An aircraft trajectory prediction method based on deep learning

Abstract

The application provides a kind of aircraft trajectory prediction method based on deep learning, it is related to machine learning technical field, specifically includes: according to six degrees of freedom aircraft model simulation to obtain flight trajectory sample data, based on attention mechanism and gated recurrent unit, construct Dual-attention GRU model;According to sample data generation training data set and test data set, according to training data set training the model, set parameter and evaluation index, select different parameters to train to determine the hyperparameter model with minimum loss on test set;Use hyperparameter model to predict aircraft trajectory.The method can not only improve the accuracy of predicting the future flight trajectory of enemy aircraft, but also improve the generalization ability of the prediction model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of machine learning, and particularly relates to a flight vehicle trajectory prediction method based on deep learning. BACKGROUND

[0002] In a flight competition, forming an observation-orientation-decision-action (OODA) cycle before an opponent is the key to victory. Accurate prediction of the trajectory of the opponent's flight vehicle will enable the prediction party to take more predictive and effective maneuvers or countermeasures in combination with the trajectory prediction result, greatly shortening the OODA cycle. For the party that cannot predict the trajectory of the flight vehicle, its OODA cycle will always lag behind the opponent, which may lead to defeat. Therefore, flight trajectory prediction is the key to evasion and countermeasures in air games and is an important part of winning.

[0003] Flight trajectory prediction is essentially a time series prediction problem. According to the historical movement trajectory of the target, the inherent information is learned, and then a reasonable prediction of the future movement trend of the target is made. Current flight trajectory prediction methods can be mainly divided into traditional model methods and deep learning methods based on data driving.

[0004] Traditional model methods mainly include Kalman filter, polynomial regression and the like. Lu Bo et al. established a multi-information extended Kalman filter prediction model to realize the trajectory prediction of a flight vehicle. Qiao Shaojie et al. proposed a dynamic trajectory prediction algorithm based on Kalman filter, which improved the prediction accuracy and real-time performance. Zhang Pingping et al. proposed a polynomial trajectory prediction method, which adopts the least square method for model parameter estimation. The traditional prediction algorithm has problems in that it cannot sufficiently learn the maneuvering characteristics of the target, the prediction accuracy is low, and usually in order to more accurately describe the movement of the target, the complexity of the model is high in modeling, which leads to poor real-time performance and generalization of the prediction algorithm, and it is difficult to meet the needs of air combat confrontation.

[0005] With the development of deep learning, many deep learning algorithms have also been introduced into the field of flight trajectory prediction and have achieved certain results. Qian Kui et al. first cluster the flight trajectory of the aircraft, and then predict the flight trajectory through the BP network, but the prediction result is not ideal because the BP network cannot model the time sequence dependence of the data. Li Nan et al. comprehensively consider the three-dimensional position, speed and other characteristics of the flight trajectory, and use the BP network to establish a flight trajectory prediction model to more accurately predict the flight trajectory of the aircraft. Wang Xin et al. sample a flight trajectory and input it into the Elman network, and use a hybrid particle swarm algorithm to optimize the initial weights of the Elman network, thereby reducing the training time of the network and improving the prediction accuracy of the network. Kou Yinxin et al. construct a KELM model and optimize its parameters using an improved bat algorithm, then use the optimized KELM neural network as a weak predictor, combine it with an ensemble learning algorithm to generate a strong predictor, and continuously optimize the structure and parameters of the strong predictor through training. The simulation results show that the target maneuvering trajectory prediction model has good prediction accuracy and generalization ability. Yang Xingyu uses a graph convolution network to model the group flight trajectory prediction problem, takes each fighter as a node of the graph structure, and each node contains the position sequence of the fighter at the historical time. The change of the graph structure is used to describe the whole flight trend, capture the spatial correlation in the combat process, and add an attention mechanism to strengthen the learning ability of the interaction between the fighters, thereby improving the prediction accuracy of the flight trajectory. Alahi et al. use a long short-term memory (LSTM) network to predict the trajectories of multiple targets, and use a pooling layer to achieve the purpose of sharing information between adjacent targets, thereby improving the accuracy of multi-target trajectory prediction. Zhang Hongpeng et al. use a gate recurrent unit (GRU) to conduct trajectory prediction research, and the results show that the average absolute error of the method is about 18 in the x-axis, about 11 in the y-axis, and about 22 in the z-axis, which is significantly smaller than that of the ordinary BP network. Zhang Zongteng uses a bidirectional GRU to predict the flight trajectory, and the results show that the average absolute error of the prediction result in the XYZ three-axis direction is about 10. SUMMARY

[0006] To solve the above technical problems, the present application provides a flight vehicle tactical trajectory prediction method based on deep learning, which can not only improve the accuracy of predicting the future flight trajectory of the enemy aircraft, but also improve the generalization ability of the prediction model.

[0007] Therefore, according to an aspect of an embodiment of the present application, a flight vehicle trajectory prediction method based on deep learning is provided, comprising:

[0008] Step S1: obtaining flight trajectory sample data according to a six-degree-of-freedom aircraft model simulation, wherein each trajectory sample comprises three-dimensional coordinates and velocity, attitude quantity of the six-degree-of-freedom aircraft model;

[0009] Step S2: constructing a Dual-attention GRU model based on an attention mechanism and a gated recurrent unit network;

[0010] Step S3: generating a training data set and a test data set according to sample data, training the model according to the training data set, setting parameters and evaluation indexes, and selecting different parameters for training to determine a hyperparameter model with minimum loss on the test set;

[0011] Step S4: predicting an aircraft trajectory using the hyperparameter model.

[0012] Optionally, before step S1, the method further comprises establishing kinematic equations and dynamic equations of the six-degree-of-freedom aircraft model, and aircraft attitude kinematic equations and dynamic equations, with an origin of an aircraft coordinate system as a mass center of the aircraft.

[0013] Optionally, step S2 further comprises:

[0014] Step S21: in an encoding stage, inputting an attention mechanism to adaptively select a relevant state feature sequence, wherein an encoder is used to encode an original sequence into a model feature representation to give a sequence feature with k dimensions and a time length of T, and according to the attention mechanism, a weight of each dimension feature is calculated in combination with hidden layer information;

[0015] Step S22: in a decoding stage, using an attention mechanism to adaptively select encoding information output by the encoder, wherein an input of a decoder is a target sequence at a previous time, a hidden state, and a scenario vector.

[0016] Optionally, step S3 comprises:

[0017] Step S31: adopting a min-max method to normalize state feature data;

[0018] Step S32: dividing the normalized state feature data using a sliding window algorithm to obtain flight trajectory samples, and setting a window size as T, wherein data at T-1 previous time is used as a model input value, and data at T time is used as a true value.

[0019] Optionally, step S3 further comprises:

[0020] Step S33: setting a hyperparameter of the Dual-Attention GRU model as an encoder GRU unit hidden state dimension equal to a decoder GRU unit hidden state dimension;

[0021] Step S34: the loss function is mean square error, and the evaluation indexes include mean absolute error, root mean square error and mean absolute percentage error.

[0022] Optionally, the method further comprises: training the model hyperparameters with different values by using an Adam optimizer according to the loss function.

[0023] Optionally, in step S33, the GRU units of the encoder and the decoder are set as p=q∈{64, 128, 256, 512}.

[0024] Optionally, the method further comprises: step S5: comparing the predicted flight trajectory and the real flight trajectory to determine the coincidence degree.

[0025] The application utilizes the recurrent neural network such as GRU and its variant structure to solve the flight trajectory prediction accuracy problem, improves the expression ability of the current trajectory prediction model, and predicts the trajectory more accurately. BRIEF DESCRIPTION OF DRAWINGS

[0026] FIG. 1 is a Dual-Attention GRU model diagram of a deep learning-based aircraft tactical trajectory prediction method according to an embodiment of the application, wherein, Figure 1-1 is an attention mechanism for the encoding stage, Figure 1-2 is an attention mechanism for the decoding stage.

[0027] FIG. 2 is a Dual-Attention GRU model training loss curve diagram of a deep learning-based aircraft tactical trajectory prediction method according to an embodiment of the application, wherein, Figure 2-1 is the X-axis, Figure 2-2 is the Y-axis, Figure 2-3 is the Z-axis.

[0028] FIG. 3 is a comparison diagram of the predicted trajectory and the flight trajectory of a Dual-Attention GRU model of a deep learning-based aircraft tactical trajectory prediction method according to an embodiment of the application, wherein, Figure 3-1 is the X-axis, Figure 3-2 is the Y-axis, Figure 3-3 is the Z-axis. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0030] A deep learning-based aircraft trajectory prediction method comprises:

[0031] Step S1: Obtain flight trajectory sample data according to a six-degree-of-freedom aircraft model simulation, wherein each trajectory sample comprises three-dimensional coordinates and velocity, attitude of the six-degree-of-freedom aircraft model;

[0032] Step S2: Construct a Dual-attention GRU model based on an attention mechanism and a gated recurrent unit (GRU); specifically, the Dual-attention GRU flight trajectory prediction model is a sequence-to-sequence model with a two-stage attention mechanism and can be divided into an encoder and a decoder. Both the encoder and the decoder use GRU as a basic unit. GRU is a variant of RNN proposed in 2014 and has the characteristics of simple structure and strong expression ability, which can effectively learn the time sequence dependence of flight trajectories.

[0033] Step S3: Generate a training data set and a test data set according to sample data, train the model according to the training data set, set parameters and evaluation indexes, and select different parameters for training to determine the hyperparameter model with the minimum loss on the test set;

[0034] Specifically, different model parameters are selected for experiments to determine the final model structure, and the results can be seen in Table 1. As can be seen from the table, the loss of the X-axis and the Y-axis is the smallest when the neuron parameter is 128, and the loss of the Z-axis is the smallest when the neuron parameter is 256 (2.53), but is close to that when the neuron parameter is 128 (2.64). Considering the model accuracy and complexity, 128 is finally selected as the final model structure.

[0035] Table 1: Test set loss of different hyperparameter models

[0036]

[0037] The loss curve of the model trained when the hyperparameters p and q are 128 is taken. After multiple rounds of training, the model reaches a state of convergence. Thus, the hyperparameter model is determined.

[0038] Step S4: Use the hyperparameter model to predict the trajectory of the aircraft.

[0039] Further, before step S1, the method further comprises establishing kinematic equations and dynamic equations of a six-degree-of-freedom aircraft model, and aircraft attitude kinematic equations and dynamic equations with the origin of the aircraft coordinate system as the aircraft center of mass. A six-degree-of-freedom aircraft model is established, a large number of trajectory samples are simulated by giving initial control quantities within a certain range, wherein each trajectory is composed of three-dimensional coordinates and velocity, attitude, etc. of the aircraft. These trajectories can basically cover various trajectory changes that may occur within a period of time under multiple states of the aircraft.

[0040] Specifically, the origin of the body coordinate system is the center of mass of the aircraft, and the specific kinematic equation of the six-degree-of-freedom aircraft model is established:

[0041]

[0042] In the formula, X, Y, and Z are the positions of the aircraft in the ground coordinate system, u, v, and w are the components of the airspeed of the aircraft in the body coordinate system, theta, phi are the pitch angle, yaw angle, and roll angle of the aircraft.

[0043] The dynamic equation is established as follows:

[0044]

[0045] In the formula, p, q, and r are the roll angular velocity, pitch angular velocity, and yaw angular velocity, F x , F y , and F z are the components of the resultant force in the body coordinate system, g is the gravitational acceleration, and m is the mass of the aircraft.

[0046] The aircraft attitude kinematic equation and dynamic equation are established as follows:

[0047]

[0048] In the formula, is the moment of inertia, I xz is the inertia product, L, M, and N are the roll moment, pitch moment, and yaw moment.

[0049] Further, step S2 further comprises:

[0050] Step S21: In the encoding stage, the attention mechanism is used to adaptively select the relevant state feature sequence, wherein the encoder is used to encode the original sequence into a model feature representation to give a k-dimensional sequence feature with a time length of T, and according to the attention mechanism, the weight of each dimensional feature is calculated in combination with the hidden layer information.

[0051] Step S22: In the decoding stage, the attention mechanism is used to adaptively select the encoding information output by the encoder, wherein the input of the decoder is the target sequence at the previous time, the hidden state, and the scene vector.

[0052] Specifically, the encoder is used to encode the original sequence into a model feature representation to give a k-dimensional sequence feature with a time length of T:

[0053]

[0054] According to the attention mechanism, the weight of each dimensional feature is calculated in combination with the hidden layer information The formula is:

[0055]

[0056]

[0057] Where h t-1 v is the hidden state of the GRU. e W e U e The normalized attention weights are obtained from the learnable parameters. The new feature sequence can then be represented as:

[0058]

[0059] For updating the hidden state of a GRU unit, we have:

[0060] h t =f1(h t-1 x t )

[0061] Where f1 is the GRU hidden state update function, the calculation process is as follows:

[0062]

[0063]

[0064] h t =tan h(W) x x t +W h (r t eh t-1 )+b h )

[0065]

[0066] Where r t ,z t To reset and update the door, For the candidate hidden state, W r W z W x W h and b r ,b z ,b h For learnable parameters, ⊙ indicates element-wise multiplication.

[0067] The decoding phase uses an attention mechanism to adaptively select the encoded information output by the encoder:

[0068] The input to the decoder is the target sequence y from the previous time step. t-1Hidden state d t-1 and scenario vector c t-1 c t-1 The calculation formula is:

[0069]

[0070] in, Based on the hidden state d of the decoder in the previous time step t-1 The encoder's feature representation of information at each time step is obtained, and the formula is:

[0071]

[0072]

[0073] Among them, v d W d U d These are learnable parameters. Similarly, using d... t =f2(y t-1 ,c t-1 ,d t-1 Update the hidden state of the GRU unit, where f2 is the decoder GRU unit hidden state update function. Then, for the model output, we have:

[0074]

[0075] in The model outputs the predicted trajectory coordinates at time T, [d T c T [] is the concatenation of the hidden state and the context vector. W y ,b w ,b v These are learnable parameters.

[0076] Further, step S3 includes:

[0077] Step S31: To reduce the impact of data range on algorithm performance and improve network convergence speed, the min-max method is used to normalize the state feature data; the formula is as follows:

[0078]

[0079] In the formula x n Here, x represents the normalized sample, and x represents the original sample. min Let x be the minimum value of a state variable in this flight trajectory. max This represents the maximum value of a state variable within the flight trajectory.

[0080] Step S32: The normalized state feature data is further divided by a sliding window algorithm to obtain flight trajectory samples, and the window size is set to T, wherein the data at the previous T-1 time is taken as the model input value, and the data at the T time is taken as the true value.

[0081] Further, step S3 further includes:

[0082] Step S33: The hyperparameters of the Dual-Attention GRU model are set as follows: the encoder GRU unit hidden state dimension is equal to the decoder GRU unit hidden state dimension; the hyperparameters of the Dual-Attention GRU model mainly include the encoder GRU unit hidden state dimension (p) and the decoder GRU unit hidden state dimension (q). In order to obtain better performance of the model parameters and take into account the training efficiency, the GRU units of the encoder and the decoder are set as p=q∈{64, 128, 256, 512}.

[0083] Step S34: The loss function is mean square error, and the evaluation indexes include mean absolute error, root mean square error and mean absolute percentage error, wherein the loss function is mean square error (MSE), and the specific formula is as follows:

[0084]

[0085] In order to more objectively and accurately evaluate the performance difference of the algorithm, three evaluation indexes are selected to evaluate the performance of the algorithm, which are:

[0086] Mean absolute error (MAE) is:

[0087]

[0088] Root mean square error (RMSE) is:

[0089]

[0090] Mean square percentage error (MAPE) is:

[0091]

[0092] Further, it further includes: according to the loss function, the model hyperparameters with different values are trained by using the Adam optimizer.

[0093] Further, it further comprises: step S5: comparing the predicted flight trajectory and the real flight trajectory to determine the coincidence degree. The comparison of the model to one of the predicted trajectory and the real trajectory is given, the fitting degree of the prediction result of the model in the three axes and the original trajectory is viewed, whether the predicted flight trajectory and the actual flight trajectory remain highly consistent, and whether the two trajectory lines are nearly coincident.

[0094] Embodiment 1

[0095] The application provides a kind of based on deep learning's aircraft trajectory prediction method, comprising:

[0096] (1) establishes six degrees of freedom aircraft model, for simulation to obtain trajectory data.

[0097] The six degrees of freedom aircraft model in the step (1) is established, and the specific steps include:

[0098] (1-1) with the origin O of body coordinate system body As aircraft mass center, establish specific kinematics equation;

[0099] (1-2) establishes aircraft attitude kinematics equation and dynamics equation.

[0100] (2) design a Dual-attention GRU model.

[0101] The Dual-attention GRU model in the step (2) is designed, and the specific steps include:

[0102] As shown in Figure 1, in the encoding stage, the input attention mechanism is introduced to adaptively select the relevant state feature sequence;In decoding stage, the attention mechanism is used to adaptively select the encoding information of the encoder output, wherein the encoder and the decoder both adopt GRU as the basic unit.

[0103] (3) data preprocessing, experimental setting and selecting appropriate evaluation index.

[0104] The data preprocessing, experimental setting and selecting appropriate evaluation index in the step (3) are as follows:

[0105] (3-1) data preprocessing: 5000 groups of flight trajectory data are obtained by flight simulation, and the interval of data is 0.05s, and each data group includes three-dimensional coordinates, speed, attitude and other state variables of aircraft. 4000 groups of trajectory data are used as training set, and 1000 groups of trajectory data are used as test set;

[0106] (3-2) the feature state data is normalized to the range of [0,1] by using min-max method;

[0107] (3-3) The normalized data is divided by using a sliding window algorithm to obtain flight trajectory samples;

[0108] (3-4) The GRU units of the encoder and the decoder are set as p=q∈{64, 128, 256, 512};

[0109] (3-5) The loss function is set as Mean Square Error (MSE), and three evaluation indexes of Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Mean Square Percentage Error (MAPE) are selected to evaluate the performance of the algorithm.

[0110] (4) Different model parameters are selected for experiments to determine the final model structure. The model with the smallest loss on the test set is selected through experiments.

[0111] The final model structure is determined and the most suitable hyperparameters are selected in step (4), and the specific steps are as follows:

[0112] (4-1) Different model parameters are selected for experiments to determine the final model structure;

[0113] (4-2) As shown in FIG. 2, the X-axis and Y-axis losses are the smallest when the neuron parameter is 128, and the Z-axis loss is the smallest when the neuron parameter is 256 (2.53), but is close to 128 (2.64). Considering the model accuracy and complexity, 128 is finally selected as the final model structure. The loss curve of the model trained when the hyperparameter p=q=128 is shown in FIG. 2. As shown in the figure, after multiple rounds of training, the model reaches a convergent state;

[0114] (5) The real flight trajectory and the predicted trajectory are compared.

[0115] The real flight trajectory and the predicted trajectory are compared in step (5), and the specific steps are as follows:

[0116] As shown in FIG. 3, the comparison between the predicted trajectory and the real trajectory of the model is given. It can be seen that the prediction results of the model on the three axes are close to the original trajectory, the predicted flight trajectory is highly consistent with the actual flight trajectory, and the two trajectory lines are almost coincident, having high prediction accuracy. The predicted flight trajectory can maintain a low error level in the turning section, and is closer to the actual trajectory. Therefore, from the three-dimensional coordinates of the trajectory, the prediction accuracy of the network can accurately determine the current position of the aircraft, meeting the requirement of positioning accuracy of the enemy aircraft in air combat.

[0117] (6) Select the same structure of BP network and GRU network for training, and compare the prediction accuracy of different methods in multiple evaluation indexes.

[0118] Embodiment 2

[0119] In this embodiment, a deep learning-based aircraft trajectory prediction method is used, and the implementation steps are as follows:

[0120] (1) Establish a six-degree-of-freedom aircraft model, and simulate a large number of trajectory samples by giving initial control quantities within a certain range.

[0121] (2) A Dual-attention GRU model is designed;

[0122] (3) Data preprocessing, experimental setting and selection of appropriate evaluation indexes;

[0123] (4) Different model parameters are selected for experiment to determine the final model structure. Train it, and select the hyperparameter model with the smallest loss on the test set through experiment.

[0124] (5) Compare the real flight trajectory and the predicted trajectory results;

[0125] (6) Select the same structure of BP network and GRU network for training, and compare the prediction accuracy of different methods in multiple evaluation indexes.

[0126] The application compares the prediction performance of BP network, GRU and LSTM and the proposed Dual-Attention GRU model in three axes in multiple evaluation indexes, wherein the number of neurons of the comparison model is set to 128, and the results are shown in Tables 2, 3 and 4. From the tables, it can be seen that the proposed Dual-Attention GRU model has significant advantages in each evaluation index. Among them, the error of the BP network is the largest because it cannot model the time sequence dependence in the trajectory data, the errors of the GRU and LSTM models are similar, which are lower than that of the BP network, and the Dual-Attention GRU model can maintain a high level of precision in flight trajectory prediction in three-axis data, which is much higher than that of the BP, GRU and LSTM models. This shows that the proposed model has significantly improved the target trajectory prediction accuracy, and verifies the significant superiority and effectiveness of the Dual-Attention GRU model compared with the BP, GRU and LSTM networks in target trajectory time series prediction.

[0127] Table 2 Comparison of performance of different models X axis

[0128]

[0129] Table 3 Different model performance comparison Y axis

[0130]

[0131] Table 4 Different model performance comparison Z axis

[0132]

[0133] The above merely describes specific embodiments of the present application, and the detailed description is not exhaustive of the conventional technology. However, the protection scope of the present application is not limited thereto, and any changes or replacements that can be easily thought of by those skilled in the art within the technical scope disclosed by the present application shall be encompassed within the protection scope of the present application. The protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A deep learning-based method for predicting aircraft trajectories, characterized in that, include: Step S1: Obtain flight trajectory sample data based on the simulation of the six-degree-of-freedom aircraft model. Each trajectory sample includes the three-dimensional coordinates, velocity, and attitude parameters of the six-degree-of-freedom aircraft model. Using the origin of the aircraft coordinate system as the aircraft's centroid, we establish the kinematic and dynamic equations of a six-degree-of-freedom aircraft model, as well as the attitude kinematic and dynamic equations of the aircraft. By establishing the six-degree-of-freedom aircraft model and simulating a large number of trajectory samples within a certain range of initial control quantities, each trajectory is composed of the aircraft's three-dimensional coordinates and velocity and attitude quantities. Using the origin of the body coordinate system as the center of mass of the aircraft, the specific kinematic equations of the six-degree-of-freedom aircraft model are established: ； In the formula, X, Y, and Z represent the aircraft's position in the ground coordinate system, u, v, and w represent the components of the aircraft's airspeed in the body coordinate system, and θ, , These are the pitch angle, yaw angle, and roll angle of the aircraft. Establish its dynamic equations: ； In the formula, p, q, and r are the roll rate, pitch rate, and yaw rate, respectively, and F is the angular velocity. x F y F z Let g be the component of the resultant force in the body coordinate system, g be the gravitational acceleration, and m be the mass of the aircraft. Establish the kinematic and dynamic equations of the aircraft's attitude: ； In the formula, For rotational inertia, I xz Let L be the product of inertia, and M and N be the rolling moment, pitching moment, and yaw moment, respectively. Step S2: Construct a Dual-attention GRU model based on the attention mechanism and gated recurrent network; Step S3: Generate training and test datasets based on sample data, train the model based on the training dataset, set parameters and evaluation metrics, and select different parameters for training to determine the hyperparameter model with the minimum loss on the test set. Step S4: Predict the aircraft trajectory using a hyperparameter model.

2. The method as described in claim 1, characterized in that, Prior to step S1, the method further includes establishing the kinematic and dynamic equations of a six-degree-of-freedom aircraft model, with the origin of the aircraft coordinate system as the aircraft's center of mass, as well as the aircraft's attitude kinematic and dynamic equations.

3. The method as described in claim 1, characterized in that, Step S2 also includes: Step S21: In the encoding stage, the input attention mechanism adaptively selects the relevant state feature sequence, where the encoder is used to encode the original sequence into a model feature representation. Given a k-dimensional sequence feature with a time length of T, the weight of each feature dimension is calculated based on the attention mechanism and the hidden layer information. Step S22: In the decoding phase, the attention mechanism is used to adaptively select the encoded information output by the encoder, where the input to the decoder is the target sequence, hidden state, and context vector from the previous time step.

4. The method as described in claim 3, characterized in that, Step S3 includes: Step S31: Normalize the state feature data using the min-max method; Step S32: Then, the normalized state feature data is divided using the sliding window algorithm to obtain flight trajectory samples. The window size is set to T, where the data at the first T-1 time is used as the model input value, and the data at time T is used as the true value.

5. The method as described in claim 1, characterized in that, Step S3 also includes: Step S33: Set the hyperparameters of the Dual-Attention GRU model so that the hidden state dimension of the encoder GRU unit is equal to the hidden state dimension of the decoder GRU unit; Step S34: The loss function is the mean squared error, and the evaluation metrics include mean absolute error, root mean square error, and mean absolute percentage error.

6. The method as described in claim 5, characterized in that, Also includes: The Adam optimizer is used to train the model hyperparameters with different values ​​based on the loss function.

7. The method according to claim 5, characterized in that, In step S33, the GRU units of the encoder and decoder are set to p = q ∈ {64,128,256,512}.

8. The method as described in claim 1, characterized in that, Also includes: Step S5: Compare the predicted flight trajectory with the actual flight trajectory to determine the degree of overlap.

Images