Multi-aircraft 4D trajectory prediction method based on graph attention hybrid neural network

Abstract

The invention discloses a multi-aircraft 4D trajectory prediction method based on a graph attention hybrid neural network, and belongs to the technical field of aviation. Historical trajectory implicit information of different aircrafts in an airspace is extracted by utilizing Transform and a long-short-term memory network, the aircrafts in the airspace are regarded as nodes, an aircraft connection network graph is constructed, and modeling is performed on the aircraft implicit information and the aircraft connection network graph through a graph attention network so as to capture interaction information between the aircrafts. The graph attention network integrates the track implicit information of all aircrafts in the airspace and shares and updates the track implicit information, finally decodes the track implicit feature information of the aircrafts through two decoders based on the long and short term memory network, fuses the output of the two decoders, and outputs a prediction track through a full connection layer. Compared with a method which does not consider aircraft space interaction information, the accuracy of flight trajectory prediction is improved, and an accurate trajectory prediction technology provides support for a subsequent risk perception technology and a conflict resolution technology.

Description

Technical Field

[0001] The invention relates to a multi-aircraft 4D trajectory prediction method based on a graph attention hybrid neural network, and belongs to the field of aviation technology. Background Art

[0002] The terminal airspace has the characteristics of high traffic, high density and high variation, which has a great impact on the safety and efficiency of the air traffic management system. Therefore, most of the decision support tool research focuses on the safe operation and efficiency of the terminal airspace. These decision tools (conflict detection and resolution, sorting and scheduling tools) rely heavily on the accuracy of four-dimensional trajectory prediction.

[0003] In actual operation, the difficulties of multi-aircraft trajectory prediction in terminal airspace include: 1) How to capture the interaction between aircraft in a dynamic environment, such as the dynamic change of the interaction information between two aircraft at the current timestamp and the exchange information at the future timestamp; 2) The need to simultaneously predict the trajectory of three different flight postures: take-off, landing, and level flight. The model needs to be able to simultaneously apply the trajectory prediction tasks of these three flight states; 3) The coupling of time correlation and spatial correlation of multi-agent trajectory prediction, for example, the movement of the current aircraft needs to consider the movement of all related aircraft.

[0004] This paper proposes a deep learning-based multi-aircraft trajectory prediction framework, which aims to achieve tactical (short-term) four-dimensional trajectory prediction in the terminal airspace. The framework consists of local feature channels and global feature channels. The local feature channel includes the extraction of spatial and temporal features, in which the LSTM model is used to capture the temporal characteristics of the aircraft trajectory, and the GAT model is used to encode the spatial interaction characteristics of the aircraft at each time step. The global feature channel is based on the Transformer encoder, which effectively captures the global information of the entire trajectory sequence through a multi-head attention mechanism. Finally, the local and global features are fused and processed through a fully connected layer to achieve accurate, stable and reliable prediction of the 4D flight trajectory. Summary of the invention

[0005] The present invention provides a multi-aircraft 4D trajectory prediction method based on a graph attention hybrid neural network, which solves the problems disclosed in the background technology.

[0006] In order to solve the above technical problems, the present invention adopts a multi-aircraft 4D trajectory prediction method based on a graph attention hybrid neural network, comprising the following steps:

[0007] S1, extracting historical trajectory data of aircraft and preprocessing the trajectory data;

[0008] S2, perform differential transformation on the data and construct a model training data set;

[0009] S3. Construct a multi-aircraft trajectory prediction model based on graph attention hybrid neural network;

[0010] S4. Train the prediction model based on the preprocessed trajectory data, and perform real-time collaborative prediction of multiple aircraft trajectories based on the trained model.

[0011] As a further improvement of the present invention, the extracting of aircraft historical trajectory data and preprocessing of the trajectory data comprises:

[0012] (1) Based on the actual operation trajectory data, extract the aircraft's flight number, model, longitude, latitude, altitude, horizontal speed, vertical speed, heading, timestamp and other data;

[0013] (2) Data cleaning of the track is performed to delete tracks with missing drift points and track points;

[0014] (3) Interpolate the track so that the interval between each track point is fixed at 5 s;

[0015] (4) Convert the processed track data into x-coordinates, y-coordinates, and z-coordinates in the Earth-centered Earth-fixed coordinate system, and extract the component velocity v in three dimensions x 、v y and v z .

[0016] As a further improvement of the present invention, the data is differentially transformed, and constructing a model training data set includes:

[0017] (1) Perform differential processing on the three-dimensional coordinates x, y, and z of the aircraft to obtain the relative trajectory position

[0018] (2) Combining the relative trajectory position and the component velocity v in three dimensions x 、v y and v z As the input features of the model, the historical trajectory information of each aircraft in the airspace is used as the input of the model, and the future three-dimensional trajectory position sequence is used as the model output to construct the model training data set.

[0019] As a further improvement of the present invention, the graph attention hybrid neural network includes a trajectory encoder module, a trajectory decoder module and a trajectory feature fusion prediction module;

[0020] Step S3: construct a multi-aircraft trajectory prediction model based on graph attention hybrid neural network, including

[0021] (1) Trajectory encoder module: Based on the input historical position status information of multiple aircraft, the trajectory is multi-encoded and feature extracted, and the spatial interaction features between aircraft are extracted through the graph attention network, integrating the historical status implicit information of multiple aircraft.

[0022] (2) Trajectory decoder module: decodes the implicit information of the historical status of multiple aircraft extracted by the encoder, and uses multiple LSTMs as decoders to map this information to the future trajectory position.

[0023] (3) Trajectory feature fusion prediction module: Integrates the output information of the track decoder. This information is fused and input into the linear layer of the neural network to output the future trajectory position to achieve the purpose of trajectory prediction.

[0024] As a further improvement of the present invention, the track encoder module comprises:

[0025] (1) LSTM-GAT encoder: embeds the historical trajectory information of the aircraft into a high-dimensional vector to obtain a high-dimensional representation of the trajectory Then, each aircraft in the airspace uses the same LSTM module (called M-LSTM) to extract the implicit information of trajectory timing. The implicit information of all aircraft at each time step will be input into the GAT network to obtain feature information that takes into account the spatial interaction between aircraft. The spatial feature information of each time step will be input into a new LSTM (called G-LSTM) to explicitly model the temporal correlation of spatial interactions and obtain spatial interaction feature information that takes time sequence into account. Finally, the implicit information of the two LSTM outputs is fused to obtain the implicit information of the fused time and space feature track at time t.

[0026] (2) Transformer encoder: This module uses a multi-head attention mechanism to extract the global deep features of the aircraft's historical trajectory state information. First, the historical trajectory state information is passed through the embedding layer to obtain an embedding vector And perform position encoding on it. The trajectory information P after position encoding i Input into the multi-head attention mechanism, and then pass through the feedforward network layer and the residual connection normalization layer to obtain the implicit information Y of the global feature of the trajectory m .

[0027] As a further improvement of the present invention, the trajectory decoder module comprises:

[0028] (1) D-LSTM decoder: uses an LSTM layer to decode the implicit features output by the LSTM-GAT encoder. Decode and output;

[0029] (2) T-LSTM decoder: Use an LSTM layer to decode the implicit feature Y output by the Transformer encoder m Decode and output.

[0030] As a further improvement of the present invention, the trajectory feature fusion prediction module is responsible for fusing the outputs of the D-LSTM decoder and the T-LSTM decoder, and outputting the future trajectory through a linear layer:

[0031]

[0032] Among them, φ5 is a linear layer, which concatenates the local features and global features of each time period and outputs the prediction result of the relative trajectory position.

[0033] As a further improvement of the present invention, the steps of model training are as follows:

[0034] (1) Initialize network model parameters;

[0035] (2) Input the processed training data into the constructed prediction model and output the relative trajectory position prediction result

[0036] (3) Convert relative position to absolute position And calculate the loss, the loss function calculation formula is:

[0037]

[0038] in, For aircraft A i The ground-truth trajectory of is the trajectory predicted by the model.

[0039] (4) The proposed graph attention network hybrid neural network model is trained by minimizing the sum of the loss functions of all aircraft in a dataset sample. This model can achieve accurate 4D trajectory prediction of multiple aircraft.

[0040] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for predicting 4D trajectories of multiple aircraft based on a graph attention hybrid neural network is implemented.

[0041] A computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implements the multi-aircraft 4D trajectory prediction method based on a graph attention hybrid neural network

[0042] Compared with the prior art, the advantages of the present invention are as follows:

[0043] 1. Use a graph attention network to capture the interaction between aircraft in a dynamic environment. This structure is more efficient than the Social-LSTM method, without using a separate LSTM model for each aircraft.

[0044] 2. The proposed method can simultaneously predict trajectories for different flight attitudes such as takeoff, landing, and level flight in the terminal area.

[0045] 3. The encoder module of the proposed method can capture the global and local features of the trajectory and enhance the expression ability of the model trajectory motion information.

[0046] 4. Differential processing of trajectories is adopted. The input of the model is the relative displacement of each aircraft relative to the previous moment. The information learned by the model is based on the motion state of the aircraft. This processing is done so that when the graph attention network aggregates features, the aircraft will consider the motion state of other aircraft more to predict the trajectory. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is the overall technical roadmap of the method of the present invention;

[0048] Figure 2 Schematic diagram of long short-term memory network extracting implicit features of trajectories;

[0049] Figure 3 Schematic diagram of spatial interaction feature extraction in graph attention network;

[0050] Figure 4 Transformer encoder structure. DETAILED DESCRIPTION

[0051] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0052] Example: Figure 1 As shown, a multi-aircraft 4D trajectory prediction method based on a graph attention hybrid neural network includes the following steps:

[0053] Step 1: Extract the historical trajectory data of the aircraft, which includes timestamp, flight line, longitude, latitude, altitude, horizontal speed, vertical speed, heading, take-off and landing airports, etc. Analyze and clean the extracted aircraft data, delete the drift points and tracks with a large number of missing track points. Interpolate the track and fix the time interval of each track point at 5s.

[0054] Convert the processed track data into x-coordinates, y-coordinates, and z-coordinates in the Earth-centered Earth-fixed coordinate system:

[0055]

[0056] Among them, R is the radius of the earth (usually the average radius is about 6371 kilometers), τ is the longitude, and λ is the latitude. At the same time, the component speed v in three dimensions is extracted x 、v y and v z , we get six key parts of the track time series:

[0057]

[0058] in, Represents the components of the velocity of aircraft i at time t in the x, y, and z dimensions.

[0059] Step 2: Build a model training data set. First, we need to determine the input data of the model. In order to better describe the motion pattern of the trajectory, we perform differential processing on the three-dimensional trajectory to obtain the three-dimensional relative trajectory position of the aircraft:

[0060]

[0061] represents the transformation of aircraft i in three dimensions from its previous position at time t. Fusion is used as the input feature of the model. Since our input is the historical trajectory information of multiple aircraft, the input of the model is as follows:

[0062]

[0063] in represents the distance between aircraft i and T obs Historical observation trajectory information over time; represents the input trajectory information of aircraft i at time t; N(t) represents the number of aircraft in the airspace at time t. The future relative three-dimensional trajectory position sequence is used as the model output:

[0064]

[0065] in For aircraft i at t to T pred Future trajectory position in time; Represents the predicted three-dimensional trajectory position of aircraft i at time t. The model training data set is constructed based on the input and output of the model:

[0066]

[0067] Step 3, build the trajectory encoder module, which is divided into LSTM-GAT encoder and Transformer encoder. We will introduce the specific structures of the two encoders respectively:

[0068] (1) LSTM-GAT encoder: Figure 2 As shown, firstly, the historical trajectory information of the aircraft is embedded into the high-dimensional vector to obtain the high-dimensional representation of the trajectory Then, each aircraft in the airspace uses the same M-LSTM module to extract the trajectory time series implicit information:

[0069]

[0070] in, represents the hidden state of LSTM at time step t, W m is the weight of M-LSTM. Figure 3 As shown in Figure 2, the implicit information of all aircraft at each time step will be input into the GAT network to obtain feature information that considers the interaction between aircraft:

[0071]

[0072] Among them, σ is a nonlinear function, is the aggregated hidden state of aircraft i at time t, which contains the spatial influence of other aircraft. The spatial feature information at each time step will be input into a new LSTM (called G-LSTM) to explicitly model the temporal correlation of spatial interactions:

[0073]

[0074] Among them, W g is the weight of G-LSTM, Represents the feature vector of node i considering interaction at time t.

[0075] Finally, the implicit information of the two LSTM outputs is fused to obtain the fused spatiotemporal feature information at time t:

[0076]

[0077] Among them, φ1(·), φ2(·) are multi-layer perceptrons. Fusion of spatiotemporal feature information vectors.

[0078] (2) Transformer encoder: This module uses a multi-head attention mechanism to extract global deep features of the aircraft’s historical trajectory state information. Figure 4 The schematic diagram of the Transformer encoder structure is as follows. First, the historical state information (relative position and velocity components) of aircraft i is embedded through trajectory encoding to obtain the embedding vector Embedding vector Add the position encoding matrix PE to get the position encoded result P i :

[0079]

[0080] Embed the trajectory information P of the aircraft trajectory i As the input of the Transformer encoder module, calculate Q i ,K i ,V i matrix:

[0081]

[0082] Among them, f Q , f K , and f V is a linear function and is shared among all aircraft. i ,K i ,V i The matrix is ​​input into the multi-head attention mechanism to effectively capture the relevant information in the overall sequence and obtain the output X of the multi-head attention. m :

[0083] MultiHead(Q,K,V)=Concat(head1,...,head h )

[0084]

[0085] After obtaining the output of the multi-head attention, the feedforward network of the Transformer model is composed of a fully connected layer. This model passes its output through the residual normalization layer and then enters the LSTM layer, which plays a role in integrating the information of the self-attention mechanism layer. The calculation process of the feedforward network layer FFN is expressed as:

[0086]

[0087] In the formula, X m represents the input of the normalization layer; γ,b x are the learnable weights and biases, μ,δ 2 Respectively represent the input X m The mean and variance of ; ε is used to prevent calculation errors caused by the variance being 0.

[0088] From this we can get the output of the feedforward network layer The output trajectory of the residual normalization contains implicit feature information

[0089]

[0090] Step 5: Construct a trajectory decoder module. The decoder is used to generate future trajectory sequences. Considering that future trajectories have strong temporal correlation, two LSTMs are used to decode the two implicit information output by the LSTM-GAT encoder and the Transformer encoder respectively. The two LSTM decoders are called D-LSTM and T-LSTM respectively.

[0091]

[0092] Among them, W d is the weight of D-LSTM, φ3 is the linear layer, is the local feature of the trajectory.

[0093]

[0094] Where W o is the weight of D-LSTM, φ4 is the linear layer, is the global feature of the trajectory.

[0095] Step 6: Construct a trajectory feature fusion prediction module, which fuses the outputs of the two DLSTMs and TLSTMs in the feature dimension and outputs the future trajectory prediction through a linear layer:

[0096]

[0097] Among them, φ5 is the linear layer, The relative trajectory position at time t+1 is obtained by splicing the local features and global features of each time period to output the prediction result of the relative trajectory position.

[0098] Step 7: Initialize the network model first, input the processed training data into the model to obtain the relative trajectory position prediction result The relative position Convert to absolute position And calculate the loss, the selected loss function calculation formula is:

[0099]

[0100] in, For aircraft A i The true trajectory of is the trajectory predicted by the model.

[0101] The proposed graph attention hybrid neural network model is trained by minimizing the sum of the loss functions of all aircraft trajectories in the dataset. The trained model can input historical aircraft status information to achieve accurate prediction of the 4D trajectories of multiple aircraft.

[0102] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are included in the scope of the claims of the present invention to be approved.

Claims

1. A multi-aircraft 4D trajectory prediction method based on graph attention hybrid neural network, characterized in that: The method comprises the following steps: S1, extracting historical trajectory data of aircraft and preprocessing the trajectory data; S2, perform differential transformation on the data and construct a model training data set; S3. Construct a multi-aircraft trajectory prediction model based on graph attention hybrid neural network; S4. Train the prediction model based on the preprocessed trajectory data, and perform real-time collaborative prediction of multiple aircraft trajectories based on the trained model.

2. The method for multi-aircraft 4D trajectory prediction based on graph attention hybrid neural network according to claim 1, characterized in that: Step S1 is specifically as follows: The longitude, latitude, altitude, horizontal speed, vertical speed, heading, and timestamp data of the aircraft were extracted, and the track data was cleaned to delete the drift points and tracks with missing track points. Then the track was interpolated to make the interval between each track point fixed at 5s. The processed track data is converted into the x-coordinate, y-coordinate and z-coordinate in the Earth-centered Earth-fixed coordinate system, and the component velocity v in three dimensions is extracted at the same time. x 、v y and v z , we get six key parts of the track: in, Represents the components of the velocity of aircraft i at time t in the x, y, and z dimensions.

3. The method for multi-aircraft 4D trajectory prediction based on graph attention hybrid neural network according to claim 1, characterized in that: In step S2, a model training data set is constructed, and the specific steps are as follows; Perform differential processing on the x, y, and z of the track data: represents the transformation of aircraft i in three dimensions from its previous position at time t, The historical trajectory information of multiple aircraft in the airspace is used as the input of the model, and the future three-dimensional trajectory position sequence is used as the model output. The model training data set is constructed based on the input and output of the model.

4. The method for multi-aircraft 4D trajectory prediction based on graph attention hybrid neural network according to claim 1, characterized in that: The graph attention hybrid neural network includes a trajectory encoder module, a trajectory decoder module and a trajectory feature fusion prediction module; Step S3: construct a multi-aircraft trajectory prediction model based on graph attention hybrid neural network, including (1) Trajectory encoder module: Based on the input historical position status information of multiple aircraft, the trajectory is multi-encoded and feature extracted, and the spatial interaction features between aircraft are extracted through the graph attention network, integrating the historical status implicit information of multiple aircraft. (2) Trajectory decoder module: decodes the implicit information of the historical status of multiple aircraft extracted by the encoder, and uses multiple LSTMs as decoders to map this information to the future trajectory position. (3) Trajectory feature fusion prediction module: Integrates the output information of the track decoder. This information is fused and input into the linear layer of the neural network to output the future trajectory position to achieve the purpose of trajectory prediction.

5. The method for multi-aircraft 4D trajectory prediction based on graph attention hybrid neural network according to claim 4, characterized in that: The trajectory encoder module for multi-aircraft 4D trajectory prediction includes: (1) LSTM-GAT encoder: First, the historical trajectory information of the aircraft is embedded into a high-dimensional vector to obtain a high-dimensional representation of the trajectory. Then, each aircraft in the airspace uses the same M-LSTM module to extract the trajectory time series implicit information: in, represents the hidden state of LSTM at time step t, W m The weights of M-LSTM, the implicit information of all aircraft at each time step will be input into the GAT network to obtain feature information that considers the interaction between aircraft: Among them, σ is a nonlinear function, is the aggregated hidden state of aircraft i at time t, which contains the spatial influence of other aircraft. The spatial feature information of each time step will be input into a new LSTM (called G-LSTM) to explicitly model the temporal correlation of spatial interactions. Among them, W g is the weight of G-LSTM, It represents the feature vector of node i considering interaction at time t. Finally, the implicit information of the two LSTM outputs is fused to obtain the fused spatiotemporal feature information at time t: Among them, φ1(·), φ2(·) are multi-layer perceptrons. Fusion of spatiotemporal feature information vectors; (2) Transformer encoder: This module uses a multi-head attention mechanism to extract the global deep features of the aircraft's historical trajectory state information. First, the historical state information of aircraft i is embedded through trajectory encoding to obtain an embedding vector Embedding vector Add the position encoding matrix PE to get the position encoded result P i : Embed the trajectory information P of the aircraft trajectory i As the input of the Transformer encoder module, calculate Q i ,K i ,V i matrix: Among them, f Q , f K , and f V is a linear function and is shared among all aircraft, Q i ,K i ,V i The matrix is ​​input into the multi-head attention mechanism to effectively capture the relevant information in the overall sequence and obtain the output X of the multi-head attention. m : MultiHead(Q,K,V)=Concat(head1,...,head h ) head i =Attention(QW i Q ,KW i K ,VW i V ) After obtaining the output of the multi-head attention, the feedforward network of the Transformer model is composed of a fully connected layer. The model passes its output through the residual normalization layer and then enters the LSTM layer, which plays a role in integrating the information of the self-attention mechanism layer. The calculation process of the feedforward network layer FFN is expressed as: In the formula, X m represents the input of the normalization layer; γ,b x are the learnable weights and biases, μ,δ 2 Respectively represent the input X m The mean and variance of ; ε is to prevent calculation errors caused by variance 0, Thus, the output of the feedforward network layer is The output trajectory of the residual normalization contains implicit feature information 6. The method for multi-aircraft 4D trajectory prediction based on graph attention hybrid neural network according to claim 4, characterized in that: Trajectory decoder module for multi-aircraft 4D trajectory prediction: The decoder is used to generate the future trajectory sequence. Considering the strong temporal correlation of the future trajectory, two LSTMs are used to decode the two implicit information output by the LSTM-GAT encoder and the Transformer encoder respectively. The two LSTM decoders are called D-LSTM and T-LSTM respectively. Among them, W d is the weight of D-LSTM, φ3 is the linear layer, is the local feature of the trajectory, Where W o is the weight of D-LSTM, φ4 is the linear layer, is the global feature of the trajectory.

7. The method for multi-aircraft 4D trajectory prediction based on graph attention hybrid neural network according to claim 4, characterized in that: Trajectory feature fusion prediction module steps: The outputs of the two D-LSTMs and T-LSTMs at each time period are fused in the feature dimension and the relative three-dimensional position of the future track is output through a linear layer. Convert the relative position to the absolute position to get the trajectory prediction result Among them, φ5 is a linear layer, which concatenates the local features and global features of each time period and outputs the prediction result of the relative trajectory position.

8. According to the method for multi-aircraft 4D trajectory prediction based on graph attention hybrid neural network in claim 1, in step S4, the prediction model is trained based on the preprocessed trajectory data: First, initialize the network model, input the processed training data into the model to obtain the predicted three-dimensional trajectory position and calculate the loss. The selected loss function calculation formula is: in, For aircraft A i The true trajectory of is the trajectory predicted by the model, The proposed graph attention hybrid neural network model is trained by minimizing the sum of the loss functions of all aircraft trajectories in the dataset. The trained model can input historical aircraft status information to achieve accurate prediction of the 4D trajectories of multiple aircraft.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the multi-aircraft 4D trajectory prediction method based on graph attention hybrid neural network as described in any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by a processor, the multi-aircraft 4D trajectory prediction method based on a graph attention hybrid neural network as described in any one of claims 1 to 8 is implemented.

Images