A Reliable Prediction Method and Apparatus for Multi-Aircraft Tracks Based on Machine Learning

Abstract

A machine learning-based method and apparatus for reliable prediction of multi-aircraft tracks can quickly predict the tracks of surrounding aircraft in the future, providing strong support for various tasks in the field of air traffic flow management, such as arrival and departure sequencing, conflict detection, flight flow management, and arrival time estimation. The method includes: (1) performing data preprocessing and constructing a multi-aircraft track dataset, converting data from multiple complex scenarios into a fixed-length track vector sequence; (2) constructing a reliable prediction model for multi-aircraft tracks, building a framework based on the spatiotemporal correlation of multi-aircraft tracks, which includes a spatial feature extraction module and a long short-term memory network (LSTM) module. The spatial feature extraction module enables the interaction of adjacent agents in the modeling space, and the LSTM module captures the temporal dependency of a single aircraft's 4D track; (3) training with historical data and using the reliable prediction model for multi-aircraft tracks to make reliable predictions of multi-aircraft tracks.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent air traffic management in civil aviation, and particularly to a reliable prediction method for multi-aircraft trajectories based on machine learning, as well as a reliable prediction device for multi-aircraft trajectories based on machine learning. Background Technology

[0002] Whether it's the current sector-based operation mode or the future trajectory-based operation (TBO) mode, reliable aircraft trajectory prediction is crucial for the efficient operation of both modes. Simultaneously, accurate reliable trajectory prediction is a key technology for various applications in the field of air traffic management (ATM), such as arrival and departure sequencing, conflict detection, flight flow management, and time of arrival estimation. It can significantly reduce the uncertainty of future aircraft flights and improve the predictability of air traffic.

[0003] Due to airspace congestion, weather, temporary military activities, and airspace restrictions, aircraft sometimes deviate from their pre-planned flight paths. This reduces the predictability of air traffic, increases the difficulty of air traffic management, and seriously hinders the efficient operation of air traffic management systems. Aircraft are also affected by other adjacent aircraft traveling at the same time. For example, controllers sometimes direct aircraft to deviate from their planned routes to alleviate airspace congestion. Such trajectory deviations cannot be predicted by observing a single aircraft; it is necessary to consider other aircraft in the vicinity that are most likely to affect it.

[0004] The difficulties in reliable prediction of multi-aircraft trajectories are: (1) It is difficult to capture the interaction between agents in a dynamic environment. For example, two agents may interact in the current time step but not in the future; (2) The action space of multiple agents is significantly larger than that of a single agent. For example, in each step, the action of one agent affects the action of its neighboring agents; (3) Temporal correlation and spatial interaction are coupled. For example, the time series motion of the current agent needs to take into account the motion of all its neighbors. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the technical problem to be solved by this invention is to provide a reliable prediction method for multi-aircraft tracks based on machine learning. This method can quickly and reliably predict the tracks of surrounding aircraft in the future, providing strong support for various tasks in the field of air traffic flow management, such as arrival and departure sequencing, conflict detection, flight flow management, and arrival time estimation.

[0006] The technical solution of this invention is: a reliable prediction method for multi-aircraft tracks based on machine learning, which includes the following steps:

[0007] (1) Perform data preprocessing and construct a multi-aircraft trajectory dataset to convert data from multiple complex scenarios into a fixed-length trajectory vector sequence to adapt to the input format of the machine learning model;

[0008] (2) Construct a reliable prediction model for multi-aircraft tracks. Based on the spatiotemporal correlation of multi-aircraft tracks, construct a framework. This framework includes a spatial feature extraction module and a long short-term memory network (LSTM) module. The spatial feature extraction module enables the interaction of adjacent agents in the modeling space, and the LSTM module captures the temporal dependency of a single aircraft's 4D track.

[0009] (3) Use historical data for training and save the trained model data. Then use the multi-aircraft trajectory reliable prediction model to make reliable predictions of multi-aircraft trajectories.

[0010] This invention constructs a reliable prediction model for multi-aircraft trajectories based on the spatiotemporal correlation of multi-aircraft trajectories. It can capture the action patterns of individual aircraft and the interactions of agents in congested scenarios using a data-driven approach, exhibiting strong generalization capabilities. It can rapidly and reliably predict multi-aircraft trajectories with a certain level of confidence and confidence interval, meeting real-time requirements and providing strong support for various tasks in air traffic flow management, such as arrival and departure sequencing, conflict detection, flight flow management, and arrival time estimation.

[0011] A machine learning-based reliable prediction device for multi-aircraft trajectories is also provided, comprising:

[0012] The data preprocessing module is configured to preprocess data and construct a multi-aircraft trajectory dataset, converting data from multiple complex scenarios into a fixed-length trajectory vector sequence to adapt to the input format of machine learning models.

[0013] The model building module is configured to build a reliable prediction model for multi-aircraft tracks. Based on the spatiotemporal correlation of multi-aircraft tracks, a framework is built, which includes a spatial feature extraction module and a Long Short-Term Memory (LSTM) network module. The spatial feature extraction module enables the interaction of neighboring agents in the modeling space, and the LSTM module captures the temporal dependencies of a single aircraft's 4D track.

[0014] The training module is configured to train the model using historical data and save the trained model data. Then, the multi-aircraft trajectory reliability prediction model is used to make reliable predictions of multi-aircraft trajectories. Attached Figure Description

[0015] Figure 1 The diagram shown is an overall flowchart of the machine learning-based reliable prediction method for multi-aircraft tracks according to the present invention.

[0016] Figure 2 The flowchart shown is a step (2) of the machine learning-based reliable prediction method for multi-aircraft tracks according to the present invention.

[0017] Figure 3 The diagram shows a flowchart of the interaction between neighboring agents in the modeling space of step (2) of the machine learning-based multi-aircraft trajectory reliability prediction method according to the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] To make the description of this disclosure more detailed and complete, illustrative descriptions of embodiments and specific examples of the present invention are provided below; however, these are not the only forms of implementing or utilizing the specific examples of the present invention. The embodiments cover features of multiple specific examples and methods and steps for constructing and operating these specific examples, and their order. However, other specific examples may also be used to achieve the same or equivalent functions and order of steps.

[0020] like Figure 1 As shown, this machine learning-based method for reliable prediction of multi-aircraft tracks includes the following steps:

[0021] (1) Perform data preprocessing and construct a multi-aircraft trajectory dataset to convert data from multiple complex scenarios into a fixed-length trajectory vector sequence to adapt to the input format of the machine learning model;

[0022] (2) Construct a reliable prediction model for multi-aircraft tracks. Based on the spatiotemporal correlation of multi-aircraft tracks, construct a framework. This framework includes a spatial feature extraction module and a long short-term memory network (LSTM) module. The spatial feature extraction module enables the interaction of adjacent agents in the modeling space, and the LSTM module captures the temporal dependency of a single aircraft's 4D track.

[0023] (3) Use historical data for training and save the trained model data. Then use the multi-aircraft trajectory reliable prediction model to make reliable predictions of multi-aircraft trajectories.

[0024] This invention constructs a reliable prediction model for multi-aircraft trajectories based on the spatiotemporal correlation of multi-aircraft trajectories. It can capture the action patterns of individual aircraft and the interactions of agents in congested scenarios using a data-driven approach, exhibiting strong generalization capabilities. It can rapidly and reliably predict multi-aircraft trajectories with a certain level of confidence and confidence interval, meeting real-time requirements and providing strong support for various tasks in air traffic flow management, such as arrival and departure sequencing, conflict detection, flight flow management, and arrival time estimation.

[0025] Preferably, step (1) includes the following sub-steps:

[0026] (1.1) Construct a 4D trajectory dataset of M aircraft, where M is the number of trajectory samples. Divide the trajectory data of multiple aircraft in the same scene into a sequence of trajectory vectors with a time interval of t, a length of L, and a maximum number of aircraft of N. The trajectory vectors include: aircraft ID, longitude, latitude, heading, and ground speed.

[0027] (1.2) Normalize the data of all track samples and scale them to the range of (0, 1) using the Min-Max normalization method;

[0028] (1.3) Select M track samples arbitrarily to construct a 4D track dataset of M multi-aircraft.

[0029] Preferably, the spatial feature extraction module in step (2) enables the interaction of adjacent agents in the modeling space, including:

[0030] The spatial grid size is set to represent the speed and heading adjustments made by the aircraft due to the situational awareness of surrounding aircraft within a certain spatial distance. The number of spatial grids, N0, represents the accuracy of the aircraft's situational awareness of surrounding aircraft. For each aircraft, the spatial grid to which other aircraft belong is calculated based on the positions of other aircraft relative to the aircraft, according to the following indicator function:

[0031]

[0032] When the position of aircraft j is within the (m,n) grid of aircraft i, The value of is 1, and at each time t, the hidden state h of the LSTM is... i t Capture the latent representation of aircraft i; construct the Social tensor This representation is used to share among adjacent aircraft to model the spatial interaction relationships between aircraft at the current moment, given the hidden state dimension D and the number of spatial grids N0. It is a tensor of size N0×N0×D, and its calculation formula is as follows:

[0033]

[0034] in This is the hidden state of aircraft j at time t-1. It is the indicator function of aircraft j relative to aircraft i in the spatial grid of (m,n).

[0035] Preferably, the LSTM module in step (2) captures the time dependency of a single aircraft's 4D trajectory, including:

[0036] The trajectory vector of length 5 obtained in step (1) is input into an embedding layer to obtain the trajectory feature vector. The embedding layer consists of a fully connected layer and a ReLU nonlinear layer, as shown in the following formula:

[0037]

[0038] Among them W e The weights of the fully connected layer are used to input the resulting Social tensor into an embedding layer to obtain a spatial feature vector. The embedding layer consists of a fully connected layer and a ReLU nonlinear layer, as shown in the following formula:

[0039]

[0040] Among them W e The weights of the fully connected layer are used to concatenate the aforementioned track feature vectors and spatial feature vectors, resulting in a vector of length 128, which serves as the input vector for the LSTM module. To Bayesianize this module, a dropout mask, identical for each aircraft and each timestamp, is set up to randomly discard a portion of the neurons representing the input vector. The LSTM module consists of a single-layer LSTM unit, a dropout applied to the LSTM parameters, and a fully connected output layer. All aircraft share a single LSTM module to capture the temporal correlation of each track, ultimately obtaining a probabilistic representation of the prediction result. One cycle of the LSTM is expressed by the formula:

[0041]

[0042] in, These represent the hidden layer state and cell state of the LSTM at time step t, and the hidden layer state and cell state of the LSTM at time step t-1, respectively.

[0043] Preferably, in the training phase of step (3), a large dataset of flight paths of multiple aircraft is prepared and divided into a training set, a validation set, and a test set in a 5:1:1 ratio. First, the model is trained on the training set, with one data point trained per round. Initially, due to the large range of the solution space, a large learning rate is adopted, and then the learning rate decay method is used to gradually reduce the learning rate to obtain stable training. Every few rounds, the model is forward propagated using the validation set. If the model loss reaches a certain threshold, the training is considered successful, and the current neural network parameters are saved. The scalability of the model is preserved, and the model is iteratively upgraded as historical data becomes richer and accumulated.

[0044] Preferably, in the testing phase of step (3), the saved neural network model is used to test in the test set, the neural network is forward propagated n times, the mean and variance of the n results are calculated, and a predicted trajectory with a certain confidence level and confidence interval is quickly obtained.

[0045] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium. When executed, the program includes the steps of the methods of the above embodiments. The storage medium can be ROM / RAM, magnetic disk, optical disk, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a machine learning-based multi-aircraft trajectory reliability prediction device, which is typically represented in the form of functional modules corresponding to the steps of the method. The device includes:

[0046] The data preprocessing module is configured to preprocess data and construct a multi-aircraft trajectory dataset, converting data from multiple complex scenarios into a fixed-length trajectory vector sequence to adapt to the input format of machine learning models.

[0047] The model building module is configured to build a reliable prediction model for multi-aircraft tracks. Based on the spatiotemporal correlation of multi-aircraft tracks, a framework is built, which includes a spatial feature extraction module and a Long Short-Term Memory (LSTM) network module. The spatial feature extraction module enables the interaction of neighboring agents in the modeling space, and the LSTM module captures the temporal dependencies of a single aircraft's 4D track.

[0048] The training module is configured to train the model using historical data and save the trained model data. Then, the multi-aircraft trajectory reliability prediction model is used to make reliable predictions of multi-aircraft trajectories.

[0049] Preferably, the spatial feature extraction module of the model building module enables the interaction of adjacent agents in the modeling space, including:

[0050] The spatial grid size is set to represent the speed and heading adjustments made by the aircraft due to the situational awareness of surrounding aircraft within a certain spatial distance. The number of spatial grids, N0, represents the accuracy of the aircraft's situational awareness of surrounding aircraft. For each aircraft, the spatial grid to which other aircraft belong is calculated based on the positions of other aircraft relative to the aircraft, according to the following indicator function:

[0051]

[0052] When the position of aircraft j is within the (m,n) grid of aircraft i, The value of is 1, representing the hidden state of the LSTM at each time t. Capture the latent representation of aircraft i; construct the Social tensor This representation is used to share among adjacent aircraft to model the spatial interaction relationships between aircraft at the current moment, given the hidden state dimension D and the number of spatial grids N0. It is a tensor of size N0×N0×D, and its calculation formula is as follows:

[0053]

[0054] in It is the hidden state of aircraft j at time t-1.

[0055] Preferably, the LSTM module in step (2) captures the time dependency of a single aircraft's 4D trajectory, including:

[0056] The trajectory vector of length 5 obtained in step (1) is input into an embedding layer to obtain the trajectory feature vector. The embedding layer consists of a fully connected layer and a ReLU nonlinear layer, as shown in the following formula:

[0057]

[0058] Among them W e The weights of the fully connected layer are used to input the resulting Social tensor into an embedding layer to obtain a spatial feature vector. The embedding layer consists of a fully connected layer and a ReLU nonlinear layer, as shown in the following formula:

[0059]

[0060] Among them W eThe weights of the fully connected layer are used to concatenate the aforementioned track feature vectors and spatial feature vectors, resulting in a vector of length 128, which serves as the input vector for the LSTM module. To Bayesianize this module, a dropout mask, identical for each aircraft and each timestamp, is set up to randomly discard a portion of the neurons representing the input vector. The LSTM module consists of a single-layer LSTM unit, a dropout applied to the LSTM parameters, and a fully connected output layer. All aircraft share a single LSTM module to capture the temporal correlation of each track, ultimately obtaining a probabilistic representation of the prediction result. One cycle of the LSTM is expressed by the formula:

[0061]

[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A reliable prediction method for multi-aircraft trajectories based on machine learning, characterized in that: It includes the following steps: (1) Perform data preprocessing and construct a multi-aircraft trajectory dataset to convert data from multiple complex scenarios into a fixed-length trajectory vector sequence to adapt to the input format of the machine learning model; (2) Construct a reliable prediction model for multi-aircraft tracks. Based on the spatiotemporal correlation of multi-aircraft tracks, construct a framework. This framework includes a spatial feature extraction module and a long short-term memory network (LSTM) module. The spatial feature extraction module enables the interaction of adjacent agents in the modeling space, and the LSTM module captures the temporal dependency of a single aircraft's 4D track. (3) Use historical data for training, save the trained model data, and then use the multi-aircraft trajectory reliable prediction model to make reliable predictions of multi-aircraft trajectories. The spatial feature extraction module in step (2) enables the interaction of adjacent agents in the modeling space, including: The spatial grid size is set to represent the adjustments the aircraft makes to its speed and heading based on the situation of surrounding aircraft within a certain spatial distance, as well as the number of spatial grids. This represents the accuracy with which the aircraft perceives the situation of surrounding aircraft. For each aircraft, the spatial grid to which other aircraft belong is calculated based on the positions of other aircraft relative to the aircraft, according to the following indicator function: ， When aircraft j is at the same position as aircraft i Within the grid, The value of is 1, and at each time t, the hidden state of the LSTM is... Capture the latent representation of aircraft i; construct the Social tensor This representation is used to share among adjacent aircraft to model the spatial interactions between aircraft at the current moment, given the hidden state dimension D and the number of spatial grids. , It is a size of The formula for calculating the tensor is as follows: ， in This is the hidden state of aircraft j at time t-1. It is the indicator function of aircraft j relative to aircraft i in the spatial grid of (m,n); The LSTM module in step (2) captures the time dependency of a single aircraft's 4D trajectory, including: The trajectory vector of length 5 obtained in step (1) is input into an embedding layer to obtain the trajectory feature vector. The embedding layer includes a fully connected layer and a ReLU nonlinear layer, which is expressed by the following formula: ， in The weights of the fully connected layer are used to input the resulting Social tensor into an embedding layer to obtain a spatial feature vector. The embedding layer consists of a fully connected layer and a ReLU nonlinear layer, as shown in the following formula: ， in The weights of the fully connected layer are used to concatenate the aforementioned track feature vectors and spatial feature vectors, resulting in a vector of length 128, which serves as the input vector for the LSTM module. To Bayesianize this module, a dropout mask, identical for each aircraft and each timestamp, is set up to randomly discard a portion of the neurons representing the input vector. The LSTM module consists of a single-layer LSTM unit, a dropout applied to the LSTM parameters, and a fully connected output layer. All aircraft share a single LSTM module to capture the temporal correlation of each track, ultimately obtaining a probabilistic representation of the prediction result. One cycle of the LSTM is expressed by the formula: ， in, These represent the hidden layer state and cell state of the LSTM at time step t, and the hidden layer state and cell state of the LSTM at time step t-1, respectively.

2. The reliable prediction method for multi-aircraft tracks based on machine learning according to claim 1, characterized in that: Step (1) includes the following sub-steps: (1.1) Construct a 4D track dataset of M aircraft, where M is the number of track samples. Divide the track data of multiple aircraft in the same scene into a sequence of track vectors with a time interval of t, a length of L, and a maximum number of aircraft of N. The track vectors include: aircraft ID, longitude, latitude, heading, and ground speed. (1.2) Normalize the data of all track samples and scale them to the range of (0, 1) using the Min-Max normalization method; (1.3) Select M track samples arbitrarily to construct a 4D track dataset of M multi-aircraft.

3. The reliable prediction method for multi-aircraft tracks based on machine learning according to claim 2, characterized in that: In the training phase of step (3), a large dataset of flight paths of multiple aircraft is prepared and divided into training, validation, and test sets in a 5:1:1 ratio. The model is first trained on the training set, with one data point trained per round. Initially, a large learning rate is adopted due to the large range of the solution space. Then, the learning rate is gradually reduced by using the learning rate decay method to obtain stable training. Every few rounds, the model is forward propagated using the validation set. If the model loss reaches a certain threshold, the training is considered successful, and the current neural network parameters are saved. The scalability of the model is preserved, and the model is iteratively upgraded as historical data becomes richer and more accumulated.

4. The reliable prediction method for multi-aircraft tracks based on machine learning according to claim 3, characterized in that: In the testing phase of step (3), the saved neural network model is tested in the test set. The neural network is forward propagated n times, and the mean and variance of the n results are calculated to quickly obtain a predicted trajectory with a certain confidence level and confidence interval.

5. A machine learning-based multi-aircraft trajectory reliability prediction device, used to execute the machine learning-based multi-aircraft trajectory reliability prediction method according to claim 1, characterized in that: It includes: The data preprocessing module is configured to preprocess data and construct a multi-aircraft trajectory dataset, converting data from multiple complex scenarios into a fixed-length trajectory vector sequence to adapt to the input format of machine learning models. The model building module is configured to build a reliable prediction model for multi-aircraft tracks. Based on the spatiotemporal correlation of multi-aircraft tracks, a framework is built, which includes a spatial feature extraction module and a Long Short-Term Memory (LSTM) network module. The spatial feature extraction module enables the interaction of adjacent agents in the modeling space, and the LSTM module captures the temporal dependencies of a single aircraft's 4D track. The training module is configured to train the model using historical data and save the trained model data. Then, the multi-aircraft trajectory reliability prediction model is used to make reliable predictions of multi-aircraft trajectories.

6. The machine learning-based multi-aircraft trajectory reliable prediction device according to claim 5, characterized in that: The data preprocessing module performs the following steps: (1.1) Construct a 4D track dataset of M aircraft, where M is the number of track samples. Divide the track data of multiple aircraft in the same scene into a sequence of track vectors with a time interval of t, a length of L, and a maximum number of aircraft of N. The track vectors include: aircraft ID, longitude, latitude, heading, and ground speed. (1.2) Normalize the data of all track samples and scale them to the range of (0, 1) using the Min-Max normalization method; (1.3) Select M track samples arbitrarily to construct a 4D track dataset of M multi-aircraft.

7. The machine learning-based multi-aircraft trajectory reliable prediction device according to claim 6, characterized in that: The spatial feature extraction module of the model building module enables the interaction of adjacent agents in the modeling space, including: The spatial grid size is set to represent the adjustments the aircraft makes to its speed and heading based on the situation of surrounding aircraft within a certain spatial distance, as well as the number of spatial grids. This represents the accuracy with which the aircraft perceives the situation of surrounding aircraft. For each aircraft, the spatial grid to which other aircraft belong is calculated based on the positions of other aircraft relative to the aircraft, according to the following indicator function: ， When aircraft j is at the same position as aircraft i Within the grid, The value of is 1, and at each time t, the hidden state of the LSTM is... Capture the latent representation of aircraft i; construct the Social tensor This representation is used to share among adjacent aircraft to model the spatial interactions between aircraft at the current moment, given the hidden state dimension D and the number of spatial grids. , It is a size of The formula for calculating the tensor is as follows: ， in It is the hidden state of aircraft j at time t-1.

8. The machine learning-based multi-aircraft trajectory reliable prediction device according to claim 7, characterized in that: The LSTM module captures the temporal dependencies of a single aircraft's 4D trajectory, including: The trajectory vector of length 5 obtained in step (1) is input into an embedding layer to obtain the trajectory feature vector. The embedding layer includes a fully connected layer and a ReLU nonlinear layer, which is expressed by the following formula: ， in The weights of the fully connected layer are used to input the resulting Social tensor into an embedding layer to obtain a spatial feature vector. The embedding layer consists of a fully connected layer and a ReLU nonlinear layer, as shown in the following formula: ， in The weights of the fully connected layer are used to concatenate the aforementioned track feature vectors and spatial feature vectors, resulting in a vector of length 128, which serves as the input vector for the LSTM module. To Bayesianize this module, a dropout mask, identical for each aircraft and each timestamp, is set up to randomly discard a portion of the neurons representing the input vector. The LSTM module consists of a single-layer LSTM unit, a dropout applied to the LSTM parameters, and a fully connected output layer. All aircraft share a single LSTM module to capture the temporal correlation of each track, ultimately obtaining a probabilistic representation of the prediction result. One cycle of the LSTM is expressed by the formula: 。

Images