: rotor interaction, rotor wake, rotor downwash, wing wake, wing downwash, and rotor wake-fuselage-ground effect;

[0118] The state parameters affected by aerodynamic effects further include: rotor wake area, rotor wake radius, angle of attack, sideslip angle, and dynamic pressure;

[0119] 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, and control surface aerodynamic forces and moments.

[0120] The interactions and dependencies between these parameters constitute the complex aerodynamic model of eVTOL.

[0121] Starting from the state and control parameters of the eVTOL, through complex nonlinear relationships, the influence of the parameters is transmitted to intermediate parameters, special aerodynamic characteristic parameters, state parameters affected by aerodynamic effects, and finally to the aerodynamic coefficients of the eVTOL, including aerodynamic force and moment coefficients.

[0122] from Figure 2 As can be seen, factors influencing each layer, such as airspeed and rotor speed of the eVTOL, do not necessarily affect all variables in the next layer. Certain input parameters, such as nacelle tilt angle and control surface angle, can influence aerodynamic parameters like rotor wake area across layers, while state variables such as airspeed and airflow angle directly affect aerodynamic coefficients. Furthermore, the inflow dynamics of the rotor—the cycle of induced velocity and thrust coefficient—and the interactions between rotors and between the rotor and the wing lead to complex coupling and cyclical relationships among the aerodynamic influencing factors.

[0123] The aerodynamic models of each component of eVTOL are a nonlinear network of relationships interwoven with the incomplete correlations and cross-layer correlations between these parameters.

[0124] Step S2: Using the aerodynamic effect analysis results, combined with multilayer perceptron neural network and long short-term memory neural network, a multi-task learning model is constructed.

[0125] In this embodiment, by analyzing the aerodynamic characteristics and influencing factors of the eVTOL multi-module, a multi-task learning method is selected. This method can effectively utilize the inherent relationship between tasks and has high model training efficiency, and is used for data-driven aerodynamic modeling of eVTOL.

[0126] Building upon this foundation, and considering the incomplete connections and cross-layer influences between input and output variables, a multi-task learning model framework based on aerodynamic characteristics was constructed. The model's input consists of some state variables and control variables of the eVTOL, while the output comprises aerodynamic forces and torque coefficients. The model, constructed based on aerodynamic characteristics and the relationships between influencing factors, employs a skip-connection and sparse-connection ABMTL neural network, enabling it to more accurately fit the relationship between nonlinear aerodynamic characteristics and influencing factors, and predict aerodynamic coefficients.

[0127] More specifically, step S2 further includes: selecting MLP (Multilayer Perceptron Neural Network) and LSTM (Long Short-Term Memory Neural Network) as the base neural networks for the multi-task learning model:

[0128] The MLP is used for aerodynamic effect modeling of control surfaces such as wings, fuselage, horizontal tail and vertical tail, and there are complex nonlinear relationships between the aerodynamic effects of these components.

[0129] The LSTM is used for modeling the aerodynamic effects of rotors. The aerodynamic effects of rotors have temporal and cyclic characteristics, and LSTM can effectively handle such nonlinear temporal coupling problems.

[0130] As fundamental neural network types in multi-task learning models, MLP and LSTM are used to construct neurons in shared layers and task-specific layers.

[0131] Step S2 further includes: using sparse connections and skip connections to achieve interlayer neuron connections.

[0132] Based on the analysis of the characteristics of eVTOL aerodynamic effects, the nonlinear relationship network of the eVTOL aerodynamic model contains two features: incomplete correlation between parameters and cross-layer correlation exists in the generation process of aerodynamic effects. In this embodiment, sparse connections and skip connections are used to realize the incomplete correlation between parameters and cross-layer correlation, respectively.

[0133] Taking the wing wake of the eVTOL aerodynamic effect as an example, the wing wake typically affects state variables such as airspeed and airflow angle of the fuselage, horizontal stabilizer, and vertical stabilizer, but does not include all state variables. For this type of aerodynamic characteristic, sparse connectivity is used for expression.

[0134] The sparse connection uses the proportion of neurons in the next layer that are sparsely connected to each neuron as a hyperparameter of the network, and disconnects the connection between the neuron and the input of the previous layer based on this, that is, sets the weight of the neuron relative to the input to zero, so as to represent the incomplete correlation between aerodynamic effect parameters.

[0135] Sparse connections limit the number of connections between neurons, avoiding the influence of irrelevant factors on intermediate variables, making the network structure closer to the real aerodynamic model, while reducing the number of parameters that need to be calculated, thereby reducing computational costs and memory usage.

[0136] 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 the aerodynamic characteristics such as rotor wake across layers. State variables such as airspeed and airflow angle can directly affect the aerodynamic coefficient. Such aerodynamic effects are best represented by a jump connection, that is, by using matrix addition as the information transmission method.

[0137] The skip connection uses a residual connection, which directly passes the output of a certain layer to the input of a deeper layer through matrix addition, without adding any parameters.

[0138] Furthermore, in this embodiment, for residual connections, the number of neurons with residual connections in each layer and the number of neurons receiving information in the last two layers are used as hyperparameters, and a random method is used for neuron selection.

[0139] Residual connections pass the output of one layer directly to the input of a deeper layer through matrix addition by constructing residual blocks, without attaching any parameters. Figure 3 A residual block structure diagram according to an embodiment of the present invention is disclosed, such as... Figure 3 As shown, a single residual block consists of two weight layers, and their information transmission method is as follows:

[0140]

[0141] Where x is the residual block input and F(x) is the residual block output. σ is the residual function, W1 and W2 are the weights, b1 and b2 are the bias terms, and σ is the activation function between layers;

[0142] In this embodiment, the ReLU function is used as the activation function.

[0143] 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 transformed by the residual function. This is very similar to the characteristic of aerodynamic characteristics where influencing factors affect other parameters through intermediate variables. Furthermore, skip connections allow information to flow freely within the network, helping to transmit input information more completely to deeper layers of the network, such as input parameters that directly affect aerodynamic coefficients, thereby improving the model's expressive power and prediction accuracy.

[0144] Existing multi-task learning methods have limitations in eVTOL aerodynamic modeling, primarily in their inability to dynamically reflect the changes in the contribution of each sub-task's aerodynamic coefficients to the total aerodynamic effect under different input conditions. To address this issue, this invention proposes a loss function that incorporates aerodynamic effect characteristics. For each sub-task corresponding to aerodynamic components in eVTOL, the weights of the contributions of different aerodynamic coefficients to the total aerodynamic effect in each sub-task need to be dynamically adjusted according to different input conditions. Finally, the losses of all aerodynamic coefficients are weighted according to their respective weights to obtain the total loss function.

[0145] Because of the complex nonlinear relationship between the input state, control variables, and output aerodynamic coefficients, the weight of each subtask's output relative to the total task changes when the input samples change. Therefore, this invention employs a data-driven method to determine the weights of each output in the MTL (Multi-Task Learning) neural network. The specific steps for determining the weights of the aerodynamic coefficients for all subtasks are as follows:

[0146] Define a function to calculate the weights of the aerodynamic coefficients for all subtasks, where each weight represents the relative contribution of a certain aerodynamic coefficient to each aerodynamic component.

[0147] For the i-th subtask of multi-task learning, the output of the j-th aerodynamic coefficient is C. i,j Then its corresponding weight output α i,j The calculation formula is shown in equation (3).

[0148]

[0149] When constructing the ABMTL network structure of this invention, the weights are used as outputs and the training sample data is used as inputs to construct the corresponding nonlinear mapping function and embed it in the forward propagation process.

[0150] Step S2 further includes:

[0151] The weighted average method is used to adjust the weights of different aerodynamic coefficients for each sub-task.

[0152] To ensure that excessive fluctuations in the weights of the aerodynamic coefficients do not negatively impact model stability and training efficiency during dynamic adjustment, this embodiment employs a weighted average method to smoothly adjust the weights. This helps maintain the stability of the weights during dynamic changes. The expression for the weighted average method is as follows:

[0153]

[0154] in, w represents the weights calculated in the first n iterations during the t-th iteration of training. n This is the corresponding weighting factor.

[0155] In this embodiment, the results of the last four weight calculations are saved, and the weight factor values ​​are set to 0.4, 0.3, 0.2 and 0.1 respectively, so that the final result depends more on the recent weights.

[0156] Step S2 further includes:

[0157] The Z-score method is used to normalize the weights of different aerodynamic coefficients for each subtask.

[0158] After determining the weights using the above methods, 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:

[0159]

[0160] Where μ is the mean of the weights, σ ​​is the standard deviation of the weights, and α i,j Let be the weight of the j-th aerodynamic coefficient in the i-th subtask of multi-task learning. These are the normalized weights.

[0161] In the loss function of the ABMTL model, these dynamically adjusted weights are used. To adjust the losses for each aerodynamic coefficient and calculate the total loss, the corresponding expression is as follows:

[0162]

[0163] Among them, L i,j The loss is the aerodynamic coefficient of the j-th component in subtask i. The weights are dynamically adjusted, and this loss function replaces the MSE loss function in traditional multi-task learning.

[0164] The ABMTL network based on improved eVTOL aerodynamic characteristics proposed in this invention has a slightly longer computation time for the loss function during training. However, compared with traditional MTL neural networks that directly apply the MSE loss function, the ABMTL model has higher accuracy. Although the ABMTL structure becomes more complex due to the incorporation of eVTOL aerodynamic effects, it requires less total training time and storage space compared to methods that train multiple single-task neural networks separately.

[0165] Furthermore, given the coupling effect between the pneumatic components of eVTOL, the same sample data may be reused when training a single-task learning network model. The ABMTL model can effectively reduce this information reuse phenomenon.

[0166] Furthermore, the multi-task learning model based on aerodynamic effect features in step S2 further includes an input layer, several shared layers, a task-specific layer, and an output layer:

[0167] The input layer receives state and control variables from aerodynamic effect analysis as input;

[0168] The shared layers are used to extract aerodynamic effect features and share parameters for different subtasks;

[0169] The task-specific layer receives the output of the last shared layer required by each subtask, completes the nonlinear transformation through its respective activation function, and obtains the aerodynamic coefficients corresponding to each subtask by weighting with different weights.

[0170] The output layer is used to output aerodynamic forces and torque coefficients.

[0171] Figure 4 A schematic diagram of the ABMTL neural network model structure according to an embodiment of the present invention is disclosed, as follows: 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:

[0172] The input layer receives state and control variables from aerodynamic effect analysis as input;

[0173] The first shared layer, the second shared layer, the third shared layer, and the task-specific layer are hidden layers;

[0174] The first shared layer, which is sparsely connected to the input layer, is used to represent intermediate parameters that affect aerodynamic characteristics;

[0175] The second and third shared layers contain both LSTM neurons and MLP neurons to extract recurrent and nonlinear characteristics.

[0176] The second shared layer, which is sparsely connected to the first shared layer, transmits the output processed by the long short-term memory network neurons and the output of the MLP neurons in a weighted combination to the third shared layer.

[0177] The third shared layer is sparsely connected to the second shared layer. The output of the second shared layer is combined with the outputs of the long short-term memory network neurons and the multilayer perceptron neurons in a weighted manner and then transmitted to the task-specific layer.

[0178] During forward propagation, the output data of the second shared layer is processed by LSTM neurons, and the output of the last time step is extracted. This output is then weighted and fed into the next layer, the third shared layer, along with the output of the MLP neurons in the same layer. These shared layers are used by all subtasks, therefore their parameters need to be trained and validated using sample data from all subtasks.

[0179] In addition, the connection between each shared layer and the next layer is sparse, and the connection between each neuron and the next two layers is skip connection. These two connection methods reflect the characteristics of influencing factors not having a complete effect on aerodynamic properties and the cross-layer effect.

[0180] In the task-specific layer, neurons of different colors represent fully connected layers specific to each subtask. The task-specific layer receives the output of the last shared layer required by each subtask, performs nonlinear transformation through its own activation function, and obtains the aerodynamic coefficients corresponding to each subtask by weighting them with different weights.

[0181] The output layer is used to output aerodynamic forces and torque coefficients.

[0182] This invention constructs a multi-task learning model (ABMTL) based on aerodynamic effect features by combining MLP and LSTM neural networks and employing sparse and skip connection methods. This model can dynamically adjust the weights of the loss function, accurately reflect the aerodynamic effects of each aerodynamic component of the eVTOL, and significantly improve the model's prediction accuracy and stability.

[0183] Step S3: Obtain the raw data of the aircraft and construct a sample dataset based on the flight status and aerodynamic characteristics;

[0184] To ensure the quality and efficiency of the sample dataset, considering the data requirements and characteristics of the ABMTL network model, it is first necessary to clarify the constraints between the input and output parameters and then use appropriate methods to partition and expand the sample data. Through analysis of the sample data requirements and in-depth research into the characteristics of the original data, different sample data are rationally partitioned, and the following sampling methods are 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.

[0185] Figure 5 A schematic diagram illustrating the sample dataset construction process according to an embodiment of the present invention is shown, such as... Figure 5 As shown, based on the requirements and characteristics of the sample data, the data can be divided into continuous variables and discrete variables. For continuous variables, given their nonlinear relationships, non-uniform sampling, such as variable-granularity sampling, can be used; when based on flight envelope constraints, uniform sampling, such as rejection sampling, can be employed. Discrete variables, on the other hand, can be sampled using Latin hypercube sampling.

[0186] More specifically, the non-uniform sampling can employ variable granularity sampling;

[0187] For example, the lift coefficient and pitching moment coefficient both exhibit approximately linear changes within a certain angle of attack range, while the drag coefficient shows more significant nonlinear characteristics in the middle section of the range.

[0188] To ensure the rationality of sample data selection, after comprehensively considering the resolution of the original data and the range of values ​​of the input variables, the nonlinear region was processed by taking the union, and then the part with stronger nonlinearity was selected. The sampling granularity of the input variable intervals with nonlinear and linear changes in mapping relationship was set to 2° and 1°, respectively.

[0189] More specifically, the uniform sampling can employ a rejection sampling method;

[0190] For two-dimensional plane constraints, a rejection sampling method is used. Specifically, a rectangle containing the entire irregular region is selected as the proposal distribution. The process of random sampling and checking the constraints is repeated continuously to filter out points that fall within the flight envelope, thereby obtaining a certain number of samples.

[0191] More specifically, the Latin hypercube sampling further includes:

[0192] The intervals in each dimension are divided equally, and then a point is randomly selected in each sub-interval to ensure that the samples in each dimension are not concentrated in a certain interval, thereby reducing the clustering of sample points and ensuring the uniformity of the distribution of sample points in the multidimensional space.

[0193] Step S3 of this invention can significantly improve the efficiency of constructing the sample dataset while preserving the original data distribution characteristics. Ultimately, the ABMTL model achieves relatively high prediction accuracy with relatively low data requirements, verifying the effectiveness of the above sampling method.

[0194] Step S4: Use the sample dataset from step S3 to train the multi-task learning model from step S2, and then validate and evaluate it to finally output a multi-task learning model for eVTOL.

[0195] Training a neural network model typically involves determining a set of hyperparameter values, using a given dataset and loss function, and then adjusting the weights and biases of the neural network through the backpropagation algorithm to minimize the loss function.

[0196] Hyperparameters not only determine the size of the model, but also relate to the calculation method from input parameters to output parameters. Given a fixed sample dataset, they have a significant impact on the training efficiency and quality of the model, as well as the model accuracy.

[0197] Therefore, it is necessary to determine the appropriate hyperparameter values ​​through scientific methods.

[0198] In this embodiment, hyperparameters include the number of inputs and outputs, training period, batch size, number of neurons, number of hidden layers, memory sequence length, and overall error. These parameters determine the model's scale, computation method, training efficiency, and quality.

[0199] Furthermore, traditional deep learning methods typically use the mean squared error function as the loss function. However, in actual aerodynamic modeling, the weights of each aerodynamic coefficient and the output of each aerodynamic component in the overall task are constantly changing. This invention employs a loss function that incorporates aerodynamic effect characteristics, dynamically adjusting the weights of each sub-task's aerodynamic coefficient based on the input-output relationship of the samples using a data-driven approach. This ensures that the loss function accurately reflects the contribution of each aerodynamic coefficient to the overall task.

[0200] Cross-validation is used to evaluate the model's generalization ability and avoid overfitting. The model's accuracy, stability, and computational efficiency in predicting eVTOL aerodynamic coefficients can be assessed by comparing it with single-task MLP and LSTM neural network aerodynamic models.

[0201] This invention, through in-depth analysis of the aerodynamic effects of eVTOL, careful design of the neural network structure, reasonable generation of training datasets, and a complete training and verification process, and by introducing a multi-task learning model and innovative design combining aerodynamic effects, can effectively handle the complex coupling relationships between various components of eVTOL, thereby improving the accuracy of aerodynamic performance prediction and optimizing aircraft design and control strategies.

[0202] Figure 6 A flowchart of a multi-task learning aerodynamic modeling process for a certain eVTOL according to an embodiment of the present invention is disclosed, such as... Figure 6 As shown, taking a certain configuration of eVTOL as an example for aerodynamic modeling, this paper illustrates the eVTOL aerodynamic modeling method based on multi-task learning proposed in this invention. Steps S1 to S4 correspond to four stages: aerodynamic effect characteristic analysis, neural network model structure construction, dataset generation and partitioning, and model training and validation.

[0203] Step S1, Analysis of Aerodynamic Effect Characteristics

[0204] Taking the tiltrotor configuration eVTOL as an example, this paper analyzes the aerodynamic characteristics of each module of the tiltrotor aircraft and the complex coupling relationship between parameters based on aerodynamic modeling methods and physical mechanisms.

[0205] The configuration characteristics of tiltrotor aircraft determine that they exhibit extremely complex aerodynamic characteristics during both the vertical takeoff and transition phases.

[0206] During rotor operation, changes in the velocity and direction of the inflow will alter the flapping dynamics of the blades, and these two factors will work together to affect the thrust and induced velocity generated by the rotor.

[0207] Meanwhile, the rotor will generate aerodynamic interference to other rotors and the airframe, including but not limited to the rotor wake acting on the wings, horizontal stabilizer and vertical stabilizer. The complex coupling relationship between these aerodynamic characteristics is the main reason why traditional physical models are difficult to achieve a high degree of realism.

[0208] By analyzing the generation mechanism and process of aerodynamic characteristics in step S1 and obtaining the analysis results, a theoretical basis is provided for subsequent model selection, construction, and data requirements.

[0209] Step S2, Neural Network Model Structure Construction

[0210] Based on the aerodynamic mechanism analyzed in step S1, a suitable data-driven method is selected to construct the neural network model structure.

[0211] In this embodiment, a multi-task learning model is used in step S2.

[0212] For eVTOLs with various complex configurations, aerodynamic modeling involves multiple related tasks, namely, separate modeling of aerodynamic components such as the eVTOL rotor, fuselage, and wings. This makes multi-task learning more suitable for aerodynamic modeling of eVTOLs than single-task learning. Furthermore, compared to training a separate model for each task, multi-task learning offers advantages such as lower model training costs, higher data utilization, and faster training speed.

[0213] Based on this, the multi-task learning model used 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 aerodynamic effect characteristics of eVTOL on the basis of the classic neural network structure.

[0214] Step S3, Dataset Generation and Partitioning

[0215] The sample dataset is mainly used for selecting hyperparameters, training, and validating network models. Its quantity and quality play a decisive role in the accuracy of the network model.

[0216] To ensure the quality and efficiency of the sample dataset, it is constructed and partitioned according to the characteristics of the original data and the specific needs of the network model. Typically, the dataset is divided into training, validation, and test sets in a 7:3:1 ratio.

[0217] Step S4, Model Training and Validation

[0218] To verify the prediction accuracy of the ABMTL model proposed in this invention, a high-precision dataset was constructed using data from NASA's publicly available XV-15 tiltrotor aircraft. Single-task MLP and LSTM neural network aerodynamic models were established for each aerodynamic component, corresponding to the characteristics of the fuselage, wings, control surfaces, and rotor. The mean squared error function (MSE) was used as the evaluation metric for model training, and the network hyperparameter combination with the minimum overall error was determined, as shown in Table 1.

[0219] Table 1 Hyperparameter Combinations of Neural Network Models

[0220]

[0221] As shown in Table 1, the ABMTL model with adaptive weight loss function constructed in this invention, although relatively complex in structure, has a lower overall error compared to the MTL model using the traditional MSE loss function. Furthermore, the error of this model is significantly lower than that of the single-task network model for each aerodynamic component, thus improving the model's accuracy to a certain extent.

[0222] Based on this, test samples with the same unsteady aerodynamic characteristics were selected to conduct comparative analysis of different models.

[0223] Figures 7a-15b The figures show the prediction results of the aerodynamic coefficients of each aerodynamic component under hovering, vertical takeoff and landing, and two transitional flight conditions. The Wind tunnel data, LSTM / MLP network prediction, and MTL network prediction results in the figure refer to the original wind tunnel test dataset, the prediction results of the single-task LSTM / MLP model of each aerodynamic component, and the prediction results of the ABMTL model of this invention, respectively.

[0224] For tiltrotor aircraft, during hovering, the rotor thrust coefficient increases almost linearly with the rotor pitch angle, while the rotor torque coefficient exhibits a non-linear relationship with the thrust coefficient. The hovering power of the rotor fluctuates after reaching a certain level as the thrust increases. The aerodynamic characteristics of the rotor during vertical takeoff and transition are similar, but in these cases, the lift of the tiltrotor aircraft primarily originates from the rotor, and the rotor aerodynamic coefficients only include the rotor thrust and torque coefficients.

[0225] Figures 7a to 7c The curve shows the change in rotor aerodynamic coefficients during hovering. Specifically, Figure 7a This is a graph showing the variation of hovering power with the tensile force coefficient under hovering conditions. Figure 7b This is a graph showing the variation of the thrust coefficient with the pitch angle during hovering. Figure 7c This is a graph showing the variation of torque coefficient with tension coefficient under hovering conditions. Figures 8a to 8b The curve shows the variation of the rotor aerodynamic coefficient under vertical takeoff and landing conditions. Specifically, Figure 8a The graph shows the variation of the thrust coefficient with pitch angle under different advance ratios during vertical takeoff and landing. Figure 8b The graph shows the variation of torque coefficient with thrust coefficient under different rotor angles of attack during vertical takeoff and landing.

[0226] like Figures 7a-8b As shown, the aerodynamic coefficients of the rotor are respectively determined by an LSTM neural network and...

[0227] ABMTL neural networks were used for prediction, and the prediction results of both neural networks can reflect the changing trend of the data well.

[0228] During vertical takeoff and transition phases, the fuselage modules also provide a portion of the aerodynamic forces and torques as the airflow angle and state variables change. Figures 9a to 15b The curves in the figure represent the prediction results of the original wind tunnel test dataset, the MLP neural network, and the multi-task learning neural network under the vertical take-off and landing conditions and the transition conditions, respectively.

[0229] Figures 9a to 9c The curve shows the variation of the longitudinal aerodynamic coefficient of the fuselage under vertical takeoff and landing conditions. Specifically, Figure 9a This is a graph showing the lift coefficient as a function of angle of attack under different flap / aileron modes during vertical takeoff and landing. Figure 9b This is a graph showing the drag coefficient as a function of angle of attack under different flap / aileron modes during vertical takeoff and landing. Figure 9c The graph shows the pitching moment coefficient as a function of angle of attack under different flap / aileron modes in vertical takeoff and landing conditions.

[0230] Figures 10a to 10c The curve showing the variation of the fuselage's lateral aerodynamic coefficient under vertical takeoff and landing conditions is as follows: Figure 10a The graph shows the variation of the lateral force coefficient with yaw angle under different flap / aileron modes in vertical takeoff and landing conditions. Figure 10bThe graph shows the variation of the roll moment coefficient with yaw angle under different flap / aileron modes in vertical takeoff and landing (VTOL) conditions. Figure 10c The graph shows the variation of yaw moment coefficient with angle of attack under different flap / aileron modes in vertical takeoff and landing conditions.

[0231] Figures 11a to 11b The curve shows the variation of the wing's aerodynamic coefficient under vertical takeoff and landing conditions. Specifically, Figure 11a This is a graph showing the variation of the wing lift coefficient with angle of attack under vertical takeoff and landing (VTOL) conditions. Figure 11b This is a graph showing the variation of the wing drag coefficient with angle of attack under vertical takeoff and landing conditions.

[0232] Figures 12a to 12b The curve showing the variation of the horizontal stabilizer aerodynamic coefficient under vertical takeoff and landing conditions is as follows: Figure 12a This is a graph showing the lift coefficient of the horizontal stabilizer as a function of angle of attack under vertical takeoff and landing conditions. Figure 12b The graph shows the variation of the horizontal stabilizer drag coefficient with angle of attack under vertical takeoff and landing conditions.

[0233] Figures 13a to 13b The curve showing the variation of the vertical tail aerodynamic coefficient under vertical takeoff and landing conditions is as follows: Figure 13a This is a graph showing the variation of the lateral force coefficient of the vertical tail as a function of the sideslip angle under vertical takeoff and landing conditions. Figure 13b The graph shows the variation of the tail drag coefficient with angle of attack under vertical takeoff and landing conditions.

[0234] Figures 14a to 14c The curve showing the change in the longitudinal aerodynamic coefficient of the fuselage under transitional operating conditions, specifically... Figure 14a The graph shows the lift coefficient versus angle of attack under different flap / aileron modes during transitional operation. Figure 14b The graph shows the drag coefficient as a function of angle of attack under different flap / aileron modes during transitional operation. Figure 14c The diagram shows the pitching moment coefficient as a function of angle of attack under different flap / aileron modes during transitional operation.

[0235] Figures 15a to 15b The curve showing the change in aerodynamic coefficient of the wing under transitional operating conditions, specifically... Figure 15a This is a graph showing the change in wing lift coefficient with angle of attack under transitional operating conditions. Figure 15b The curve shows the change of wing lift coefficient with angle of attack under transitional operating conditions.

[0236] Table 2 shows the maximum and average relative errors 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 5% accuracy requirement of the simulation model. In contrast, the ABMTL network model of this invention can more accurately predict the nonlinear relationship between influencing factors and aerodynamic coefficients, with its maximum relative error controlled within 5% and its average relative error accuracy maintained at around 1%. In eVTOL simulations or actual systems, these errors typically accumulate over time or with increasing iterations. Therefore, even if the difference in predicted relative errors is small, after multiple iterations or long simulation periods, these errors can accumulate to a level sufficient to cause significant changes in the generated aerodynamic forces, resulting in an overall decrease in the realism of the flight simulation.

[0237] Table 2 Maximum Relative Error / Average Relative Error of Aerodynamic Coefficients

[0238]

[0239] Based on this, the two modeling schemes were compared in terms of various indicators other than accuracy, namely, network training time, network model storage space (saved in .pth format), and the number of training and validation sample data points required to achieve similar prediction accuracy. Table 3 shows that MLP1, MLP2, MLP3, and LSTM represent the single-task learning network models of the fuselage, wing, control surfaces, and rotor module, respectively, while the ABMTL model represents the multi-task learning network model based on aerodynamic effects proposed in this invention.

[0240] Table 3 Evaluation Indicators for Neural Network Models

[0241]

[0242]

[0243] As can be seen from Table 3, ABMTL and LSTM neural networks are more complex in structure than MLP neural networks, resulting in their average training time and network model size being several times that of MLP neural networks.

[0244] However, the single-task model only considers one aerodynamic component of the eVTOL. If we consider the overall aerodynamic model of the eVTOL, i.e., calculating the training time and storage space occupied by all single-task learning network models, their sum is 1265 seconds and 5026KB respectively. This is equivalent to 1.32 times the training time consumed by ABMTL and 1.85 times the storage space occupied by the model. Furthermore, the total number of training and validation sample points used to achieve the current training result in single-task learning is 1.369 × 10⁻⁶. 6The number of samples used is 1.65 times that of ABMTL. Despite this, the ABMTL network model has higher accuracy, while the sample data for single-task learning contains a large amount of repetitive information.

[0245] In summary, single-task learning is typically suitable for situations where only the aerodynamic characteristics of a portion of the eVTOL components need to be modeled. Its model structure is relatively simple, enabling quick and accurate acquisition of the corresponding degrees of freedom aerodynamic models. However, ABMTL's network model training efficiency is higher. Its advantage lies in the fact that it only requires building a single neural network structure and a dataset for network training and validation, resulting in less data requirements and higher data utilization. Compared to single-task learning, it further reduces the time, computation, and storage costs of aerodynamic modeling. Furthermore, by optimizing and improving the structure of the ABMTL neural network through detailed analysis of aerodynamic characteristics, the shared features of different sub-tasks can be made clearer, thereby improving the model's generalization ability.

[0246] This invention also provides an aerodynamic modeling system for eVTOL based on multi-task learning. This system includes a memory and a processor. The memory stores a program, and the processor executes the program. When the processor executes the program, it implements the aerodynamic modeling method for eVTOL based on multi-task learning provided by this invention.

[0247] The present invention also provides a computer-readable storage medium, which is a non-volatile storage medium or a non-transient storage medium, and stores computer instructions thereon. When the computer instructions are executed, they perform the steps corresponding to any of the above methods, which will not be described in detail here.

[0248] When the implementation process document of the eVTOL aerodynamic modeling method based on multi-task learning is a computer program, it can also be stored as an article of manufacture in a computer-readable storage medium. For example, computer-readable storage media can include, but are not limited to, magnetic storage devices (e.g., hard disks, floppy disks, magnetic stripes), optical discs (e.g., compact discs (CDs), digital multifunction discs (DVDs)), smart cards, and flash memory devices (e.g., electrically erasable programmable read-only memory (EPROM), cards, sticks, key drives). Furthermore, the various storage media described herein can represent one or more devices and / or other machine-readable media used 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) capable of storing, containing, and / or carrying code and / or instructions and / or data.

[0249] The present invention provides an aerodynamic modeling method for eVTOL based on multi-task learning, which has the following advantages:

[0250] 1) Compared with traditional physical mechanism modeling methods, this invention has stronger nonlinear modeling capabilities and does not require explicit expression of the mathematical model of the system, thus having better generalization ability.

[0251] 2) Compared with the computational simulation methods of fluid dynamics software, the present invention significantly reduces computational costs while improving modeling efficiency;

[0252] 3) Compared with using a single neural network model or a single-task aerodynamic modeling method, the training efficiency of the present invention based on multi-task learning is higher. Its advantage is that only one neural network structure needs to be built and a unified dataset 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, computing cost and storage cost of aerodynamic modeling.

[0253] 4) Compared with the classic multi-task learning neural network model, the ABMTL neural network established by this invention through detailed analysis of aerodynamic characteristics can more accurately reflect the aerodynamic interference and coupled aerodynamic characteristics in the original data, thereby accurately predicting the nonlinear relationship between aerodynamic coefficients and influencing factors and significantly improving the generalization ability of the model.

[0254] Although the methods described above are illustrated and depicted as a series of actions for the sake of simplicity, it should be understood and appreciated that these methods are not limited by the order of the actions, as some actions may occur in a different order and / or concurrently with other actions from the illustrations and descriptions herein or not illustrated and described herein but which may be understood by those skilled in the art, according to one or more embodiments.

[0255] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0256] Those skilled in the art will understand that information, signals, and data can be represented using any of a variety of different techniques and arts. 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, light fields or optical particles, or any combination thereof.

[0257] Those skilled in the art will further appreciate that the various illustrative logic blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or a combination of both. To clearly illustrate this interchangeability between hardware and software, the various illustrative components, blocks, modules, circuits, and steps are described above in a generalized manner in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. Those skilled in the art may implement the described functionality in different ways for each specific application, but such implementation decisions should not be construed as departing from the scope of the invention.

[0258] The various illustrative logic modules and circuits described in conjunction with the embodiments disclosed herein may be implemented or performed 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. The general-purpose processor may be a microprocessor, but in alternatives, it 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 cooperating with a DSP core, or any other such configuration.

[0259] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may 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, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor such that the processor can read and write information to / from the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and storage medium may reside as discrete components in the user terminal.

[0260] 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 changes 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 to the above embodiments, but should be the maximum scope that conforms to the innovative features mentioned in the claims.

Claims

1. A multi-task learning-based eVTOL aerodynamic modeling method, characterized in that, Includes the following steps: Step S1: Analyze the aerodynamic effects and interactions of each aerodynamic component of the eVTOL, identify key characteristic parameters affecting the aerodynamic performance of the aircraft, and obtain the aerodynamic effect analysis results. Step S2: Using the aerodynamic effect analysis results, combined with multilayer perceptron neural network and long short-term memory neural network, a multi-task learning model based on aerodynamic effect characteristics is constructed. Step S3: Obtain the raw data of the aircraft and construct a sample dataset based on the flight status and aerodynamic characteristics; Step S4: Use the sample dataset from step S3 to train the multi-task learning model from step S2, and perform validation and evaluation to obtain and output the eVTOL multi-task learning model. 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, the process passes through intermediate parameters, special aerodynamic characteristic parameters, and state parameters affected by aerodynamic effects, until the aerodynamic coefficient, thus forming a coupling and cyclical relationship among the factors influencing aerodynamic effects. The multi-task learning model based on aerodynamic effect features in step S2 further includes an input layer, several shared layers, a task-specific layer, and an output layer: The input layer receives state and control variables from aerodynamic effect analysis as input; The shared layers are used to extract aerodynamic effect features and share parameters for different subtasks; The shared layers include long short-term memory network neurons and multilayer perceptron neurons, which are used to extract the cyclic characteristics and nonlinear features of aerodynamic properties, respectively. The plurality of shared layers includes a first shared layer, a second shared layer, and a third shared layer: The first shared layer, which is sparsely connected to the input layer, is used to represent intermediate parameters that affect aerodynamic characteristics; The second shared layer, which is sparsely connected to the first shared layer, transmits the output processed by the long short-term memory network neurons and the output of the multilayer perceptron neurons in a weighted combination to the third shared layer. The third shared layer is sparsely connected to the second shared layer. The output of the second shared layer is combined with the output of the long short-term memory network neurons and the multilayer perceptron neurons in a weighted combination and then transmitted to the task-specific layer. The task-specific layer receives the output of the last shared layer required by each subtask, performs nonlinear transformation through its respective activation function, and obtains the aerodynamic coefficients corresponding to each subtask by weighting with different weights. The output layer is used to output aerodynamic forces and torque coefficients.

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 various pneumatic components of the eVTOL further include: The aerodynamic effects of isolated rotors, the mutual influence between rotors, and the aerodynamic effects of components identical to those in fixed-wing aircraft.

3. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 1, 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 velocity, and center of gravity position; The specific aerodynamic characteristic parameters further include: rotor interaction, rotor wake, rotor downwash, wing wake, wing downwash, and 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 forces and moments, wing aerodynamic forces and moments, rotor aerodynamic forces and moments, and control surface aerodynamic forces and moments.

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

5. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 1, characterized in that, Step S2 further includes using sparse connections and skip connections to achieve neuron connections between network layers: The sparse connection uses the proportion of the number of neurons in the next layer that are sparsely connected to each neuron to the next layer to the total number of neurons in the next layer as a hyperparameter of the network, and disconnects the connections of neurons to the previous layer's input. The skip connection uses a residual connection, which directly passes the output of a certain layer to the input of a deeper layer through matrix addition, without adding any parameters.

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

7. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 6, characterized in that, Step S2 further includes: The weighted average method is used to adjust the weights of different aerodynamic coefficients for each sub-task.

8. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 6, characterized in that, Step S2 further includes: The Z-score method is used to normalize the weights of different aerodynamic coefficients for each subtask.

9. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 1, characterized in that, Step S3 further includes: The sample data requirements and characteristics of the original data were analyzed, and the different sample data were divided into continuous variables and discrete variables. For continuous variables, non-uniform sampling is performed based on aerodynamic characteristics, while uniform sampling is performed based on flight envelope constraints. For discrete variables, Latin hypercube sampling is used.

10. The eVTOL aerodynamic modeling method based on multi-task learning according to claim 9, characterized in that, The non-uniform sampling includes variable granularity sampling, and the uniform sampling includes rejection sampling.

11. An eVTOL aerodynamic modeling system based on multi-task learning, characterized in that, Including memory and processor: The memory is used to store instructions that can be executed by a processor; The processor is configured to execute the instructions to implement the method as described in any one of claims 1-10.

12. A computer storage medium having stored thereon computer instructions, wherein when the computer instructions are executed by a processor, the method as described in any one of claims 1-10 is performed.

Images