Digital twin of aircraft, model construction method, data processing method and device

Abstract

The application discloses a digital twin of an aircraft, a model construction method, a data processing method and device. The proxy model is deployed in a digital twin system of an electric vertical take-off and landing aircraft. The proxy model construction method comprises: obtaining original training data, each training sample in the original training data comprising flight state parameters, flight environment parameters and structure parameters of the electric vertical take-off and landing aircraft; processing the original training data based on three different fidelity fluid mechanics numerical simulation methods, and correspondingly obtaining low-fidelity training data, medium-fidelity training data and high-fidelity training data, each training sample in each fidelity training data comprising aerodynamic load data simulated by the fluid mechanics numerical simulation method based on the corresponding fidelity according to the training sample. Based on the low-fidelity training data, a Bayesian neural network model is pre-trained to obtain a pre-trained model; based on the medium-fidelity training data and the high-fidelity training data, transfer learning is performed on the pre-trained model to obtain a multi-fidelity proxy model, and the multi-fidelity proxy model predicts flight physical data such as aerodynamic load according to actual flight parameters of the electric vertical take-off and landing aircraft in the digital twin system.

Description

Technical Field

[0001] This invention belongs to the field of aircraft simulation technology, and in particular relates to a digital twin of an aircraft, a model building method, a data processing method and device. Background Technology

[0002] Existing simulation models for electric vertical take-off and landing (eVTOL) aircraft either rely on high-fidelity (HF) data, resulting in high computational costs and long development cycles, or on low-fidelity (LF) data, leading to insufficient accuracy. Most existing eVTOL simulation models are "static, one-time modeling," unable to adapt to customized modifications and configuration adjustments throughout the entire lifecycle of eVTOL aircraft, requiring repeated modeling and exhibiting poor adaptability. Existing simulation models struggle to cover the diverse configurations of multi-rotor eVTOL aircraft, dynamic rotor speeds or accelerations, fault scenarios (e.g., rotor failure), and unsteady aerodynamic effects (e.g., rotor aerodynamic hysteresis, transient load fluctuations). Furthermore, existing literature lacks effective modeling methods for rotor dynamic responses, failing to support safe design under complex operating conditions. The airworthiness certification of electric vertical take-off and landing aircraft has extremely high requirements for the accuracy of simulation models, the completeness of scene coverage, and the controllability of risks. Traditional models, due to insufficient accuracy, incomplete scene coverage, and inability to quantify uncertainties, are unable to provide the required load data and safety analysis results, resulting in long certification cycles and high costs.

[0003] Traditional databases are typically generated through wind tunnel experiments or low-fidelity numerical simulations, which often have limited data coverage and cannot fully encompass the dynamic responses of aircraft under various operating conditions and attitudes in complex environments. While wind tunnel experiments offer high accuracy, they are expensive and time-consuming. High-fidelity methods in numerical simulations, although accurate, are computationally complex and require substantial computing resources, hindering the rapid acquisition of necessary data. Given the significant differences in accuracy and complexity between data generated by different simulation methods, effectively integrating data of varying fidelity to balance high fidelity and computational efficiency remains a challenge in current technology. Summary of the Invention

[0004] This invention provides a digital twin, model building method, data processing method, and apparatus for an aircraft, which solves the technical problem of the imbalance between fidelity and efficiency in the simulation models of electric vertical take-off and landing aircraft in the prior art, and the inability to cover the dynamic response of the aircraft under multiple operating conditions and attitudes in complex environments.

[0005] According to a first aspect of the present invention, a method for constructing a proxy model is provided, the proxy model being deployed in a digital twin system of an electric vertical takeoff and landing aircraft, the method comprising: Obtain raw training data, wherein each training sample in the raw training data includes flight state parameters, flight environment parameters, and structural parameters of the electric vertical takeoff and landing aircraft; The original training data is processed using three different fluid dynamics numerical simulation methods with varying fidelity to obtain low-fidelity training data, medium-fidelity training data, and high-fidelity training data. The training samples in each type of training data include aerodynamic load data simulated by the corresponding fluid dynamics numerical simulation method. Based on the low-fidelity training data, the Bayesian neural network model is pre-trained to obtain a pre-trained model. Based on the medium-fidelity training data and the high-fidelity training data, the pre-trained model is transferred to obtain a multi-fidelity proxy model. The multi-fidelity proxy model is used in the digital twin system to predict aerodynamic load data based on the actual flight parameters of the electric vertical take-off and landing aircraft.

[0006] In conjunction with the first aspect, in some embodiments, the original training data is processed using three different fidelity fluid dynamics numerical simulation methods to obtain low-fidelity training data, medium-fidelity training data, and high-fidelity training data, including: The original training data is processed using the facet method to obtain the low-fidelity training data; The original training data is processed using the vortex particle method to obtain the medium-fidelity training data. The original training data is processed using turbulence simulation methods to obtain the high-fidelity training data.

[0007] In conjunction with the first aspect, in some embodiments, the step of performing transfer learning on the pre-trained model based on the medium-fidelity training data and the high-fidelity training data to obtain a multi-fidelity agent model includes: Freeze the weight parameters of the bottom N layers in the pre-trained model, and fine-tune the weight parameters of the top M layers based on the mid-fidelity training data to obtain the intermediate proxy model; Freeze the weight parameters of the bottom N layers and at least one intermediate layer in the intermediate proxy model, and fine-tune the weight parameters of the remaining intermediate layers and the top layer in the intermediate proxy model based on the high-fidelity training data to obtain the multi-fidelity proxy model.

[0008] In conjunction with the first aspect, in some embodiments, the proxy model construction method further includes: The intermediate proxy model is evaluated to obtain its performance. Based on the model performance of the intermediate proxy model, determine at least one intermediate layer that needs to be frozen.

[0009] In conjunction with the first aspect, in some embodiments, the proxy model construction method further includes: The performance of the multi-fidelity proxy model is evaluated to obtain the model performance of the multi-fidelity proxy model; If the performance of the multi-fidelity proxy model is lower than a preset performance threshold, an iterative search is performed based on a Bayesian optimization strategy to determine the next optimal combination of hyperparameters. Based on the next optimal combination of hyperparameters, as well as the low-fidelity training data, the medium-fidelity training data, and the high-fidelity training data, the Bayesian neural network model is retrained and transfer learned until the preset number of iterations or convergence conditions are reached.

[0010] In conjunction with the first aspect, in some embodiments, the iterative search based on the Bayesian optimization strategy to determine the next optimal combination of hyperparameters includes: Based on the initial multiple hyperparameter combinations and the estimated model performance under each initial hyperparameter combination, a probabilistic surrogate model is constructed, wherein the probabilistic surrogate model is the relationship between hyperparameters and model performance. Candidate hyperparameter combinations are selected by the acquisition function, and the model performance under the candidate hyperparameter combinations is evaluated by the objective function and the constructed probabilistic surrogate model until the maximum number of iterations is reached or the estimated model performance reaches the performance convergence threshold, thus obtaining the next set of optimal hyperparameter combinations.

[0011] In conjunction with the first aspect, in some embodiments, the proxy model construction method, by performing performance evaluation on the multi-fidelity proxy model to obtain the model performance of the multi-fidelity proxy model, includes: Each sample to be predicted in the validation set is input into the multi-fidelity proxy model to obtain the probability distribution parameters of the predicted aerodynamic load for each sample to be predicted. For each of the samples to be predicted, the probability distribution parameters of the aerodynamic load prediction values ​​of the sample to be predicted are sampled by Monte Carlo sampling for the multi-fidelity surrogate model to obtain multiple aerodynamic load prediction values ​​for the sample to be predicted. Based on the multiple aerodynamic load prediction values, the aerodynamic load prediction result and confidence interval for the sample to be predicted are obtained. The model performance of the multi-fidelity surrogate model is determined based on the aerodynamic load prediction results and / or confidence intervals corresponding to each sample to be predicted in the validation set.

[0012] According to a second aspect of the present invention, a data processing method for an electric vertical takeoff and landing (EVTOL) aircraft is provided, comprising: Acquire manipulation control signals and atmospheric disturbance signals, wherein the atmospheric disturbance signals are generated based on environmental parameters; The control signals and atmospheric disturbance signals are input to the flight dynamics module of the digital twin system of the electric vertical takeoff and landing aircraft. The flight dynamics module deploys a multi-fidelity proxy model, which is obtained based on the proxy model construction method described in any embodiment of the first aspect. The flight mechanics module determines the aircraft state parameters based on the control signals, and the multi-fidelity proxy model predicts the aerodynamic load data of the electric vertical takeoff and landing aircraft based on the aircraft state parameters, the atmospheric disturbance signals, and the structural parameters of the electric vertical takeoff and landing aircraft. The flight mechanics module simulates the target response signal of the electric vertical takeoff and landing aircraft based on the aerodynamic load data.

[0013] According to a third aspect of the present invention, a proxy model construction apparatus is provided, wherein the proxy model is deployed in a digital twin system of an electric vertical takeoff and landing aircraft, characterized in that the proxy model construction apparatus comprises: The raw data acquisition unit is used to acquire raw training data, wherein each training sample in the raw training data includes flight state parameters, flight environment parameters, and structural parameters of the electric vertical takeoff and landing aircraft. The training data preparation unit is used to process the original training data based on three different fluid dynamics numerical simulation methods with different fidelity, so as to obtain low-fidelity training data, medium-fidelity training data and high-fidelity training data. The training samples in each fidelity training data include aerodynamic load data simulated based on the corresponding fluid dynamics numerical simulation method. The pre-training unit is used to pre-train the Bayesian neural network model based on the low-fidelity training data to obtain a pre-trained model. The transfer learning unit is used to perform transfer learning on the pre-trained model based on the medium-fidelity training data and the high-fidelity training data to obtain a multi-fidelity proxy model. The multi-fidelity proxy model is used in the digital twin system to predict aerodynamic load data based on the actual flight parameters of the electric vertical take-off and landing aircraft.

[0014] According to a fourth aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the proxy model construction method described in any embodiment of the first aspect, or the data processing method for an electric vertical take-off and landing aircraft described in any embodiment of the second aspect.

[0015] According to a fifth aspect of the present invention, a digital twin system for an electric vertical takeoff and landing aircraft is provided, comprising: The flight mechanics module and the aeroelasticity module are independently configured, and the flight mechanics module and the aeroelasticity module communicate with each other through a standardized interface; The multi-fidelity proxy model is obtained based on the proxy model construction method described in any embodiment of the first aspect. The multi-fidelity proxy model is deployed in the flight mechanics module of the digital twin system and is used to predict the aerodynamic load data of the electric vertical take-off and landing aircraft based on the flight environment data, structural data and flight state data of the electric vertical take-off and landing aircraft, so that the flight mechanics module can simulate the target response signal of the electric vertical take-off and landing aircraft based on the aerodynamic load data.

[0016] The one or more technical solutions provided in the embodiments of the present invention achieve at least the following technical effects or advantages: In this embodiment of the invention, raw training data is obtained. Each training sample in the raw training data includes flight state parameters, flight environment parameters, and structural parameters of the electric vertical takeoff and landing (EVTOL) aircraft. The raw training data is processed using three different fidelity hydrodynamic numerical simulation methods to obtain low-fidelity training data, medium-fidelity training data, and high-fidelity training data. The training samples in each fidelity training data include aerodynamic load data simulated by the corresponding fidelity hydrodynamic numerical simulation method. Based on the low-fidelity training data, a Bayesian neural network model is pre-trained to obtain a pre-trained model. Based on the medium-fidelity training data and the high-fidelity training data, the pre-trained model is transferred to obtain a multi-fidelity surrogate model. The multi-fidelity surrogate model is used in the digital twin system to predict aerodynamic loads and other flight physical data based on the actual flight parameters of the EVTOL aircraft. The aforementioned technical solution achieves iterative optimization of the surrogate model in the Digital Twin (DT) system of an electric vertical takeoff and landing (eVTOL) aircraft, based on a low-fidelity pre-trained model and integrating limited medium-fidelity (MF) and high-fidelity (HF) training data. This enables the DT system to operate throughout the entire development, certification, and operation phases of the eVTOL aircraft, dynamically integrating numerical simulation and experimental data to form an adaptively adjustable DT system. When the eVTOL undergoes customized modifications or configuration adjustments, data-driven optimization of the surrogate model ensures real-time updates to maintain the applicability of the DT system. This addresses the issues of short model lifecycles and poor adaptability inherent in static, one-time modeling, improving data utilization and effectively integrating data from different levels. The surrogate model construction achieves a balance between high fidelity and computational efficiency. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 A flowchart of a proxy model construction method according to some embodiments of the present invention is shown; Figure 2A It shows Figure 1 A schematic diagram of pre-training and transfer learning; Figure 2B The prediction results of the multi-fidelity proxy model and the proxy models constructed by the three existing methods are shown. Figure 3 A flowchart of a data processing method for an electric vertical takeoff and landing aircraft according to some embodiments of the present invention is shown; Figure 4 A schematic diagram of the architecture of a data processing method for an electric vertical takeoff and landing aircraft according to some embodiments of the present invention is shown; Figure 5 A system architecture diagram of a digital twin system for an electric vertical takeoff and landing aircraft according to some embodiments of the present invention is shown; Figure 6 A schematic diagram of the proxy model construction apparatus according to some embodiments of the present invention is shown; Figure 7 A structural diagram of a computer device according to some embodiments of the present invention is shown. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention provides a method for constructing a proxy model, which is then deployed in the digital twin system of an electric vertical takeoff and landing aircraft. Figure 1 A flowchart of a proxy model construction method according to some embodiments of the present invention is shown. For example... Figure 1 As shown, the proxy model construction method includes the following steps S101-S104.

[0021] Step S101: Obtain the original training data, wherein each training sample in the original training data includes the flight state parameters, flight environment parameters, and structural parameters of the electric vertical take-off and landing aircraft.

[0022] Understandably, the original training data is obtained through flight tests, wind tunnel tests, and simulations of the electric vertical takeoff and landing (EVTOL) aircraft. Each training sample in the original training data includes the flight state parameters, flight environment parameters, and structural parameters of the EVTOL aircraft. The original training data is the basic data support for building the digital twin and Bayesian neural network surrogate model of the aircraft.

[0023] In some embodiments, flight state parameters include position, linear velocity, angular velocity, acceleration, attitude angles (pitch, roll, yaw), sideslip angle, etc., reflecting the operating state of eVTOL and serving as the core input for model prediction of the next operating state. In some embodiments, flight environment parameters include atmospheric pressure, wind field parameters (wind direction, wind speed, wind shear), turbulence parameters (turbulence intensity, turbulence scale), etc., reflecting external disturbance factors and serving as key variables explaining differences in aircraft state under the same structure and control commands. Structural parameters are the geometric parameters of lifting components (including rotors or wings), such as diameter, airfoil, fuselage material, etc.

[0024] Step S102: Based on three different fluid dynamics numerical simulation methods with varying fidelity, generate low-fidelity training data, medium-fidelity training data, and high-fidelity training data from the original training data.

[0025] In some embodiments, the original training data is processed using the panel method to obtain the low-fidelity training data. The panel method is a numerical method in Computational Fluid Dynamics (CFD) based on potential flow theory. For each training sample in the original training data, the surface of the lifting component (rotor) is discretized into several panels based on the structural parameters of that training sample. Panel parameters are calculated for each panel. Singularities (sources, sinks, vortices, etc.) are placed on each panel to simulate the flow of fluid around the lifting component. Boundary conditions are set based on the flight environment parameters and flight state parameters of the training sample to establish governing equations. The singularity strength is solved based on the governing equations, and low-fidelity aerodynamic load data is calculated based on the singularity strength. Each training sample in the low-fidelity training data includes flight state parameters, flight environment parameters, structural parameters as input data, and low-fidelity aerodynamic load data as the true label.

[0026] In some embodiments, the original training data is processed using the Vortex Particle Method (VPM) to obtain the mid-fidelity training data. The Vortex Particle Method is a fluid simulation method based on vortex dynamics. For each training sample in the original training data, numerical simulation is performed based on the flight state parameters, flight environment parameters, and structural parameters of that training sample to obtain mid-fidelity aerodynamic load data. It is understood that each training sample in the mid-fidelity training data includes flight state parameters, flight environment parameters, and structural parameters as input data, as well as mid-fidelity aerodynamic load data as the true label.

[0027] In some embodiments, the original training data is processed using a turbulence simulation method to obtain the high-fidelity training data. It is understood that the turbulence simulation method used can be any one of Reynolds Average Navier-Stokes (RANS), Direct Numerical Simulation (DNS), Large Eddy Simulation (LES), and Detached Eddy Simulation (DES). For each training sample in the original training data, numerical simulation is performed using any of the above-mentioned turbulence simulation methods based on the flight state parameters, flight environment parameters, and structural parameters of the training sample to obtain high-fidelity aerodynamic load data. It is understood that each training sample in the high-fidelity training includes flight state parameters, flight environment parameters, and structural parameters as input data, and high-fidelity aerodynamic load data as the true label.

[0028] Understandably, the aerodynamic load data for each fidelity level (low-fidelity, medium-fidelity, and high-fidelity) includes at least one of the following types of forces on the rotor or wing: lift, drag, lateral force, and moments about the three principal axes. The moments about the three principal axes include pitch moment, roll moment, and yaw moment.

[0029] Step S103: Based on the low-fidelity training data, pre-train the Bayesian Neural Network (BNNs) model to obtain the pre-trained model.

[0030] In some embodiments, the Bayesian neural network model is a fully connected Bayesian neural network model. Figure 2A It shows Figure 1 A schematic diagram illustrating pre-training and transfer learning. (For example...) Figure 2AAs shown in this embodiment of the invention, the topological structure of the Bayesian neural network model includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is consistent with the feature dimension of the input data of the training samples, and it is used to receive the feature vector after preprocessing the input data. The hidden layer can be set to X layers, and each hidden layer contains Y neurons. The activation function can be a rectified linear unit (ReLU) or a hyperbolic tangent function (Tanh). The weights and biases of each neuron follow a Gaussian prior distribution.

[0031] In this embodiment of the invention, the prediction of aerodynamic loads (such as predicting rotor thrust, fuselage surface pressure, aerodynamic torque, etc.) is a regression task, and the number of nodes in the output layer is consistent with the dimension of the prediction target. For example, if lift, drag, and pitching moment are predicted simultaneously, then the output layer has 3 nodes. A linear activation function is used, and the output layer outputs the probability distribution parameters of the aerodynamic load prediction results.

[0032] Taking the prediction of unsteady thrust of the rotor of an eVTOL as an example, the topology of the Bayesian neural network model can be as follows: the input layer has 8 nodes (corresponding to the feature dimensions of the input data: rotor speed, wind speed, blade angle of attack, vortex particle vortex intensity, vortex core radius, atmospheric density, blade diameter, and forward velocity); the hidden layer has 3 layers, each with 64 neurons, and the activation function is ReLU; the output layer has 2 nodes (corresponding to the Gaussian distribution parameters of the mean and variance of the output rotor thrust).

[0033] In a Bayesian neural network model, all weights θ are random variables following an independent and identically distributed Gaussian distribution. The uncertainty of the weights is described by the prior distribution, likelihood function, and posterior distribution. The specific steps are as follows: 1. Prior distribution p(θ): It is assumed that all weights follow a Gaussian distribution with mean μ and variance σ², which is used to encode the initial uncertainty of the weights; 2. Likelihood function p(D|θ): describes the probability of the observed data D (the correspondence between input x and output y) occurring given weight θ. For regression tasks, it is assumed that the prediction error follows a Gaussian distribution, i.e., p(y|x,θ)~N(f(x;θ), σ²), where f(x;θ) is the deterministic output of the Bayesian neural network and σ is the noise variance. 3. According to Bayes' theorem, the posterior distribution is proportional to the product of the prior distribution and the likelihood function, i.e., p(θ|D)∝p(D|θ)p(θ), which represents the uncertainty of the weights after training. The prediction distribution of a Bayesian neural network is obtained by integrating the posterior distribution of the weights: p(y|x,D) = ∫p(y|x,θ)p(θ|D)dθ(1) Variational inference is used to approximate the posterior distribution, constructing a tractable approximate distribution q(θ). By minimizing the Kullback-Leibler divergence between q(θ) and the true posterior p(θ|D), the approximate distribution q(θ) approximates the posterior distribution p(θ|D) as closely as possible. The Kullback-Leibler divergence (KLD), also known as relative entropy, is used to quantify the difference between two probability distributions and is defined as follows: KL(q(θ)∥p(θ|D))=∫q(θ)log[q(θ) / p(θ|D)]dθ(2) Through Bayes' theorem transformation, minimizing the KL divergence is equivalent to maximizing the Evidence Lower Bound (ELBO). In actual training, the total loss function of the Bayesian neural network combines the mean-square error (MSE) of the regression task with the KL divergence term, balancing prediction accuracy with the accuracy of uncertainty quantification. The total loss function is defined as follows: L(θ)=L+KL(q(θ)∥p(θ))(3) Where L is the mean squared error loss, which is the error between the aerodynamic load prediction result of the Bayesian neural network on the input data and the true label: (4) in, It is a Bayesian data neural network for input data The aerodynamic load prediction results N represents the true labels, and N is the number of training samples in the current round.

[0034] After pre-training the Bayesian neural network using low-fidelity training data, the resulting pre-trained model can be applied to scenarios requiring low-precision prediction, such as the development phase of eVTOL (eVTOL) systems, where the pre-trained model can predict the aerodynamic loads on the rotor or wing. Understandably, pre-training the Bayesian neural network using low-fidelity training data stops when a first preset stopping condition is met, resulting in the pre-trained model.

[0035] In some embodiments, the first preset stopping condition is: after pre-training, the prediction accuracy of the Bayesian neural network reaches a preset accuracy and the proportion of confidence intervals containing true values ​​is close to a preset proportion threshold, or the number of iterations reaches a preset number of iterations. After the pre-training of the Bayesian neural network based on low-fidelity training data is completed, step S104 is executed.

[0036] Step S104: Based on the medium-fidelity training data and the high-fidelity training data, transfer learning is performed on the pre-trained model to obtain a multi-fidelity proxy model. The multi-fidelity proxy model predicts aerodynamic load data in the digital twin system based on the actual flight parameters of the electric vertical take-off and landing aircraft.

[0037] In some embodiments, by performing transfer learning on the pre-trained model based on the medium-fidelity training data and the high-fidelity training data, it is possible to reuse the general basic features learned by the pre-trained model, adapt high-order features only for the target task, and balance the modeling efficiency and prediction accuracy of the surrogate model. This allows for customized modifications throughout the entire lifecycle of the eVTOL, eliminating the need for repeated modeling in later stages, thus improving the adaptability of the surrogate model and reducing data utilization. The challenge lies in effectively integrating data at different levels to train the surrogate model at different stages. Transfer learning on the pre-trained model based on medium-fidelity and high-fidelity training data achieves the fusion of training data from multiple sources with different fidelities, enabling the multi-fidelity surrogate model to comprehensively cover the dynamic responses of the aircraft under various operating conditions and attitudes in complex environments.

[0038] In some embodiments, performing transfer learning on the pre-trained model based on the medium-fidelity training data and the high-fidelity training data may include the following steps S1031 and S1032.

[0039] Step S1031: Freeze the weight parameters of the bottom N layers in the pre-trained model, and fine-tune the weight parameters of the top M layers based on the mid-fidelity training data to obtain the intermediate proxy model.

[0040] It is understandable that the bottom N layers include the input layer and N-1 hidden layers connected sequentially after the input layer, while the top M layers are the output layer and M-1 hidden layers connected sequentially before the output layer. The bottom N layers and the top M layers together constitute the entirety of the Bayesian neural network model. After pre-training, by freezing the weight parameters of the bottom N layers in the pre-trained model, the weight parameters of the bottom N layers are no longer updated during the fine-tuning stage using the aforementioned high-fidelity training data, thus preventing the destruction of general features.

[0041] For example, such as Figure 2A As shown, the Bayesian neural network model used has a total of 5 layers: layer 1 is the input layer, layers 2-4 are hidden layers, and layer 5 is the output layer. For example, if N=2 and M=3, the weight parameters of layers 1-2 in the pre-trained model are frozen, and the weight parameters of layers 3-5 in the pre-trained model are fine-tuned based on the mid-fidelity training data. It should be understood that the topology of the Bayesian neural network model used in this embodiment of the invention can have far more than 5 layers.

[0042] Step S1032: Freeze the weight parameters of the bottom N layers and at least one intermediate layer in the intermediate proxy model, and fine-tune the weight parameters of the remaining intermediate layers and the top layer in the intermediate proxy model based on the high-fidelity training data to obtain the multi-fidelity proxy model.

[0043] You can fine-tune only the weight parameters of the top layer based on high-fidelity training data, or you can fine-tune the weight parameters of the top layer (output layer) and at least one intermediate layer other than the bottom N layers based on high-fidelity training data.

[0044] like Figure 2A As shown, continuing with the example of an intermediate proxy model topology including an input layer (layer 1), hidden layer 1 (layer 2), hidden layer 2 (layer 3), hidden layer 3 (layer 4), and output layer (layer 5), the intermediate proxy model is obtained after pre-training with low-fidelity training data and fine-tuning with medium-fidelity training data. Setting P=4, which freezes the bottom four layers: corresponding to layers 1 (input layer), 2 (hidden layer 1), 3 (hidden layer 2), and 4 (hidden layer 2) of the intermediate proxy model, the weight parameters of layer 5 (output layer) are then fine-tuned based on high-fidelity training data.

[0045] Understandably, when fine-tuning is performed based on high-fidelity training data until the second preset stopping condition is met, the fine-tuning of the weight parameters of the remaining intermediate layers and the top layer, excluding the bottom N layers and at least one intermediate layer, is stopped.

[0046] In some embodiments, a dedicated MF-Baynet library is developed to support the training, testing, and deployment of Bayesian neural networks (BNNs) with different fidelity training data. The intermediate proxy model fine-tuned in step S1031 can be applied to other stages after the R&D stage, such as the certification stage. The multi-fidelity proxy model fine-tuned in step S1032 can be applied to the operation stage.

[0047] In some embodiments, when fine-tuning weight parameters based on high-fidelity training data, the number of intermediate layers to be selected for weight parameter fine-tuning can be determined based on the model performance of the intermediate surrogate model after weight parameter fine-tuning based on medium-fidelity training data. Specifically, the intermediate surrogate model can be performance evaluated to obtain its model performance; based on the model performance of the intermediate surrogate model, the intermediate layers that need fine-tuning can be determined.

[0048] In some embodiments, the performance of the intermediate proxy model can be evaluated from one or more evaluation dimensions using a validation set partitioned from the intermediate-fidelity training data. The evaluation dimensions include one or more of the following: fitting accuracy, generalization ability, feature contribution, and uncertainty quantification reliability.

[0049] Fitting accuracy can be characterized by the root mean square error (RMSE). The validation set, derived from the mid-fidelity training data, is input into the intermediate surrogate model. The RMSE is calculated based on the dynamic load prediction results and true labels of each sample in the validation set, measuring the deviation between the intermediate surrogate model's dynamic load predictions and the true labels. Generalization ability can be characterized by the generalization error, which is the difference between the RMSE of the intermediate surrogate model on the validation set and the RMSE on the training set. A smaller difference indicates stronger generalization ability. Feature contribution can be represented by gradient-weighted class activation mapping (Grad-CAM) weights. The gradient contribution of each intermediate layer's output feature to the dynamic load prediction results output by the intermediate surrogate model is calculated. A higher contribution value indicates a more critical feature, and the feature contribution value can pinpoint which intermediate layer has learned the core patterns of the mid-fidelity data. Uncertainty quantification reliability can be characterized by Prediction Interval Coverage Probability (PICP), which is the proportion of true labels falling within the confidence interval of the intermediate surrogate model output. The ideal value is close to 95%, which measures whether the variance of the intermediate surrogate model output is reasonable.

[0050] By fine-tuning in steps S1031-S1032, only the weights of the later layers of the network are updated, preserving the general knowledge from the pre-training stage. Simultaneously, the vortex flow field characteristics of medium-fidelity data and the viscous effects and compressibility details of high-fidelity data are incorporated, improving the model's prediction accuracy for complex flow fields. Transfer learning is used to integrate the complementary advantages of low, medium, and high-fidelity training data, achieving a balance between accuracy and efficiency through phased training.

[0051] In some embodiments, after fine-tuning the multi-fidelity proxy model using high-fidelity training data to obtain the model performance, the multi-fidelity proxy model is evaluated to obtain the model performance. If the model performance is lower than a preset performance threshold, an iterative search is performed based on a Bayesian optimization strategy to determine the next optimal hyperparameter combination. Based on this optimal hyperparameter combination, as well as the low-fidelity, medium-fidelity, and high-fidelity training data, the Bayesian neural network model is re-pre-trained and transfer-learned until a preset number of iterations or convergence conditions are reached.

[0052] In some embodiments, when evaluating the performance of the multi-fidelity proxy model, a validation set partitioned from the high-fidelity training data can be used to evaluate the performance of the multi-fidelity proxy model. Evaluating the performance of the multi-fidelity proxy model using a validation set partitioned from the high-fidelity training data may include: inputting each sample to be predicted from the validation set partitioned from the high-fidelity training data into the multi-fidelity proxy model to obtain the probability distribution parameters of the aerodynamic load prediction values ​​for each sample to be predicted; for each sample to be predicted, sampling the probability distribution parameters of the aerodynamic load prediction values ​​corresponding to the sample to be predicted by the multi-fidelity proxy model using Monte Carlo sampling to obtain multiple aerodynamic load prediction values ​​for the sample to be predicted; obtaining the aerodynamic load prediction result and confidence interval for the sample to be predicted based on the multiple aerodynamic load prediction values; and determining at least one performance evaluation index of the multi-fidelity proxy model's model performance based on the aerodynamic load prediction results and / or confidence intervals corresponding to each sample to be predicted in the validation set. It is understood that the performance evaluation metrics of the multi-fidelity proxy model include one or more dimensions such as fitting accuracy, generalization ability, feature contribution, and uncertainty quantification reliability. The process of obtaining the various performance evaluation metrics of the multi-fidelity proxy model is similar to the implementation method of obtaining the model performance of the intermediate proxy model, except that the source of the validation set used is different. For the sake of brevity, it will not be described in detail here.

[0053] Understandably, multiple samples are extracted from the Bayesian neural network through Monte Carlo sampling to generate multiple aerodynamic load prediction values. Then, the aerodynamic load mean and standard deviation are calculated. The standard deviation is used to distinguish between cognitive uncertainty and accidental uncertainty, providing a quantitative basis for the reliability of the dynamic load prediction results. The specific process is as follows: input the sample to be predicted into a multi-fidelity surrogate model, perform M forward propagations through Monte Carlo sampling to generate M aerodynamic load prediction values, calculate the aerodynamic load prediction mean μ and prediction standard deviation σ based on the M aerodynamic load prediction values, and output the aerodynamic load prediction mean μ and confidence interval [μ-1.96σ, μ+1.96σ].

[0054] In some embodiments, determining the next optimal hyperparameter combination based on the performance metrics and the Bayesian optimization (BO) strategy may include: constructing a probabilistic surrogate model based on the initial multiple hyperparameter combinations and the model performance under each initial hyperparameter combination, wherein the probabilistic surrogate model is the relationship between hyperparameters and model performance; selecting candidate hyperparameter combinations through a collection function, and evaluating the model performance under the candidate hyperparameter combinations through an objective function and the constructed probabilistic surrogate model, until the maximum number of iterations is reached or the model performance reaches a performance convergence threshold, thereby obtaining the next optimal hyperparameter combination.

[0055] In some embodiments, the probabilistic surrogate model may be established through a Gaussian process (GP), and the acquisition function may be the expected improvement (EI), the probability of improvement (PI), or the upper confidence interval (UCB).

[0056] The predefined hyperparameter space includes the type and value range of each hyperparameter (e.g., learning rate, optimizer type, number of hidden layers, number of neurons, activation function). Multiple initial hyperparameter combinations are sampled from the predefined hyperparameter space (e.g., 5 to 20 initial hyperparameter combinations are randomly sampled). The model performance of the surrogate model under each initial hyperparameter combination is obtained. Then, based on each initial hyperparameter combination and the model performance of the surrogate model under each initial hyperparameter combination, a probabilistic surrogate model is established using a Gaussian process. The established probabilistic surrogate model represents the probabilistic mapping relationship between hyperparameters and model performance.

[0057] In some embodiments, for each initial hyperparameter combination selected in a predefined hyperparameter space, a Bayesian neural network model is trained (including pre-training and transfer learning), and the RMSE of the trained surrogate model under each initial hyperparameter combination is calculated to obtain the initial sample set D0{(x1,f(x1)),(x2,f(x2)),...(x i , f(x i ))}. Construct a probabilistic proxy model based on the initial sample set. The objective function is f(x) = RMSE(x), where x is a combination of hyperparameters, and the optimization objective is to minimize RMSE.

[0058] Taking the expected value improvement (EVI) of the acquisition function as an example, the EI value is calculated for all candidate hyperparameter combinations in the hyperparameter space. The combination with the largest EI value is the next sampling point for the hyperparameter combination. The value of the corresponding objective function at this sampling point is calculated to obtain the new RMSE value. The new sample (x) is then used to calculate the next sampling point. i+1 , f(x i+1 Add samples to the sample set to obtain a new sample set, and update the probabilistic surrogate model with the new sample set. Then, use the acquisition function to select the next sampling point for the hyperparameter combination. Repeat the above steps until the stopping condition is met. After iteration, select the sampling point of the hyperparameter combination that optimizes the objective function from all the sampling points of all hyperparameter combinations, and use this as the next optimal hyperparameter combination. Use the acquisition function (e.g., to improve EI) to guide the next hyperparameter selection, avoiding blind searching.

[0059] Hyperparameters are crucial to the performance of multifidelity transfer. This invention employs a Bayesian optimization strategy instead of traditional grid search, iteratively exploring the hyperparameter space to efficiently find the optimal hyperparameter combination. It utilizes the probabilistic characteristics of Bayesian neural network models to quantify prediction uncertainty, addressing the problem that traditional models cannot assess confidence levels.

[0060] According to one or more embodiments of the present invention, based on massive amounts of leaf element method data, a pre-trained BNNs model is used to capture the general laws of aerodynamic performance of lift components, providing a stable initial model for subsequent transfer learning; the bottom weights of the pre-trained model are frozen, and only the top weights are trained, so that the model can adapt to the three-dimensional flow field effect reflected by vorticity particle method data while retaining the general laws; the number of frozen layers is reduced, allowing some intermediate layers to participate in training, so that the model can accurately fit the Reynolds average method, the model performance is improved through Bayesian optimization, and the probabilistic characteristics of the BNNs model are used to quantify the prediction uncertainty.

[0061] Figure 2B The prediction results of the multi-fidelity proxy model and three existing proxy models are compared. The multi-fidelity proxy model constructed according to the embodiments of the present invention (…) Figure 2B The BNN-TL shown was compared with the proxy models constructed in the other three ways in a laboratory setting. The comparison results are shown in [link to comparison]. Figure 2B As shown. Figure 2B This paper presents the prediction results of the multi-fidelity surrogate model constructed in this embodiment of the invention (the prediction result corresponds to the blue curve) and the surrogate models constructed in three different ways, and compares them with real experimental data (validation data). The horizontal axis represents rotational speed, which is the input variable. The vertical axis represents tension, which is the output variable and the target predicted by the surrogate model. The other three methods are: a surrogate model obtained by training a Bayesian neural network (BNN-LF) based on low-fidelity data, with the prediction result corresponding to the pink curve; a surrogate model obtained by training a Bayesian neural network (BNN-MF) based on medium-fidelity data, with the prediction result corresponding to the green curve; and a surrogate model obtained by co-training with Kriging (COKriging), with the prediction result corresponding to the yellow curve. It can be seen that the prediction result of the multi-fidelity surrogate model constructed in this embodiment of the invention is closer to the real experimental results than the surrogate models constructed in the other three ways.

[0062] Based on the same inventive concept, a data processing method for an electric vertical takeoff and landing aircraft is provided. Figure 3 A flowchart of a data processing method for an electric vertical takeoff and landing aircraft according to some embodiments of the present invention is shown. Figure 4 A schematic diagram of the architecture of a data processing method for an electric vertical takeoff and landing aircraft according to some embodiments of the present invention is shown. For example... Figure 3 As shown, the data processing method includes the following steps S301 to S304: S301: Acquire manipulation control signals and atmospheric disturbance signals, wherein the atmospheric disturbance signals are generated based on environmental parameters; S302: Input the control signals and atmospheric disturbance signals into the flight dynamics module of the digital twin system of the electric vertical takeoff and landing aircraft. The flight dynamics module deploys a multi-fidelity proxy model, which is obtained by the proxy model construction method described in any of the preceding embodiments. S303: The flight mechanics module determines the aircraft state parameters based on the control signals, and the multi-fidelity proxy model predicts the aerodynamic load data of the electric vertical takeoff and landing aircraft based on the aircraft state parameters, the atmospheric disturbance signals and the structural parameters of the electric vertical takeoff and landing aircraft. S304: The flight mechanics module simulates the target response signal of the electric vertical takeoff and landing aircraft based on the aerodynamic load data.

[0063] like Figure 4 As shown, flight environment data serves as external disturbance input and is transmitted to the atmospheric disturbance model. The control signal layer integrates input signals, generates unified control commands, and transmits them to the flight dynamics module. The flight dynamics module is the core calculation unit for the motion state of the electric vertical takeoff and landing (EVT) aircraft. It includes a weight and balance submodule, a landing gear submodule, a multi-fidelity surrogate model, and a propulsion system model. The weight and balance submodule calculates basic parameters such as the aircraft's mass distribution and center of gravity; the landing gear module simulates the mechanical characteristics of the landing gear (such as forces during takeoff and landing); the multi-fidelity surrogate model predicts core aerodynamic loads such as aerodynamic lift, drag, and torque of lifting components like the rotor; and the propulsion system model simulates power output such as engine thrust. It is understood that the propulsion system model can also be constructed based on the aforementioned surrogate model construction method, with only the original training data differing. The original training data is the engine's test data, which will not be elaborated upon here for the sake of brevity. Finally, the motion state of the EVT aircraft is obtained through the calculation of resultant force and resultant torque, and the calculation of the six-degree-of-freedom motion equations. The atmospheric disturbance model receives flight environment data to simulate environmental disturbances such as wind and turbulence, outputs atmospheric disturbance signals, and transmits them to the flight dynamics module for calculations. Through these calculations, the flight dynamics module ultimately outputs the target response signal, which represents the aircraft's target motion state (such as attitude, velocity, and position).

[0064] Figure 5 A system architecture diagram of a digital twin system for an electric vertical takeoff and landing aircraft according to some embodiments of the present invention is shown. For example... Figure 5As shown, the digital twin system of the electric vertical takeoff and landing aircraft includes multiple independently configured functional modules. The functional modules communicate with each other through standardized interfaces. At least one proxy model constructed in the above embodiments is deployed in at least one functional module.

[0065] In some embodiments, a digital twin system for an electric vertical takeoff and landing (EVTOL) aircraft provided by the present invention may include: an independently configured flight mechanics module and an aeroelasticity module, wherein the flight mechanics module and the aeroelasticity module communicate with each other via a standardized interface; and a multi-fidelity proxy model, obtained based on the proxy model construction method described in any embodiment of the first aspect, wherein the multi-fidelity proxy model is deployed in the flight mechanics module of the digital twin system, and is used to predict the aerodynamic load data of the EVTOL aircraft based on the flight environment data, structural data, and flight state data of the EVTOL aircraft, so that the flight mechanics module can simulate the target response signal of the EVTOL aircraft based on the aerodynamic load data.

[0066] Understandably, the flight mechanics module can be a flight mechanics solver module built on MATLAB-SIMULINK. It is used for trim calculations and specified trajectory tracking in eVTOL, incorporating the lowest-order elastic modes of the structure to account for both rigid body motion and elastic deformation effects. The core design is based on the six-degree-of-freedom equations of motion of a rigid body, using quaternions to describe the aircraft's attitude (avoiding the singularity problem of Euler angles at large attitude angles). Simultaneously, it combines aerodynamic load data output from multi-fidelity surrogate models to improve the simulation realism of the flight mechanics module, adapting it to complex flight conditions such as eVTOL vertical takeoff and landing and hovering.

[0067] The aeroelasticity module analyzes static deformation, linear stability, stability margin, and dynamic response by linearizing the system in the time or frequency domain, or by treating it as a fully nonlinear system, covering the influence of key parameters such as rotor speed and acceleration variations. It analyzes the aeroelastic response of the flexible structures (wings and rotors) of the eVTOL to mitigate risks such as flutter, divergence, and whirl flutter. The aeroelasticity module includes a finite element method (FEM) wing model, key rigid body modes, and rotor elasticity model.

[0068] In some embodiments, a proxy model can also be deployed in the aeroelastic module. The construction method is similar to that of the multi-fidelity proxy model described above. The aeroelastic module improves the accuracy of coupling analysis by integrating the modeling advantages of the mid-fidelity VPM method for vortex flow fields with the viscosity effect characteristics of high-fidelity data.

[0069] Existing eVTOL simulation systems have independent modules for flight mechanics and aeroelasticity, resulting in ineffective data exchange and difficulty in reflecting multiphysics coupling characteristics. Furthermore, the system architecture lacks flexibility, leading to high coupling and low efficiency when developed in parallel by multidisciplinary teams (aerodynamics, structure, control). In contrast, the digital twin system for electric vertical takeoff and landing aircraft provided in this invention employs a modular design, enabling high flexibility and scalability, facilitating multidisciplinary team collaboration and functional upgrades.

[0070] It should be noted that the difference between the flight mechanics module and the aeroelasticity module in this embodiment of the invention and the prior art lies in the use of a multi-fidelity surrogate model (such as a multi-fidelity surrogate model for aerodynamic load prediction) instead of a traditional physical model, in order to overcome the problems of insufficient calculation accuracy or low efficiency of physical models in the prior art. Figure 5 The calculation process and data transfer of other models can be found in the relevant technologies, and will not be elaborated here for the sake of brevity.

[0071] Based on the same inventive concept, embodiments of the present invention provide a proxy model construction device, wherein the proxy model is deployed in the digital twin system of an electric vertical takeoff and landing aircraft. Figure 6 A schematic diagram of the proxy model construction apparatus according to some embodiments of the present invention is shown. For example... Figure 6 As shown, the proxy model construction device provided in this embodiment of the invention includes: a raw data acquisition unit 601, used to acquire raw training data, wherein each training sample in the raw training data includes flight state parameters, flight environment parameters, and structural parameters of an electric vertical takeoff and landing (EVTOL) aircraft; a training data preparation unit 602, used to process the raw training data based on three different fidelity fluid dynamics numerical simulation methods to obtain low-fidelity training data, medium-fidelity training data, and high-fidelity training data, wherein the training samples in each fidelity training data include aerodynamic load data simulated based on the corresponding fidelity fluid dynamics numerical simulation method; a pre-training unit 603, used to pre-train a Bayesian neural network model based on the low-fidelity training data to obtain a pre-trained model; and a transfer learning unit 604, used to perform transfer learning on the pre-trained model based on the medium-fidelity training data and the high-fidelity training data to obtain a multi-fidelity proxy model, wherein the multi-fidelity proxy model is used in the digital twin system to predict aerodynamic load data based on the actual flight parameters of the EVTOL aircraft.

[0072] In some embodiments, the training data preparation unit 602 is configured to: process the original training data using the surface element method to obtain the low-fidelity training data; process the original training data using the eddy particle method to obtain the medium-fidelity training data; and process the original training data using the turbulence simulation method to obtain the high-fidelity training data.

[0073] In some embodiments, the transfer learning unit 604 is configured to: freeze the weight parameters of the bottom N layers in the pre-trained model, fine-tune the weight parameters of the top M layers based on the mid-fidelity training data to obtain an intermediate proxy model; freeze the weight parameters of the bottom N layers and at least one intermediate layer in the intermediate proxy model, fine-tune the weight parameters of the remaining intermediate layers and the top layer in the intermediate proxy model based on the high-fidelity training data to obtain the multi-fidelity proxy model.

[0074] In some embodiments, the proxy model construction apparatus further includes a freeze layer determination unit, configured to: perform performance evaluation on the intermediate proxy model to obtain the model performance of the intermediate proxy model; and determine the at least one intermediate layer that needs to be frozen based on the model performance of the intermediate proxy model.

[0075] In some embodiments, the proxy model construction apparatus further includes: an evaluation unit, configured to evaluate the performance of the multi-fidelity proxy model to obtain the model performance of the multi-fidelity proxy model; an optimization unit, configured to perform iterative search based on a Bayesian optimization strategy to determine the next optimal hyperparameter combination if the model performance of the multi-fidelity proxy model is lower than a preset performance threshold; and a retraining unit, configured to retrain and transfer learn the Bayesian neural network model based on the next optimal hyperparameter combination, the low-fidelity training data, the medium-fidelity training data, and the high-fidelity training data, until a preset number of iterations or convergence condition is reached.

[0076] In some embodiments, the optimization unit is configured to: construct a probabilistic surrogate model based on an initial set of multiple hyperparameter combinations and the estimated model performance under each initial hyperparameter combination, wherein the probabilistic surrogate model is a relationship between hyperparameters and model performance; select candidate hyperparameter combinations through a collection function, and evaluate the model performance under the candidate hyperparameter combinations through an objective function and the constructed probabilistic surrogate model, until the maximum number of iterations is reached or the estimated model performance reaches a performance convergence threshold, thereby obtaining the next set of optimal hyperparameter combinations.

[0077] In some embodiments, the frozen layer determination unit is configured to: input each sample to be predicted in the validation set into the multi-fidelity surrogate model to obtain the probability distribution parameters of the aerodynamic load prediction value of each sample to be predicted, wherein the validation set is partitioned from the high-fidelity training data; for each sample to be predicted, sample the probability distribution parameters of the aerodynamic load prediction value of the sample to be predicted by the multi-fidelity surrogate model using Monte Carlo sampling to obtain multiple aerodynamic load prediction values ​​for the sample to be predicted; obtain the aerodynamic load prediction result and confidence interval for the sample to be predicted based on the multiple aerodynamic load prediction values; and determine the model performance of the multi-fidelity surrogate model according to the aerodynamic load prediction result and / or confidence interval corresponding to each sample to be predicted in the validation set.

[0078] This device is for implementing the aforementioned proxy model construction method. More implementation details of this device can be found in the aforementioned proxy model construction method. For the sake of brevity, these details will not be repeated here.

[0079] Based on the same inventive concept, embodiments of the present invention provide a computer device. Figure 7 Structural diagrams of a computer device according to some embodiments of the present invention are shown. For example... Figure 7 As shown, it includes: a memory 701, a processor 702, and code stored in the memory and executable on the processor 702. When the processor 702 executes the code, it implements the proxy model construction method described in any of the foregoing embodiments, or the data processing method of the electric vertical take-off and landing aircraft described in any of the foregoing embodiments.

[0080] Among them, Figure 7 In this document, a bus architecture (represented by bus 700) is used. Bus 700 may include any number of interconnected buses and bridges, linking various circuits including one or more processors represented by processor 702 and memory represented by memory 701. Bus 700 may also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 705 provides an interface between bus 700 and receiver 703 and transmitter 704. Receiver 703 and transmitter 704 may be the same element, i.e., a transceiver, providing a unit for communicating with various other devices over a transmission medium. Processor 702 is responsible for managing bus 700 and general processing, while memory 701 can be used to store data used by processor 702 during operation.

[0081] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable code.

[0082] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer instructions. These computer instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0083] These computer instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0084] These computer instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0085] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0086] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for constructing a proxy model, wherein the proxy model is deployed in the digital twin system of an electric vertical takeoff and landing aircraft, characterized in that, The method includes: Obtain raw training data, wherein each training sample in the raw training data includes flight state parameters, flight environment parameters, and structural parameters of the electric vertical takeoff and landing aircraft; The original training data is processed using three different fluid dynamics numerical simulation methods with varying fidelity to obtain low-fidelity training data, medium-fidelity training data, and high-fidelity training data. The training samples in each type of training data include aerodynamic load data simulated by the corresponding fluid dynamics numerical simulation method. Based on the low-fidelity training data, the Bayesian neural network model is pre-trained to obtain a pre-trained model. Based on the medium-fidelity training data and the high-fidelity training data, the pre-trained model is transferred to obtain a multi-fidelity proxy model. The multi-fidelity proxy model is used in the digital twin system to predict aerodynamic load data based on the actual flight parameters of the electric vertical take-off and landing aircraft.

2. The proxy model construction method as described in claim 1, characterized in that, The fluid dynamics numerical simulation method based on three different fidelity levels processes the original training data to obtain low-fidelity training data, medium-fidelity training data, and high-fidelity training data, including: The original training data is processed using the facet method to obtain the low-fidelity training data; The original training data is processed using the vortex particle method to obtain the medium-fidelity training data. The original training data is processed using turbulence simulation methods to obtain the high-fidelity training data.

3. The proxy model construction method as described in claim 1, characterized in that, The step of performing transfer learning on the pre-trained model based on the medium-fidelity training data and the high-fidelity training data to obtain a multi-fidelity agent model includes: Freeze the weight parameters of the bottom N layers in the pre-trained model, and fine-tune the weight parameters of the top M layers based on the mid-fidelity training data to obtain the intermediate proxy model; Freeze the weight parameters of the bottom N layers and at least one intermediate layer in the intermediate proxy model, and fine-tune the weight parameters of the remaining intermediate layers and the top layer in the intermediate proxy model based on the high-fidelity training data to obtain the multi-fidelity proxy model.

4. The proxy model construction method as described in claim 3, characterized in that, Also includes: The intermediate proxy model is evaluated to obtain its performance. Based on the model performance of the intermediate proxy model, determine at least one intermediate layer that needs to be frozen.

5. The proxy model construction method as described in claim 3, characterized in that, Also includes: The performance of the multi-fidelity proxy model is evaluated to obtain the model performance of the multi-fidelity proxy model; If the performance of the multi-fidelity proxy model is lower than a preset performance threshold, an iterative search is performed based on a Bayesian optimization strategy to determine the next optimal combination of hyperparameters. Based on the next optimal combination of hyperparameters, as well as the low-fidelity training data, the medium-fidelity training data, and the high-fidelity training data, the Bayesian neural network model is retrained and transfer learned until the preset number of iterations or convergence conditions are reached.

6. The proxy model construction method as described in claim 5, characterized in that, The iterative search based on the Bayesian optimization strategy to determine the next optimal combination of hyperparameters includes: Based on the initial multiple hyperparameter combinations and the estimated model performance under each initial hyperparameter combination, a probabilistic surrogate model is constructed, wherein the probabilistic surrogate model is the relationship between hyperparameters and model performance. Candidate hyperparameter combinations are selected by the acquisition function, and the model performance under the candidate hyperparameter combinations is evaluated by the objective function and the constructed probabilistic surrogate model until the maximum number of iterations is reached or the estimated model performance reaches the performance convergence threshold, thus obtaining the next set of optimal hyperparameter combinations.

7. The proxy model construction method as described in claim 5, characterized in that, The step of evaluating the performance of the multi-fidelity proxy model to obtain its performance includes: Each sample to be predicted in the validation set is input into the multi-fidelity proxy model to obtain the probability distribution parameters of the predicted aerodynamic load value of each sample to be predicted. The validation set is divided from the high-fidelity training data. For each of the samples to be predicted, the probability distribution parameters of the aerodynamic load prediction values ​​of the sample to be predicted are sampled by Monte Carlo sampling for the multi-fidelity surrogate model to obtain multiple aerodynamic load prediction values ​​for the sample to be predicted. Based on the multiple aerodynamic load prediction values, the aerodynamic load prediction result and confidence interval for the sample to be predicted are obtained. The model performance of the multi-fidelity surrogate model is determined based on the aerodynamic load prediction results and / or confidence intervals corresponding to each sample to be predicted in the validation set.

8. A data processing method for an electric vertical takeoff and landing aircraft, characterized in that, include: Acquire manipulation control signals and atmospheric disturbance signals, wherein the atmospheric disturbance signals are generated based on environmental parameters; The control signals and atmospheric disturbance signals are input to the flight dynamics module of the digital twin system of the electric vertical takeoff and landing aircraft. The flight dynamics module deploys a multi-fidelity proxy model, which is obtained based on the proxy model construction method described in any one of claims 1-7. The flight mechanics module determines the aircraft state parameters based on the control signals, and the multi-fidelity proxy model predicts the aerodynamic load data of the electric vertical takeoff and landing aircraft based on the aircraft state parameters, the atmospheric disturbance signals, and the structural parameters of the electric vertical takeoff and landing aircraft. The flight mechanics module simulates the target response signal of the electric vertical takeoff and landing aircraft based on the aerodynamic load data.

9. A proxy model construction apparatus, wherein the proxy model is deployed in a digital twin system of an electric vertical takeoff and landing aircraft, characterized in that, The proxy model construction device includes: The raw data acquisition unit is used to acquire raw training data, wherein each training sample in the raw training data includes flight state parameters, flight environment parameters, and structural parameters of the electric vertical takeoff and landing aircraft. The training data preparation unit is used to process the original training data based on three different fluid dynamics numerical simulation methods with different fidelity, so as to obtain low-fidelity training data, medium-fidelity training data and high-fidelity training data. The training samples in each fidelity training data include aerodynamic load data simulated based on the corresponding fluid dynamics numerical simulation method. The pre-training unit is used to pre-train the Bayesian neural network model based on the low-fidelity training data to obtain a pre-trained model. The transfer learning unit is used to perform transfer learning on the pre-trained model based on the medium-fidelity training data and the high-fidelity training data to obtain a multi-fidelity proxy model. The multi-fidelity proxy model is used in the digital twin system to predict aerodynamic load data based on the actual flight parameters of the electric vertical take-off and landing aircraft.

10. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the proxy model construction method of any one of claims 1-7, or the data processing method of the electric vertical take-off and landing aircraft of claim 8.

11. A digital twin system for an electric vertical takeoff and landing aircraft, characterized in that, include; The flight mechanics module and the aeroelasticity module are independently configured, and the flight mechanics module and the aeroelasticity module communicate with each other through a standardized interface; A multi-fidelity proxy model, obtained based on any of the proxy model construction methods in claims 1-7, is deployed in the flight dynamics module of the digital twin system. It is used to predict the aerodynamic load data of the electric vertical takeoff and landing aircraft based on the flight environment data, structural data, and flight state data of the electric vertical takeoff and landing aircraft, so that the flight dynamics module can simulate the target response signal of the electric vertical takeoff and landing aircraft based on the aerodynamic load data.

Images