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

Abstract

The application discloses a kind of based on graph attention hybrid neural network's multi-aircraft 4D trajectory prediction method, belong to aviation technical field.Utilize Transformer and long short-term memory network to extract different aircraft historical trajectory implicit information in airspace, aircraft in airspace is regarded as node and constructs aircraft connection network graph, aircraft implicit information and aircraft connection network graph are modeled by graph attention network, for capturing the interaction information between aircraft.Graph attention network will integrate the trajectory implicit information of all aircraft in airspace and carry out sharing and updating, finally by two long short-term memory network-based decoder, aircraft trajectory implicit feature information is decoded, the output of the fusion of two decoders is output by fully connected layer, predicted trajectory, compared with the method not considering aircraft spatial interaction information, improve the accuracy of flight trajectory prediction, and accurate trajectory prediction technology provides support for subsequent risk perception technology and conflict resolution technology.

Description

TECHNICAL FIELD

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

[0002] The terminal airspace has high traffic flow, high density and high altitude change, and has a great influence on the safety and efficiency of the air traffic management system, so most of the decision support tool researches focus on the safe operation and efficiency of the terminal airspace, and the decision tools (conflict detection and resolution, sequencing and scheduling tools) greatly depend on the accuracy of the four-dimensional trajectory prediction.

[0003] In actual operation, the difficulties of multi-aircraft trajectory prediction in the terminal airspace include: 1) how to capture the interaction between aircrafts in a dynamic environment, such as the dynamic change of the interaction information of two aircrafts at the current timestamp and the exchange information at the future timestamp; 2) it is necessary to simultaneously predict the trajectories of take-off, landing and flat flight in three different flight attitudes, and the model needs to be suitable for trajectory prediction tasks in the three flight states; 3) the coupling of time correlation and space correlation of multi-agent trajectory prediction, for example, the motion of the current aircraft needs to consider the motion of all related aircrafts.

[0004] This paper proposes a multi-aircraft trajectory prediction framework based on deep learning, aiming to realize the tactical (short-term) four-dimensional trajectory prediction of the terminal airspace. The framework is composed 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 features of the aircraft trajectory, and the GAT model is used to encode the spatial interaction features 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 the multi-head attention mechanism. Finally, the local and global features are fused and processed through the full connection layer to realize the accurate, stable and reliable prediction of the 4D flight trajectory. SUMMARY

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

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

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

[0008] S2, differentiating and transforming the data to construct a model training data set;

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

[0010] S4, train the prediction model based on the preprocessed trajectory data, and perform real-time collaborative prediction on the multi-aircraft trajectory based on the trained model.

[0011] As a further improvement of the application, the extracted aircraft historical trajectory data, and the preprocessing of the trajectory data includes:

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

[0013] (2) Data cleaning of the flight path, deleting the flight path with missing points and trajectory points;

[0014] (3) Interpolation of the flight path, making the interval between each flight path point fixed at 5s;

[0015] (4) Convert the processed flight path data into x, y and z coordinates in the geocentric coordinate system, and extract the three-dimensional velocity v x , v y and v z .

[0016] As a further improvement of the application, the data is differentially transformed to construct a model training data set, which includes:

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

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

[0019] As a further improvement of the application, 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 a graph attention hybrid neural network, including

[0021] (1) Trajectory encoder module: According to the input multi-aircraft historical position state information, the trajectory is multiple encoded and feature extracted, and the spatial interaction features between the aircrafts are extracted through the graph attention network, and the historical state implicit information of the multi-aircraft is integrated,

[0022] (2) Trajectory decoder module: The multi-aircraft historical state implicit information extracted by the encoder is decoded, and multiple LSTMs are used as the decoder to map the information to the future trajectory position,

[0023] (3) Trajectory feature fusion prediction module: The output information of the trajectory decoder is integrated, which is input to the neural network linear layer through fusion, so as to output the future trajectory position to achieve the purpose of trajectory prediction.

[0024] As a further improvement of the application, the trajectory encoder module comprises:

[0025] (1) LSTM-GAT encoder: embedding 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 extracts trajectory temporal implicit information through the same LSTM (referred to as M-LSTM) module The implicit information of all aircrafts at each time step is input into the GAT network to obtain feature information considering the spatial interaction between aircrafts The spatial feature information at each time step is input into a new LSTM (referred to as G-LSTM) to explicitly model the time correlation of spatial interaction to obtain spatial interaction feature information considering time sequence Finally, the two LSTM output implicit information is fused to obtain the fusion time and space feature trajectory implicit information at t time

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

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

[0028] (1) D-LSTM decoder: using an LSTM layer to decode the implicit features output by the LSTM-GAT encoder; Output;

[0029] (2) T-LSTM decoder: using an LSTM layer to decode the hidden features Y output by the Transformer encoder m to output.

[0030] As a further improvement of the present application, 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] Wherein, φ5 is a linear layer, which splices the local features and global features of each time period and outputs the prediction results of the relative trajectory position.

[0033] As a further improvement of the present application, 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 results

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

[0037]

[0038] Wherein, is the ground-truth trajectory of the aircraft A i , and is the trajectory predicted by the model.

[0039] (4) Train the proposed graph attention hybrid neural network model by minimizing the sum of the loss functions of all aircrafts in a data set sample. The model can realize accurate 4D trajectory prediction of multiple aircrafts.

[0040] An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the multi-aircraft 4D trajectory prediction method based on the graph attention hybrid neural network when executing the program.

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

[0042] Compared with the prior art, the present application has the following advantages:

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

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

[0045] 3. The encoder module of the proposed method can capture both global and local features of the trajectory, enhancing the ability to express the motion information of the model trajectory.

[0046] 4. Differential trajectory processing was adopted. The model input 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 to allow the aircraft to consider the motion state of other aircraft more when the graph attention network aggregates features to predict the trajectory. Attached Figure Description

[0047] Figure 1 This is a diagram illustrating the overall technical roadmap of the method of the present invention;

[0048] Figure 2 A schematic diagram of trajectory latent feature extraction using a Long Short-Term Memory network;

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

[0050] Figure 4 Transformer encoder structure. Detailed Implementation

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

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

[0053] Step 1: Extract historical aircraft trajectory data. This data includes timestamps, flight path, longitude, latitude, altitude, horizontal speed, vertical speed, heading, and departure and arrival airports. Analyze and clean the extracted aircraft data, removing drift points and tracks with significant gaps. Interpolate the tracks to fix the time interval between each point at 5 seconds.

[0054] The processed track data is then converted into x, y, and z coordinates in a geocentric coordinate system.

[0055]

[0056] where R is the radius of the earth (usually take the average radius of about 6371 km), τ is the longitude, λ is the latitude. At the same time, the three-dimensional component velocity v x , v y and v z are extracted, and six key parts of the time sequence of the track are obtained:

[0057]

[0058] where v represents the component of the speed of aircraft i at time t in each dimension of x, y, z.

[0059] Step 2, build the model training data set, first determine the input data of the model, in order to better describe the motion pattern of the trajectory, we carry out differential processing on the three-dimensional trajectory to obtain the three-dimensional relative trajectory position of the aircraft:

[0060]

[0061] represents the change of aircraft i at time t and the last position in three dimensions. The three-dimensional relative trajectory features and are fused as the input features of the model. Since our input is the historical trajectory information of multiple aircrafts, the input of the model is as follows:

[0062]

[0063] where represents the historical observation trajectory information of aircraft i from t to T obs ; represents the input trajectory information of aircraft i at time t; N(t) represents the number of aircrafts in the airspace at time t. The future relative three-dimensional trajectory position sequence is taken as the output of the model:

[0064]

[0065] where is the future trajectory position of aircraft i from t to T pred ; represents the predicted three-dimensional trajectory position of aircraft i at time t. According to the input and output of the model, the model training data set is constructed:

[0066]

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

[0068] (1) LSTM-GAT encoder: as shown in Figure 2 , first embed the historical trajectory information of the aircraft into a high-dimensional vector to get the high-dimensional representation of the trajectory Then each aircraft in the airspace extracts the trajectory temporal hidden information through the same M-LSTM module:

[0069]

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

[0071]

[0072] Among them, σ is a nonlinear function, is the aggregated hidden state of aircraft i at t time, 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 time correlation of spatial interaction:

[0073]

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

[0075] Finally, the hidden information output by the two LSTM is fused to obtain the fused spatio-temporal feature information at t time:

[0076]

[0077] Among them, φ1(·), φ2(·) are multi-layer perceptron, fusion spatio-temporal feature information vector.

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

[0079]

[0080] The embedding trajectory information P i of the aircraft trajectory is added to the position encoding matrix PE to obtain the position encoded result P i ,K i ,V i matrix:

[0081]

[0082] where f Q , f K , and f V are linear functions and are shared among all aircraft. The Q i ,K i ,V i matrix is input to the multi-head attention mechanism to effectively capture relevant information in the entire sequence, and the output of the multi-head attention mechanism is X m :

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

[0084]

[0085] After obtaining the output of the multi-head attention mechanism, the feedforward network of the Transformer model is composed of a fully connected layer. The output of the model is input to the LSTM layer through the residual normalization 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 represented as:

[0086]

[0087] In the formula, X m represents the input of the normalization layer; γ, b x are the weights and biases that can be learned; μ, δ 2 represent the mean and variance of the input X m , respectively; and ε is a calculation error to prevent the variance from being 0.

[0088] Thus, we can obtain the output of the feedforward network layer and the trajectory hidden feature information of the output of the residual normalization

[0089]

[0090] Step 5, build trajectory decoder module, the decoder is used to generate future trajectory sequence, considering the strong time correlation of future trajectory, two LSTM are used to decode the two hidden information output by LSTM-GAT encoder and Transformer encoder respectively, and the two LSTM decoders are called D-LSTM and T-LSTM respectively.

[0091]

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

[0093]

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

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

[0096]

[0097] where φ5 is a linear layer, is the relative trajectory position at t+1 time, the local feature and global feature of each time period are spliced to output the prediction result of relative trajectory position

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

[0099]

[0100] where, is the real trajectory of aircraft A i , and is the trajectory predicted by the model.

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

[0102] The above merely illustrates the embodiments of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the scope of the claims of the present application.

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 aircraft historical trajectory data, preprocessing the trajectory data; S2, differential transformation is performed on the data to construct a model training data set; S3, a multi-aircraft trajectory prediction model based on a graph attention hybrid neural network is constructed; S4, the prediction model is trained based on the preprocessed trajectory data, and the multi-aircraft trajectory is predicted in real time based on the trained model; Wherein, the graph attention hybrid neural network comprises a trajectory encoder module, a trajectory decoder module and a trajectory feature fusion prediction module; Step S3, constructing a multi-aircraft trajectory prediction model based on a graph attention hybrid neural network, comprising (1) the trajectory encoder module: according to the input of the multi-aircraft historical position state information, the trajectory is multi-encoded and feature extracted, and the spatial interaction features between the aircrafts are extracted through the graph attention network, and the historical state implicit information of the multi-aircraft is integrated, (2) the trajectory decoder module: the multi-aircraft historical state implicit information extracted by the encoder is decoded, and a plurality of LSTMs are used as the decoder to map the information to the future trajectory position, (3) the trajectory feature fusion prediction module: the output information of the trajectory decoder is integrated, which is input into the neural network linear layer through fusion, so as to output the future trajectory position and achieve the purpose of trajectory prediction; The trajectory encoder module of the multi-aircraft 4D trajectory prediction comprises: (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 , and then each aircraft in the airspace extracts the trajectory temporal hidden information through the same M-LSTM module, where, denotes the hidden state of the t-th time step LSTM, are the weights of the M-LSTM, and all the hidden information of the aircraft at each time step will be input into the GAT network to obtain the feature information considering the interaction between the aircrafts: wherein, is a non-linear function, is an aircraft The aggregated hidden state 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 time correlation of spatial interaction, wherein, is the weight of represents the feature vector of node i considering interaction at time t, and finally the fusion of the two LSTM output hidden information is obtained to get the fusion spatio-temporal feature information at time t: wherein, is a multilayer perceptron, fusing the spatiotemporal feature information vectors; (2) Transformer encoder: This module uses a multi-head attention mechanism to extract global deep features of the aircraft historical trajectory state information. First, the historical state information of aircraft i is embedded into an embedding vector by trajectory encoding . The embedding vector is added to the position encoding matrix PE to obtain the position encoded result : Embedded trajectory information for an aircraft trajectory As input to the Transformer encoder module, compute Matrix: wherein , , and are linear functions and shared across all aircrafts, with matrix input to multi-head attention mechanism to effectively capture relevant information in the whole sequence, obtaining the output of multi-head attention : After obtaining the output of the multi-head attention, the feedforward network of the Transformer model is composed of a fully connected layer, and the model outputs it through a residual normalization layer and enters an 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 represented as: In the formula, represents the input of the normalization layer; are weights and biases that can be learned, respectively represent the mean and variance of the input ; is a calculation error to prevent variance from being 0, Thus, the output of the feed-forward network layer and the trajectory-implicit feature information of the output of the residual normalization : ； The trajectory decoder module of the multi-aircraft 4D trajectory prediction: The decoder is used to generate a future trajectory sequence. Considering the strong time 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. The two LSTM decoders are called D-LSTM and T-LSTM respectively, wherein, is the weight of, is a linear layer, is a trajectory local feature, wherein are weights of the T-LSTM, is a linear layer, is a trajectory global feature.

2. The method of claim 1, wherein: Step S1 is as follows: The longitude, latitude, altitude, horizontal speed, vertical speed, heading and timestamp data of the aircraft are extracted, the trajectory data is cleaned, the trajectory points missing trajectory points are deleted, and then the trajectory is interpolated, so that the interval between each trajectory point is fixed at 5s, The processed track data is converted into x coordinate, y coordinate and z coordinate in the geocentric geodetic coordinate system, and the sub-speeds in three dimensions are extracted , and , to obtain six key parts of the track: wherein, denotes the components of the velocity of the aircraft i in the respective dimensions x, y, z at the time t.

3. The method of claim 1, wherein: In step S2, the model training data set is constructed, and the specific steps are as follows; The x, y and z of the trajectory data are differentially processed: , , represents the transformation of the aircraft i in three dimensions at time t from the previous position, The historical trajectory information of multiple aircrafts in the airspace is used as the input of the model, and the future three-dimensional trajectory position sequence is used as the output of the model. The model training data set is constructed according to the input and output of the model.

4. The method of claim 3, wherein: Step of the trajectory feature fusion prediction module: The outputs of the two D-LSTMs and T-LSTMs at each time period are fused in the feature dimension and output the relative three-dimensional position of the future trajectory through a linear layer The relative position is converted into absolute position to obtain the trajectory prediction result , wherein, For the linear layer, the local features and global features of each time period are spliced to output the prediction result of the relative trajectory position.

5. According to the multi-aircraft 4D trajectory prediction method based on the graph attention hybrid neural network of claim 1, in step S4, the prediction model is trained based on the preprocessed trajectory data: First, the network model is initialized, and after the processed training data is input into the model, the predicted three-dimensional trajectory position is obtained and the loss is calculated. The selected loss function calculation formula is: wherein, for an aircraft a true trajectory, a model predicted trajectory, The proposed graph attention hybrid neural network model is trained by minimizing the sum of the loss functions of all aircraft trajectories in the data set. The trained model can input historical aircraft state information to achieve accurate 4D trajectory prediction of multiple aircrafts.

6. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that: The processor implements the multi-aircraft 4D trajectory prediction method based on the graph attention hybrid neural network in any one of claims 1 to 5 when executing the program.

7. A computer readable storage medium having stored thereon computer instructions, characterized in that: The computer instructions implement the multi-aircraft 4D trajectory prediction method based on the graph attention hybrid neural network in any one of claims 1 to 5 when executed by the processor.

Images