Helicopter take-off and landing dynamic response prediction method and system based on physical constraints

Abstract

The present application belongs to the technical field of computer simulation and aircraft dynamics, and discloses a helicopter take-off and landing dynamics response prediction method and system based on physical constraints. The present application obtains a training data set, the input of which is the coordinates of the target environmental wind field area and the three-dimensional space transient flow field data, and the output of which is the corresponding helicopter dynamics response data; a long short-term memory neural network model is constructed; a hybrid loss function composed of a data loss term and a physical constraint loss term is designed, the physical constraint loss term at least including a residual based on a pitch angle balance equation, used to constrain the physical consistency between the rotor total pitch, lateral cyclic pitch and longitudinal cyclic pitch predicted by the model; the model is trained using the training data set and the hybrid loss function; the target environmental wind field data to be predicted is input into the trained model, and the helicopter take-off and landing dynamics response prediction result is output, thereby improving the accuracy and physical consistency in the neural network model training process.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computer simulation and aircraft dynamics, specifically involving a method and system for predicting the dynamic response of helicopter take-off and landing based on physical constraints, which is particularly suitable for rapid prediction of the take-off and landing status of helicopters in complex wind fields. Background Technology

[0002] The dynamic response of a helicopter during takeoff and landing is strongly coupled with its own aerodynamic characteristics, control inputs, and the external wind field. Currently, there are three main methods for predicting the dynamic response of this process: High-fidelity numerical computation (CFD / CSD coupled simulation): This method establishes an accurate physical model based on computational fluid dynamics (CFD) and computational structural dynamics (CSD). Although it has high accuracy, it consumes huge computational resources, and a single simulation can take several hours or even days. It cannot meet the needs of real-time or rapid prediction, such as in flight simulators, real-time flight control systems or rapid mission planning.

[0003] Traditional flight dynamics methods (blade element theory / inflow model): This method is based on a simplified physical model and reduces the order of complex flows through quasi-steady assumptions and low-order inflow models. These methods are computationally efficient and suitable for flight dynamics modeling and control analysis, but are limited by model assumptions and cannot accurately describe complex unsteady flows, exhibiting limited accuracy under complex operating conditions.

[0004] Pure data-driven models (such as traditional LSTM, ANN, etc.): These methods rely entirely on historical flight data to train black-box models. Although they are computationally fast, their predictions lack physical consistency. In conditions not covered by the training data (such as extreme wind fields), the model may produce predictions that violate physical laws (such as non-conservation of energy and momentum), resulting in unreliable predictions and poor generalization ability.

[0005] Existing research is mostly based on physical modeling frameworks, employing multibody dynamics methods to model complex rotor blade configurations, or using mathematical models to solve nonlinear equations to correct parameter errors in the dynamic model. These methods typically rely on an explicit physical modeling process, establishing structural dynamics and aerodynamic models to describe the helicopter's dynamic behavior. In complex flow environments (such as wake interference from ship decks), the modeling accuracy of these methods depends heavily on the completeness and parameter accuracy of the physical model. This invention, while preserving physical constraints, introduces a data-driven mechanism to achieve efficient modeling of complex dynamic responses, thus improving the limitations of traditional physical models in balancing accuracy and efficiency to some extent.

[0006] Existing data-driven methods for predicting helicopter dynamic parameters mainly focus on dynamic parameter identification or control variable prediction. Their core objective is usually to modify existing models, predicting the control variable at the next moment from the helicopter's current control variable; or to calculate the helicopter's dynamic response using traditional dynamic equations based on predictions of ship motion and the external environment, rather than directly modeling the dynamic response under complex flow field conditions. Furthermore, while some methods introduce physical constraints, these constraints apply to the control response, which differs from the method of physical constraints based on helicopter flight dynamics used in this invention.

[0007] Therefore, existing methods still face a significant trade-off between computational efficiency and physical reliability: high-fidelity methods, while possessing good physical consistency, suffer from excessively high computational costs; low-order physical models, while computationally efficient, have limited accuracy; and purely data-driven methods, while capable of rapid prediction, lack physical constraints and reliable generalization capabilities. Thus, there is an urgent need in this field to develop a novel fusion method that, while ensuring computational efficiency, introduces physical constraint mechanisms to achieve efficient and physically consistent predictions of helicopter takeoff and landing dynamics.

[0008] Based on the above analysis, the problems and shortcomings of the existing technology are as follows: (1) The contradiction between computational efficiency and physical reliability is prominent: high-fidelity numerical computation consumes huge resources, and a single simulation takes several hours to several days, which cannot meet the needs of flight simulators, real-time flight control and other applications for rapid prediction; traditional flight dynamics methods are limited by quasi-steady assumptions and low-order inflow models, making it difficult to accurately describe complex unsteady flows, and their accuracy is limited under complex working conditions.

[0009] (2) Pure data-driven models lack physical constraints and have poor generalization ability: black box models that rely entirely on historical data lack physical consistency in their prediction results; under conditions where training data is not covered, predictions may be made that violate physical laws such as energy conservation and momentum conservation, resulting in unreliable results.

[0010] (3) Existing research has failed to effectively integrate physical mechanisms and data-driven methods: Physical modeling-based research is highly dependent on the completeness of the physical model and the accuracy of the parameters, and its accuracy is limited in complex flow environments. Existing data-driven methods are mostly focused on parameter identification or control quantity prediction. The physical constraints introduced are only for control response, rather than directly for helicopter dynamic response modeling under complex flow field conditions. It is necessary to predict the environment first and then calculate the response through traditional dynamic equations, which fails to achieve end-to-end physical-guided rapid prediction. Summary of the Invention

[0011] To overcome the problems existing in related technologies, the present invention discloses a method and system for predicting the dynamic response of helicopter takeoff and landing based on physical constraints. The technical solution is as follows: This invention is implemented as follows: a method for predicting the dynamic response of helicopter takeoff and landing based on physical constraints, comprising the following steps: S1. Obtain a training dataset, which includes training input data and training output data; wherein, the training input data is the coordinate data of the wind field area of ​​the target environment and the three-dimensional transient flow field data, and the training output data is the helicopter dynamic response data corresponding to the training input data; S2. Construct a long short-term memory neural network model, wherein the long short-term memory neural network model is used to predict the training output data based on the training input data; S3. Design a hybrid loss function, which is composed of a data loss term and a physical constraint loss term. The data loss term is used to measure the difference between the model prediction results and the actual output data. The physical constraint loss term includes at least the residual based on the pitch angle balance equation, which is used to constrain the physical consistency between the rotor collective pitch, lateral cyclic pitch and longitudinal cyclic pitch predicted by the model. S4. Train the long short-term memory neural network model using the training dataset and the hybrid loss function, and update the network parameters; S5. Input the target environment wind field data to be predicted into the trained long short-term memory neural network model, and output the corresponding helicopter take-off and landing dynamic response prediction results.

[0012] In step S1, the training dataset includes: Using the coordinate data of the target environment wind field region and the three-dimensional transient flow field data As the training input for the model, where, These represent the spatial coordinates of points within the wind field region. These represent the horizontal, vertical, and lateral wind speed components at the spatial coordinates of a point in the wind field region, respectively. Indicates the timestamp of the wind field area. Indicates the spatial point number. Indicates the total number of points in space; Using helicopter dynamic response data As the training output of the model, These represent the helicopter's pitch and roll angles, respectively. Indicates rotor collective pitch, Indicates lateral periodic pitch, Indicates longitudinal periodic pitch, This indicates the collective pitch of the tail rotor.

[0013] In step S1, before model training, the model is initially set up, including setting the physical parameters such as the blade root mounting angle. Blade torque Rotor blade azimuth angle Distance from the blade profile to the center of the rotor hub Rotor radius Tail rotor radius rotor advance ratio rotor cone angle Rotor equivalent induced velocity Rotor speed helicopter weight The physical parameters are used to define the initial state of the helicopter rotor; the model hyperparameter settings include: the number of neural network layers. Number of hidden layer units Used to specify the neural network structure; training learning rate Used to set the gradient descent step size; number of training rounds Used to set the maximum number of training iterations; batch size. Used to specify the number of training data samples; minimum residual , used for training convergence criteria.

[0014] Furthermore, the neural network is initialized: weight parameters The Xavier method is used for random initialization. Xavier initialization ensures that the variances of the physical state quantities (pitch angle and rotor thrust) remain stable during forward propagation, avoiding exponential attenuation or explosion of the signal as it propagates deep within the LSTM network. This provides a numerically stable initial state for calculating the residuals of the dynamic equations in the physical constraint loss term; bias... Set to zero; hidden state With memory unit Initialize the network to a zero vector to allow it to learn input and output data features without prior knowledge; set the optimizer to the Adam optimizer; and read the physical parameters: blade root installation angle. Blade torque Rotor blade azimuth angle Distance from the blade profile to the center of the rotor hub Rotor radius Tail rotor radius rotor advance ratio rotor cone angle Rotor equivalent induced velocity Rotor speed helicopter weight Read model hyperparameters: number of neural network layers Number of hidden layer units Training learning rate Number of training rounds Batch size , the minimum residual .

[0015] In step S2, the long short-term memory neural network model is used to predict the training output data based on the training input data, including: The long short-term memory neural network model is trained based on the coordinate data of the target environmental wind field region and the three-dimensional transient flow field data. Establishing helicopter dynamic response data through a gating mechanism The mapping relationship between them; After training, the long short-term memory neural network model yields preliminary prediction results of helicopter dynamic response. ,in, These represent the predicted pitch and roll angles of the helicopter, respectively. This indicates the predicted rotor collective pitch. This indicates the prediction results of the lateral periodicity variation. This indicates the prediction results of longitudinal periodic pitch. This indicates the predicted collective pitch of the tail rotor.

[0016] Furthermore, during training, the helicopter rotor pitch angle equation and aerodynamic equation are embedded as physical constraints into the loss function, which consists of a data loss term. With physical constraint loss term Together they constitute; among them, data loss It is used to measure the difference between the model's predicted output and the real training samples, and is expressed in the form of mean squared error: ; in, Represents helicopter dynamic response data. This indicates the predicted results of the helicopter's dynamic response. Indicates the spatial point number. Indicates the total number of points in space; The physical constraint loss term is used to correct the physical consistency of the prediction results by introducing constraints from the helicopter dynamics equations; this loss is derived from the residuals of the pitch angle equilibrium equations. Horizontal force equilibrium equation residuals Longitudinal force balance equation residuals The residuals of the pitch angle equilibrium equation constitute As a physical constraint on the helicopter rotor control input of the model, the pitch angle is used. The solution process ensures the rotor collective pitch. Lateral periodic pitch Longitudinal periodic pitch If a balance relationship is satisfied during the prediction process, the residual is expressed as: ; Among them, the pitch angle The equilibrium equation is expressed as: ; in, The pitch angle is calculated from the predicted results, and the equilibrium equation is expressed as: ; in, This indicates the distance from the blade profile to the center of the blade hub. Indicates the rotor radius. Indicates blade torque. Indicates the blade azimuth angle; Transverse force equilibrium equation residuals Represented as: ; in, Indicates rotor thrust. Indicates the rotor shaft tilt angle. This indicates the chamfering after the rotor flaps. Indicates the weight of the helicopter. Indicates the helicopter's pitch angle; Longitudinal force balance equation residuals Represented as: ; in, Indicates the bevel angle of the rotor flapping. Indicates tail rotor thrust. Indicates the helicopter's pitch angle; The rotor thrust The function mapping relationship is expressed as The tail rotor thrust The function mapping relationship is expressed as ,in, Indicates the tail rotor radius; The rotor blades chamfered after flapping Side bevel with rotor flapping The blade flapping motion equation is calculated from the aerodynamic equation. The blade flapping motion equation is a preliminary step in the residual calculation and is directly calculated from the prediction results. By uniformly modeling the longitudinal and lateral spatial tilt of the rotor disk and the aerodynamic load distribution, the physical coupling constraint between the flapping motion and the dynamic response of the airframe is realized. ; ; in, Indicates the rotor advance ratio, Indicates the blade root mounting angle. Indicates the rotor cone angle. Indicates the rotor's equivalent induced velocity. Indicates the rotor speed; The physical constraint loss term Represented as: ; Overall model loss Represented as: .

[0017] In step S4, the long short-term memory neural network model is trained using the training dataset and the hybrid loss function, and the network parameters are updated including: The long short-term memory neural network model processes the coordinate data of the target environmental wind field region and the three-dimensional transient flow field data in the input sequence. The forward propagation process of the neural network is executed to extract the temporal features of the wind field and output the predicted results of the helicopter dynamic response. ; Calculate the weight parameters using the backpropagation algorithm. Bias Initial hidden state With initial memory unit Gradient, and using the Adam adaptive optimizer to adjust the weight parameters Bias Initial hidden state With initial memory unit Perform iterative updates, the iterative update process is represented as: ; ; ; ; in, This represents the overall loss of the model. Represents the weight parameters. Indicates bias. express Hide your status at all times. express Hide your status at all times. express The residual corresponding to the calculation result at time step, express Time-based memory unit express Time-based memory unit This represents the total duration of the input data segment, with the timestamp range expressed as: .

[0018] During training, the learning rate decays using an exponential decay method with a decay rate of 0.95. After each training cycle, the current learning rate is multiplied by 0.95. Setting the value to 0.001 helps the model fine-tune network weights in the later stages of training, avoiding oscillations in the non-convex optimization surface of the physical constraint loss term; the regularization strategy uses a weight decay method, with a regularization coefficient of 1×10. -4 This prevents the model from overfitting specific wind field patterns, thereby enhancing its generalization ability under different inflow conditions.

[0019] Furthermore, the model training process also includes the use of learning rate decay and regularization strategies to achieve a balance between data fitting accuracy and physical consistency.

[0020] In step S5, the target environmental wind field data to be predicted is input into the trained long short-term memory neural network model to quickly predict the helicopter takeoff and landing dynamic response results, including: When inputting the coordinate data of the new target environment wind field region and the three-dimensional transient flow field data... At that time, the model outputs the predicted results of the helicopter dynamic response. .

[0021] Another objective of this invention is to provide a helicopter takeoff and landing dynamics response prediction system based on physical constraints. This system is used to implement the aforementioned helicopter takeoff and landing dynamics response prediction method based on physical constraints. The system includes: The data acquisition module is used to acquire a training dataset, which includes training input data and training output data. The training input data consists of coordinate data of the wind field area of ​​the target environment and three-dimensional transient flow field data. The training output data consists of helicopter dynamic response data corresponding to the training input data. A model building module is used to build a long short-term memory neural network model, which is used to predict the training output data based on the training input data. The loss function construction module is used to design a hybrid loss function, which is composed of a data loss term and a physical constraint loss term. The data loss term is used to measure the difference between the model prediction results and the actual output data. The physical constraint loss term includes at least the residual based on the pitch angle balance equation, which is used to constrain the physical consistency between the model's predicted rotor collective pitch, lateral cyclic pitch and longitudinal cyclic pitch. The model training module is used to train the long short-term memory neural network model using the training dataset and the hybrid loss function, and to update the network parameters. The prediction module is used to input the target environment wind field data to be predicted into the trained long short-term memory neural network model and quickly output the corresponding helicopter take-off and landing dynamic response prediction results.

[0022] Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, this invention employs physically constrained temporal neural network technology, introducing physical residual constraint design based on dynamic control equations into the LSTM structure to achieve a synergistic integration of data-driven modeling and physical consistency. Based on the LSTM temporal network, this invention improves the loss function design of the neural network model. The loss function of this invention is composed of data loss terms and physical constraint loss terms. The physical constraint loss terms include helicopter dynamic equations such as the propeller pitch angle equation and the helicopter rotor thrust equation, thereby improving the accuracy and physical consistency of the neural network training process.

[0023] Secondly, this invention achieves a rapid prediction process for helicopter take-off and landing dynamics response based on physical constraints by introducing dynamic equations and pitch angle constraints into the neural network. This replaces the traditional complex dynamic simulation process with neural network forward propagation calculation, which greatly improves the efficiency of single prediction. It can be directly applied in flight simulators, real-time flight control, and mission planning.

[0024] Third, this invention embeds dynamic constraints such as the pitch angle equation and the helicopter rotor thrust equation into the loss function, so that the prediction results strictly follow the laws of helicopter dynamics while ensuring speed. On the basis of numerical similarity, the prediction results also ensure the balance of dynamic equations, avoiding non-physical problems such as non-conservation of force and torque. Attached Figure Description

[0025] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the disclosure of this invention and, together with the description, serve to explain the principles of the disclosure of this invention. Figure 1 This is a flowchart of a method for predicting the dynamic response of helicopter takeoff and landing based on physical constraints, provided in an embodiment of the present invention. Figure 2 This is a structural diagram of a rapid prediction model for helicopter takeoff and landing dynamics response based on physical constraints of blade pitch angle provided in an embodiment of the present invention; Figure 3 These are comparison charts of prediction results provided by the present invention method and traditional methods. The traditional method selects a Long Short-Term Memory Neural Network (LSTM). Among them, (a) is a comparison chart of rotor collective pitch; (b) is a comparison chart of tail collective pitch; (c) is a comparison chart of lateral periodic pitch; (d) is a comparison chart of longitudinal periodic pitch; (e) is a comparison chart of helicopter roll angle; and (f) is a comparison chart of helicopter pitch angle. Figure 4 This is a comparison chart of the prediction errors of the method provided in this embodiment of the invention and the traditional method; Figure 5 This is a comparison of the training time of the method provided in this embodiment of the invention with that of the traditional method; Figure 6 These are comparison charts of the extrapolation verification experiment prediction results of this method and the traditional method under the condition of 0° headwind as the training set and 15° oblique wind as the verification set, provided by the embodiments of the present invention; wherein, (a) is a comparison chart of rotor collective pitch; (b) is a comparison chart of tail collective pitch; (c) is a comparison chart of lateral periodic pitch; (d) is a comparison chart of longitudinal periodic pitch; (e) is a comparison chart of helicopter roll angle; and (f) is a comparison chart of helicopter pitch angle. Detailed Implementation

[0026] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0027] The innovation of this invention lies in its use of physically constrained temporal neural network technology. It introduces physical residual constraint design based on dynamic control equations into the LSTM structure, achieving a synergistic integration of data-driven modeling and physical consistency. Specifically, this invention incorporates physical residual constraint design of the helicopter rotor pitch angle dynamic control equation and rotor thrust equation into the LSTM network structure. More specifically, this invention improves the loss function design of the neural network model based on the LSTM temporal network. The loss function is composed of a data loss term and a physical constraint loss term, where the physical constraint loss term includes helicopter dynamic equations such as the rotor pitch angle equation and the helicopter rotor thrust equation, thus improving the accuracy and physical consistency of the neural network training process.

[0028] Example 1, such as Figure 1 As shown, the helicopter takeoff and landing dynamics response prediction method based on physical constraints provided in this embodiment of the invention includes the following steps: S1. Obtain a training dataset, which includes training input data and training output data; wherein, the training input data is the coordinate data of the wind field area of ​​the target environment and the three-dimensional transient flow field data, and the training output data is the helicopter dynamic response data corresponding to the training input data; Using the coordinate data of the target environment wind field region and the three-dimensional transient flow field data As the training input of the model, where These represent the spatial coordinates of points within the wind field region. These represent the horizontal, vertical, and lateral wind speed components at the spatial coordinates of a point in the wind field region, respectively. Indicates the timestamp of the wind field area. Indicates the spatial point number. This represents the total number of points in space. (Based on helicopter dynamic response data.) As the training output of the model, These represent the helicopter's pitch and roll angles, respectively. Indicates rotor collective pitch, Indicates lateral periodic pitch, Indicates longitudinal periodic pitch, Indicates the collective pitch of the tail rotor. Indicates the spatial point number. This represents the total number of points in space.

[0029] Before training the model, initial settings are required. The physical parameter settings needed for training include: setting the blade root installation angle. Blade torque Rotor blade azimuth angle Distance from the blade profile to the center of the rotor hub m, rotor radius m, tail rotor radius m, rotor advance ratio rotor cone angle Rotor equivalent induced velocity The rotor speed is approximated using conventional methods. rad / s, helicopter weight N, the above physical parameters are used to define the initial state of the helicopter rotor; model hyperparameter settings include: number of neural network layers. Number of hidden layer units Used to specify the neural network structure; training learning rate Used to set the gradient descent step size; number of training rounds Used to set the maximum number of training iterations; batch size. Used to specify the number of training data samples; minimum residual , used for training convergence criteria.

[0030] Before model training, the neural network is initialized: its weight parameters... Random initialization using the Xavier method, bias Set to 0 to ensure stable gradient propagation; hidden state With memory unit The network is uniformly initialized with a zero vector, allowing it to learn input and output data features without prior knowledge; the optimizer is set to the Adam optimizer. Physical parameters are read: blade root installation angle. Blade torque Rotor blade azimuth angle Distance from the blade profile to the center of the rotor hub Rotor radius Tail rotor radius rotor advance ratio rotor cone angle Rotor equivalent induced velocity Rotor speed helicopter weight Read model hyperparameters: number of neural network layers Number of hidden layer units Training learning rate Number of training rounds Batch size , the minimum residual After completing the above initialization steps, the model training process can begin.

[0031] S2. Construct a long short-term memory neural network model, wherein the long short-term memory neural network model is used to predict the training output data based on the training input data; like Figure 2 As shown, the model consists of six modules: environmental wind field information input module, LSTM gating structure, fully connected layer structure, loss function calculation process, backpropagation module, and dynamic response output module.

[0032] The model is based on the coordinate data of the target environment wind field region and the three-dimensional transient flow field data as training input. Establishing helicopter dynamic response data through a gating mechanism The mapping relationship between them, where These represent the spatial coordinates of points within the wind field region. These represent the horizontal, vertical, and lateral wind speed components at the spatial coordinates of a point in the wind field region, respectively. Indicates the timestamp of the wind field area. Indicates the spatial point number. Indicates the total number of points in space; These represent the helicopter's pitch and roll angles, respectively. Indicates rotor collective pitch, Indicates lateral periodic pitch, Indicates longitudinal periodic pitch, This represents the collective pitch of the tail rotor. This step follows the traditional LSTM training model to obtain preliminary helicopter dynamic response prediction results. This provides an optimization foundation for the subsequent training process based on physical constraints, among which These represent the predicted pitch and roll angles of the helicopter, respectively. Indicates the rotor collective pitch prediction result, Indicates the results of lateral periodic pitch prediction. Indicates the longitudinal periodic pitch prediction results. This indicates the predicted collective pitch of the tail rotor. Indicates the spatial point number. This represents the total number of points in space.

[0033] S3. Design a hybrid loss function, which is composed of a data loss term and a physical constraint loss term. The data loss term is used to measure the difference between the model prediction results and the actual output data. The physical constraint loss term includes at least the residual based on the pitch angle balance equation, which is used to constrain the physical consistency between the rotor collective pitch, lateral cyclic pitch and longitudinal cyclic pitch predicted by the model. To address the lack of physical consistency in traditional LSTM training, this invention incorporates the helicopter rotor pitch angle equation and aerodynamic equations as physical constraints into the loss function during training. This ensures that the prediction results maintain accuracy while adhering to the fundamental laws of helicopter dynamics, thereby achieving a physically guided learning process from wind field input to helicopter dynamics output. Therefore, this invention designs a hybrid loss function, which consists of a data loss term... With physical constraint loss term Together they constitute.

[0034] Among them, data loss It is used to measure the difference between the model's predicted output and the real training samples, and is usually expressed in the form of mean squared error (MSE), as follows: ; in, Represents helicopter dynamic response data. This indicates the predicted results of the helicopter's dynamic response. Indicates the spatial point number. This represents the total number of points in space.

[0035] The physical constraint loss term is introduced by incorporating constraints from the helicopter dynamics equations to correct for physical consistency in the prediction results. This loss is derived from the residuals of the pitch angle equilibrium equations. Horizontal force equilibrium equation residuals Longitudinal force balance equation residuals The residuals of the pitch angle equilibrium equation constitute Represented as: ; in, Expressed as the pitch angle, its equilibrium equation is as follows: ; in, The pitch angle calculated from the predicted results is expressed by the following equilibrium equation: ; in, Indicates the collective pitch of the rotor. Indicates the lateral periodic pitch. Indicates longitudinal periodic pitch. This indicates the predicted rotor collective pitch. This indicates the prediction results of the lateral periodicity variation. This indicates the prediction results of longitudinal periodic pitch. This indicates the distance from the blade profile to the center of the blade hub. Indicates the rotor radius. Indicates blade torque. This indicates the azimuth angle of the propeller blades.

[0036] Transverse force equilibrium equation residuals Represented as: ; in, Indicates rotor thrust. Indicates the rotor shaft tilt angle. This indicates the chamfer after the rotor flaps (relative to the rotor's structural plane). Indicates the weight of the helicopter. Indicates the helicopter's pitch angle. Indicates the spatial point number. This represents the total number of points in space.

[0037] Longitudinal force balance equation residuals Represented as: ; in, Indicates rotor thrust. This indicates the rotor flapping side chamfer (relative to the rotor's structural plane). Indicates tail rotor thrust. Indicates the weight of the helicopter. This indicates the helicopter's pitch angle. It should be noted that this refers to rotor thrust. With tail rotor thrust Based on the fundamental helicopter dynamics equations for rotor and tail rotor thrust, this invention describes the basic calculation process using functional relationships, where the functional mapping relationship of rotor thrust is expressed as follows: The functional mapping relationship of the tail rotor thrust is expressed as: ,in These represent the horizontal, vertical, and lateral wind speed components at the spatial coordinates of a point in the wind field region, respectively. These represent the predicted pitch and roll angles of the helicopter, respectively. Indicates the rotor collective pitch prediction result, Indicates the results of lateral periodic pitch prediction. Indicates the longitudinal periodic pitch prediction results. This indicates the predicted collective pitch of the tail rotor. Indicates the spatial point number. Indicates the tail rotor radius. The chamfer after rotor flapping. Side bevel with rotor flapping The blade flapping motion equations are derived from the aerodynamic equations. These equations, as a preliminary step in residual calculation, can be directly calculated from the prediction results. ; ; in, Indicates the rotor advance ratio, Indicates the blade root mounting angle. Indicates blade torque. Indicates the rotor cone angle. Indicates the rotor's equivalent induced velocity. Indicates the rotor speed. Indicates the rotor radius. This indicates the prediction results of the lateral periodicity variation. This indicates the prediction result of longitudinal periodic pitch.

[0038] Therefore, the physical constraint loss term Represented as ,in The residuals of the pitch angle equilibrium equations are represented by... Represents the residuals of the lateral force equilibrium equations, This represents the residuals of the longitudinal force balance equations. The overall model loss. Represented as ,in Indicates data loss terms and This represents the physical constraint loss term.

[0039] S4. Train the long short-term memory neural network model using the training dataset and the hybrid loss function, and update the network parameters; During model training, the model first processes the coordinate data of the target environment wind field region and the three-dimensional transient flow field data in the input sequence. Perform the basic forward propagation process of the neural network to extract the temporal features of the wind field and output the predicted results of the helicopter dynamic response. ,in These represent the horizontal, vertical, and lateral wind speed components at the spatial coordinates of a point in the wind field region, respectively. Indicates the timestamp of the wind field area. Indicates the spatial point number. Represents the total number of points in space. These represent the predicted pitch and roll angles of the helicopter, respectively. Indicates the rotor collective pitch prediction result, Indicates the results of lateral periodic pitch prediction. Indicates the longitudinal periodic pitch prediction results. This represents the predicted collective pitch of the tail rotor. The weight parameters are then calculated using the backpropagation time-sharing (BPTT) algorithm. Bias Hidden state With memory unit Gradient, and using the Adam adaptive optimizer to adjust the weight parameters Bias Hidden state With memory unit Iterative updates are performed; during training, learning rate decay and regularization strategies are further adopted to improve convergence stability, so as to achieve a balance between data fitting accuracy and physical consistency, thereby ensuring that the model achieves stable convergence and high-precision prediction under limited data conditions.

[0040] S5. Input the target environment wind field data to be predicted into the trained long short-term memory neural network model, and output the corresponding helicopter take-off and landing dynamic response prediction results. After model training and optimization, the rapid prediction model for helicopter takeoff and landing dynamics based on dynamic equations and physical constraints can be directly used for rapid prediction. When new target environment wind field coordinate data and three-dimensional transient flow field data are input... At that time, the model can output helicopter dynamic response prediction results in a very short time. Compared to traditional numerical simulation methods, this significantly reduces computational load and improves real-time performance. These represent the spatial coordinates of points within the wind field region. These represent the horizontal, vertical, and lateral wind speed components at the spatial coordinates of a point in the wind field region, respectively. Indicates the timestamp of the wind field area. Indicates the spatial point number. Indicates the total number of points in space; These represent the predicted pitch and roll angles of the helicopter, respectively. Indicates the rotor collective pitch prediction result, Indicates the results of lateral periodic pitch prediction. Indicates the longitudinal periodic pitch prediction results. This indicates the predicted collective pitch of the tail rotor.

[0041] With the same model hyperparameters (number of neural network layers) Number of hidden layer units Training learning rate Number of training rounds Batch size , the minimum residual Under these conditions, a fast prediction model for helicopter takeoff and landing dynamic response based on dynamic equations and physical constraints and a traditional scheme are trained using coordinate data of the wind field region in the target environment and three-dimensional transient flow field data. As input to the model, the model's prediction results are then compared. These represent the spatial coordinates of points within the wind field region. These represent the horizontal, vertical, and lateral wind speed components at the spatial coordinates of a point in the wind field region, respectively. Indicates the timestamp of the wind field area. Indicates the spatial point number. This represents the total number of points in space.

[0042] Example 2: The helicopter takeoff and landing dynamics response prediction system based on physical constraints provided in this embodiment of the invention includes: The data acquisition module is used to acquire a training dataset, which includes training input data and training output data. The training input data consists of coordinate data of the wind field area of ​​the target environment and three-dimensional transient flow field data. The training output data consists of helicopter dynamic response data corresponding to the training input data. A model building module is used to build a long short-term memory neural network model, which is used to predict the training output data based on the training input data. The loss function construction module is used to design a hybrid loss function, which is composed of a data loss term and a physical constraint loss term. The data loss term is used to measure the difference between the model prediction results and the actual output data. The physical constraint loss term includes at least the residual based on the pitch angle balance equation, which is used to constrain the physical consistency between the model's predicted rotor collective pitch, lateral cyclic pitch and longitudinal cyclic pitch. The model training module is used to train the long short-term memory neural network model using the training dataset and the hybrid loss function, and to update the network parameters. The prediction module is used to input the target environment wind field data to be predicted into the trained long short-term memory neural network model and quickly output the corresponding helicopter take-off and landing dynamic response prediction results.

[0043] To further demonstrate the positive effects of the above embodiments, the present invention conducts the following experiments based on the above technical solutions.

[0044] Comparison of prediction results between the present invention and traditional methods Figure 3 As shown (see details) Figure 3 Figure (a) in the middle - Figure 3 (Figure f) shows a comparison of prediction errors with traditional methods. Figure 4 As shown in the figure. Comparing the model outputs, the fast prediction model for helicopter takeoff and landing dynamics based on dynamic equations and physical constraints proposed in this invention shows a significant improvement in prediction performance compared to traditional schemes. The average relative root mean square error (RMSE) of the prediction results for each control variable in this invention is 8.64%, while the average RMSE of the prediction results for each control variable in traditional schemes (represented by Long Short-Term Memory Neural Networks (LSTM)) is 16.11%. Regarding efficiency, the Newton iteration method, as a typical iterative solution method for nonlinear dynamic equations, is therefore used as a comparative verification method. The comparison results are shown in the figure. Figure 5 As shown, under the same parameter input conditions, the single run time of the present invention is 1.12s, while the single run time of the dynamic method using Newton's iteration method is 4.26s, thus proving the effectiveness of the present invention.

[0045] To verify the effectiveness of the physical constraints of the present invention when the input wind field data exceeds the range of the training set, the following extrapolation verification experiment was designed: Training set data: The three-dimensional transient flow field data used for model training is simulation data under a 0° headwind condition; Validation set data: The three-dimensional transient flow field data used for model validation is simulation data under a 15° oblique wind condition. Under the same training set data input conditions, the model proposed in this invention and the Long Short-Term Memory (LSTM) neural network model without physical constraints are trained simultaneously. After training, the model error is compared using the same validation set data.

[0046] The comparison results are as follows Figure 6 As shown (see details) Figure 6 Figure (a) in the middle - Figure 6 (Figure (f)) Although the input wind field of the validation dataset exceeds the distribution of the training dataset, the model output prediction is forced to satisfy the dynamic equation due to the physical residual constraints of the pitch angle equation and rotor thrust equation. The relative root mean square error between the control prediction result and the true value is 12.72%, which is less than the relative root mean square error of 21.37% of the prediction result of the Long Short-Term Memory Neural Network (LSTM) model without physical constraints. This proves that the model improvement method proposed in this invention has a certain generalization ability.

[0047] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the dynamic response of helicopter takeoff and landing based on physical constraints, characterized in that, The method includes the following steps: S1. Obtain a training dataset, which includes training input data and training output data; wherein, the training input data is the coordinate data of the wind field region of the target environment and the three-dimensional transient flow field data, and the training output data is the helicopter dynamic response data corresponding to the training input data; S2. Construct a long short-term memory neural network model, wherein the long short-term memory neural network model is used to predict the training output data based on the training input data; S3. Design a hybrid loss function, which is composed of a data loss term and a physical constraint loss term. The data loss term is used to measure the difference between the model prediction results and the actual output data. The physical constraint loss term includes at least the residual based on the pitch angle balance equation, which is used to constrain the physical consistency between the rotor collective pitch, lateral cyclic pitch and longitudinal cyclic pitch predicted by the model. S4. Train the long short-term memory neural network model using the training dataset and the hybrid loss function, and update the network parameters; S5. Input the target environment wind field data to be predicted into the trained long short-term memory neural network model, and output the corresponding helicopter take-off and landing dynamic response prediction results.

2. The method for predicting helicopter takeoff and landing dynamics response based on physical constraints according to claim 1, characterized in that, In step S1, the training dataset includes: Using the coordinate data of the target environment wind field region and the three-dimensional transient flow field data As the training input for the model, These represent the spatial coordinates of points within the wind field region. These represent the horizontal, vertical, and lateral wind speed components at the spatial coordinates of a point in the wind field region, respectively. Indicates the timestamp of the wind field area. Indicates the spatial point number. Indicates the total number of points in space; Using helicopter dynamic response data As the training output of the model, These represent the helicopter's pitch and roll angles, respectively. Indicates rotor collective pitch, Indicates lateral periodic pitch, Indicates longitudinal periodic pitch, This indicates the collective pitch of the tail rotor.

3. The method for predicting helicopter takeoff and landing dynamics response based on physical constraints according to claim 1, characterized in that, In step S1, before model training, the model is initially set up, including setting the physical parameters such as the blade root mounting angle. Blade torque Rotor blade azimuth angle Distance from the blade profile to the center of the rotor hub Rotor radius Tail rotor radius rotor advance ratio rotor cone angle Rotor equivalent induced velocity Rotor speed helicopter weight The physical parameters are used to define the initial state of the helicopter rotor; the model hyperparameter settings include: the number of neural network layers. Number of hidden layer units Used to specify the neural network structure; training learning rate Used to set the gradient descent step size; number of training rounds Used to set the maximum number of training iterations; batch size. Used to specify the number of training data samples; minimum residual , used for training convergence criteria.

4. The method for predicting helicopter takeoff and landing dynamics response based on physical constraints according to claim 3, characterized in that, Initialize the neural network: weight parameters The Xavier method is used for random initialization. The Xavier method initialization ensures that the variance of the physical state quantities of pitch angle and rotor thrust remains stable during forward propagation, avoiding exponential attenuation or explosion of the signal when it is propagated deep in the LSTM network, and providing a numerically stable initial state for the calculation of the residual of the dynamic equation in the physical constraint loss term. bias Set to zero; hidden state With memory unit The network is uniformly initialized as a zero vector, enabling it to learn the features of the input and output data without prior knowledge. Set the optimizer to Adam optimizer; read the physical parameters: blade root installation angle. Blade torque Rotor blade azimuth angle Distance from the blade profile to the center of the rotor hub Rotor radius Tail rotor radius rotor advance ratio rotor cone angle Rotor equivalent induced velocity Rotor speed helicopter weight Read model hyperparameters: number of neural network layers Number of hidden layer units Training learning rate Number of training rounds Batch size , the minimum residual .

5. The method for predicting helicopter takeoff and landing dynamics response based on physical constraints according to claim 2, characterized in that, In step S2, the long short-term memory neural network model is used to predict the training output data based on the training input data, including: The long short-term memory neural network model is trained based on the coordinate data of the target environmental wind field region and the three-dimensional transient flow field data. Establishing helicopter dynamic response data through a gating mechanism The mapping relationship between them; After training, the long short-term memory neural network model yields preliminary prediction results of helicopter dynamic response. ,in, These represent the predicted pitch and roll angles of the helicopter, respectively. This indicates the predicted rotor collective pitch. This indicates the prediction results of the lateral periodicity variation. This indicates the prediction results of longitudinal periodic pitch. This indicates the predicted collective pitch of the tail rotor.

6. The method for predicting helicopter takeoff and landing dynamics response based on physical constraints according to claim 5, characterized in that, During training, the helicopter rotor pitch angle equation and aerodynamic equation are embedded as physical constraints into the loss function, which consists of a data loss term. With physical constraint loss term Together they constitute; among them, data loss It is used to measure the difference between the model's predicted output and the real training samples, and is expressed in the form of mean squared error: ； in, Represents helicopter dynamic response data. This indicates the predicted results of the helicopter's dynamic response. Indicates the spatial point number. Indicates the total number of points in space; The physical constraint loss term is used to correct the physical consistency of the prediction results by introducing constraints from the helicopter dynamics equations; this loss is derived from the residuals of the pitch angle equilibrium equations. Horizontal force equilibrium equation residuals Longitudinal force balance equation residuals The residuals of the pitch angle equilibrium equation constitute As a physical constraint on the helicopter rotor control input of the model, the pitch angle is used. The solution process ensures the rotor collective pitch. Lateral periodic pitch Longitudinal periodic pitch If a balance relationship is satisfied during the prediction process, the residual is expressed as: ； Among them, the pitch angle The equilibrium equation is expressed as: ； in, The pitch angle is calculated from the predicted results, and the equilibrium equation is expressed as: ； in, This indicates the distance from the blade profile to the center of the blade hub. Indicates the rotor radius. Indicates blade torque. Indicates the blade azimuth angle; Transverse force equilibrium equation residuals Represented as: ； in, Indicates rotor thrust. Indicates the rotor shaft tilt angle. Indicates the chamfering after the rotor flaps. Indicates the weight of the helicopter. Indicates the helicopter's pitch angle; Longitudinal force balance equation residuals Represented as: ； in, Indicates the bevel angle of the rotor flapping. Indicates tail rotor thrust. Indicates the helicopter's pitch angle; The rotor thrust The function mapping relationship is expressed as The tail rotor thrust The function mapping relationship is expressed as ,in, Indicates the tail rotor radius; The rotor blades chamfered after flapping Side chamfer with rotor flapping The blade flapping motion equation is calculated from the aerodynamic equation. The blade flapping motion equation is a preliminary step in the residual calculation and is directly calculated from the prediction results. By uniformly modeling the longitudinal and lateral spatial tilt of the rotor disk and the aerodynamic load distribution, the physical coupling constraint between the flapping motion and the dynamic response of the airframe is realized. ； ； in, Indicates the rotor advance ratio, Indicates the blade root mounting angle. Indicates the rotor cone angle. Indicates the rotor's equivalent induced velocity. Indicates the rotor speed; The physical constraint loss term Represented as: ； Overall model loss Represented as: 。 7. The method for predicting helicopter takeoff and landing dynamics response based on physical constraints according to claim 6, characterized in that, In step S4, the long short-term memory neural network model is trained using the training dataset and the hybrid loss function, and the network parameters are updated including: The long short-term memory neural network model processes the coordinate data of the target environmental wind field region and the three-dimensional transient flow field data in the input sequence. The forward propagation process of the neural network is executed to extract the temporal features of the wind field and output the predicted results of the helicopter dynamic response. ; Calculate the weight parameters using the backpropagation algorithm. Bias Initial hidden state With initial memory unit Gradient, and using the Adam adaptive optimizer to adjust the weight parameters Bias Initial hidden state With initial memory unit Perform iterative updates, the iterative update process is represented as: ； ； ； ； in, This represents the overall loss of the model. Represents the weight parameters. Indicates bias. express Hide your status at all times. express Hide your status at all times. express The residual corresponding to the calculation result at time step, express Time-based memory unit express Time-based memory unit This represents the total duration of the input data segment, with the timestamp range expressed as: ; During training, the learning rate decays using an exponential decay method with a decay rate of 0.

95. After each training cycle, the current learning rate is multiplied by 0.

95. Setting the value to 0.001 helps the model fine-tune network weights in the later stages of training, avoiding oscillations in the non-convex optimization surface of the physical constraint loss term; the regularization strategy uses a weight decay method, with a regularization coefficient of 1×10. -4 This prevents the model from overfitting specific wind field patterns, thereby enhancing its generalization ability under different inflow conditions.

8. The method for predicting helicopter takeoff and landing dynamics response based on physical constraints according to claim 7, characterized in that, The model training process also includes the use of learning rate decay and regularization strategies to achieve a balance between data fitting accuracy and physical consistency.

9. The method for predicting helicopter takeoff and landing dynamics response based on physical constraints according to claim 7, characterized in that, In step S5, the target environmental wind field data to be predicted is input into the trained long short-term memory neural network model to quickly predict the helicopter takeoff and landing dynamic response results, including: When inputting the coordinate data of the new target environment wind field region and the three-dimensional transient flow field data... At that time, the model outputs the predicted results of the helicopter dynamic response. .

10. A helicopter takeoff and landing dynamics response prediction system based on physical constraints, characterized in that, This system is used to implement the helicopter takeoff and landing dynamics response prediction method based on physical constraints as described in any one of claims 1 to 9, and the system comprises: The data acquisition module is used to acquire a training dataset, which includes training input data and training output data. The training input data consists of coordinate data of the wind field area of ​​the target environment and three-dimensional transient flow field data. The training output data consists of helicopter dynamic response data corresponding to the training input data. A model building module is used to build a long short-term memory neural network model, which is used to predict the training output data based on the training input data. The loss function construction module is used to design a hybrid loss function, which is composed of a data loss term and a physical constraint loss term. The data loss term is used to measure the difference between the model prediction results and the actual output data. The physical constraint loss term includes at least the residual based on the pitch angle balance equation, which is used to constrain the physical consistency between the model's predicted rotor collective pitch, lateral cyclic pitch and longitudinal cyclic pitch. The model training module is used to train the long short-term memory neural network model using the training dataset and the hybrid loss function, and to update the network parameters. The prediction module is used to input the target environment wind field data to be predicted into the trained long short-term memory neural network model and quickly output the corresponding helicopter take-off and landing dynamic response prediction results.

Images