Aircraft trajectory prediction method and system based on representation learning
By adopting a two-stage learning method based on characterization learning in aircraft trajectory prediction, the problems of low prediction efficiency and low accuracy in the prior art are solved, and efficient and high-precision trajectory prediction is achieved.
Patent Information
- Application Number
- CN202510059974.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-15
- Publication Date
- 2025-05-23
AI Technical Summary
The prior art has problems of low prediction efficiency and low prediction accuracy in aircraft trajectory prediction.
The two-stage learning method based on representation learning is adopted. First, the characterization learning of trajectory data is performed through self-supervised pre-training, and the trajectory data is reconstructed; then in the fine-tuning stage of the prediction task, weight initialization is used to perform high-precision trajectory prediction.
The model's ability to characterize trajectory data is improved, and efficient and high-precision aircraft trajectory prediction is achieved.
Smart Images
Figure CN120030881A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of space target trajectory prediction, and in particular to an aircraft trajectory prediction method and system based on representation learning. Background Art
[0002] Space trajectory prediction refers to analyzing the three-dimensional trajectory of a target in the past period of time, and then predicting its flight trajectory in the future. In the air battlefield, the trajectory prediction of the target can effectively perform key tasks such as threat assessment, target tracking, and maneuver decision-making. How to achieve accurate prediction of target movement has become a major issue in current space trajectory prediction.
[0003] In the early days, target trajectory prediction methods relied on aerodynamic models and state estimation algorithms. Aerodynamic models modeled the motion of aircraft to achieve trajectory prediction by affecting aircraft performance, state, environment, intention, and other aspects; state estimation algorithms estimated the state of the aircraft at the next moment by constructing state transitions. However, these methods have a large number of modeling parameters and limited prediction length, making them difficult to apply to real environments.
[0004] In recent years, with the rise of artificial intelligence technology, deep neural networks have begun to be applied to target trajectory prediction. In the early stages of this direction, various sequence models such as LSTM, RNN, GRU, etc. were used for trajectory prediction. Shi et al. used LSTM neural network to build a 4D trajectory prediction model, which achieved obvious advantages over the traditional Markov model. Sekhon et al. used LSTM network with attention mechanism for autonomous ship prediction. The model combines spatial and temporal attention mechanisms and spatial and temporal weights so that different areas of information can be focused on. With the higher requirements for trajectory prediction length and accuracy, research began to use Transformer model for long-sequence trajectory prediction. Nguyen et al. proposed TrAISformer model, which obtains data representation that is more conducive to model learning by gridding trajectory data embedding and uses Transformer network for iterative ship trajectory prediction. Guo et al. proposed FlightBERT model for civil aviation trajectory prediction task. The model proposed a new attention block called attribute correlation attention to clearly capture the influence of speed attribute on the corresponding position dimension, and improve the ability of Transformer model to mine trajectory attribute information.
[0005] However, the above algorithms are weak in their ability to represent the intrinsic characteristics of trajectories and mine structural information, and there are still inaccurate predictions when performing prediction tasks. Summary of the invention
[0006] In order to overcome the shortcomings of the prior art, the present invention provides an aircraft trajectory prediction method and system based on representation learning to solve the problems of low prediction efficiency and low prediction accuracy in the prior art.
[0007] The technical solution adopted by the present invention to solve the above problems is:
[0008] A method for predicting aircraft trajectory based on representation learning predicts aircraft trajectory data based on a self-supervised learning method for aircraft trajectory data.
[0009] As a preferred technical solution, the following steps are included:
[0010] A. Self-supervised pre-training: Use self-supervised learning methods to learn the representation of trajectory data and reconstruct trajectory data;
[0011] B, Prediction task fine-tuning: perform trajectory prediction on the reconstructed trajectory data.
[0012] As a preferred technical solution, step A comprises the following steps:
[0013] A1, trajectory data mask processing: Get input trajectory data Z based on trajectory dataset Z mask , randomly add masks of different lengths to the trajectory to realize trajectory data masking. The formula is:
[0014]
[0015] Where i represents the number of the trajectory point in the trajectory set Z, represents the trajectory point data after mask processing of the ith trajectory set Z, z i represents the i-th original trajectory point data in the trajectory set Z, and Mask represents the randomly generated mask position sequence;
[0016] A2, trajectory interpolation reconstruction: based on Z and Z mask Perform trajectory interpolation reconstruction.
[0017] As a preferred technical solution, step A2 includes the following steps:
[0018] A21, Z mask With PE mask Embedded into feature space:
[0019]
[0020] Where T represents the trajectory sequence length, d embed represents the embedding dimension, represents the embedding vector representation, Embedding(·) represents the embedding operation, PE represents the position encoding, and PE maskrepresents a positional encoding with a mask;
[0021] A22, will Perform encoding processing and output the encoding matrix
[0022] Where N represents the number of layers of the embedding network, Represents the encoding matrix of the Nth layer embedding network output, EncodingLayer N (·) represents the Nth layer embedding operation;
[0023] A23, Restore to reconstruct trajectory data Reconstruct Z by interpolation mask The missing mask data in is filled and the trajectory is reconstructed.
[0024] As a preferred technical solution, in step A2, an encoder network is used to perform trajectory interpolation reconstruction, and the encoder network includes an embedding layer, N encoding layers, and a linear projection layer that are sequentially connected in communication; wherein N≥1 and N is an integer;
[0025] In step A21, the trajectory position code and the data code are embedded into the feature space through an embedding layer;
[0026] In step A22, the N-layer coding layer is used to to process;
[0027] In step A23, the linear projection layer is used to transform Restore to reconstruct trajectory data
[0028] As a preferred technical solution, in step A2, the trajectory data learned by training representation is calculated by the formula of reconstruction loss of training:
[0029]
[0030] Where n represents the number of trajectory points, α represents the weight coefficient, i and j represent the numbers of trajectory points, express The i-th reconstruction trajectory in express The jth reconstruction trajectory in .
[0031] As a preferred technical solution, in step B, the input trajectory X and the predicted trajectory Y are divided from the trajectory data set Z, X and Y are used as training data sets, the reconstructed trajectory data is trained, and the trained predicted trajectory is generated based on X
[0032] As a preferred technical solution, in step B, the calculation formula of the loss function for training the reconstructed trajectory data is:
[0033]
[0034] Among them, L MSE represents the MSE loss, i represents the i-th sample number, n represents the total number of samples, and y i represents the i-th element in Y, express The i-th element in .
[0035] As a preferred technical solution, in step A, before adopting the self-supervised learning method to perform characterization learning of the trajectory data, sliding window processing is performed on the trajectory data.
[0036] A system for predicting aircraft trajectory based on representation learning, used to implement the method for predicting aircraft trajectory based on representation learning, comprises the following modules which are sequentially connected in communication:
[0037] Self-supervised pre-training module: used to learn the representation of trajectory data using self-supervised learning methods and reconstruct trajectory data;
[0038] Prediction task fine-tuning module: used to perform trajectory prediction on the reconstructed trajectory data.
[0039] Compared with the prior art, the present invention has the following beneficial effects:
[0040] The present invention uses two-stage learning to perform trajectory prediction tasks. In the first stage, self-supervised pre-training is used to improve the model's ability to represent trajectory data. In the second stage, weight initialization is used to fine-tune the prediction task to achieve high-precision trajectory prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 A prediction flow chart of an aircraft trajectory prediction method based on representation learning according to the present invention;
[0042] Figure 2 This is a comparison diagram of trajectory prediction results using the Transformer model between the present invention and the prior art;
[0043] Figure 3 This is the second comparison diagram of trajectory prediction results using the Transformer model between the present invention and the prior art. DETAILED DESCRIPTION
[0044] The present invention will be further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.
[0045] It is worth noting that: the following embodiments take an aircraft as an example; however, in view of the technical problem to be solved, the technical solution of the present invention can also be used in other aircraft to achieve the technical effect of the present invention, and these technical solutions for other aircraft are also within the inventive concept of the present invention.
[0046] Example 1
[0047] like Figures 1 to 3 As shown, in recent years, representation learning technology has been applied in many fields. The technology aims to automatically discover useful features or representations of data instead of relying on manually designed features. Through the learned representation, the model can better understand and process the data, thereby improving the performance of the task. In the field of deep learning, representation learning uses self-supervised learning or semi-supervised learning technology to directly learn the internal information of the data and obtain effective representation of the data without labels. The present invention applies the representation learning technology based on self-supervised learning to trajectory prediction, providing a new research direction for aircraft trajectory information mining.
[0048] This paper studies the task of aircraft trajectory prediction and proposes a self-supervised learning method to characterize trajectory data, thereby improving the model's ability to understand trajectory data and achieving more efficient and accurate aircraft trajectory prediction. Specifically, this work implements a two-stage trajectory data characterization, which includes model self-supervised pre-training and model prediction task fine-tuning.
[0049] The main innovation of this invention is to apply pre-trained representation learning to trajectory prediction. The model uses two-stage learning to perform trajectory prediction tasks. In the first stage, self-supervised pre-training is used to improve the model's ability to represent trajectory data. In the second stage, weight initialization is used to fine-tune the prediction task to achieve high-precision trajectory prediction.
[0050] The aircraft trajectory data consists of the longitude lat, latitude lon, altitude alt and speed vel where the aircraft is located. Its data format Traj can be regarded as a multivariate time series with a dimension of [4,N]:
[0051] Traj = [lat i ,lon i ,alt i ,vel i ],ii=1,…,N
[0052] Among them, i represents the i-th trajectory point, N represents the number of trajectory points in the entire trajectory, and [4,N] represents N points, each with 4 parameters.
[0053] The trajectory prediction task requires predicting future unknown trajectories based on some known trajectories. Generally speaking, the flight trajectories of aircraft vary in length and have a large number of sampling points. Before prediction, the trajectories need to be processed by sliding windows. Sliding window processing can provide a training data set with consistent length and sufficient data volume for model training. For each sliding window trajectory, it can be regarded as a combination of an input sequence and a label sequence. The input sequence is a historical observation sequence. Where C is the variable dimension, L is the size of the look-back window, and the trajectory prediction task uses the input sequence to predict the future sequence. Where T is the prediction length.
[0054] In the present invention, the trajectory data is firstly pre-trained by model self-supervision. At this stage, the present invention adopts self-supervision learning method to learn the representation of trajectory data. For trajectory data Z, the present invention obtains input trajectory data Z through random masking module (which can be realized by existing technology, and its specific working principle and working process are not repeated here) mask ,The random masking module randomly adds masks of different lengths to the trajectory to achieve trajectory data hollowing.
[0055]
[0056] Then through Z and Z mask Constructing the trajectory interpolation reconstruction task, the present invention adopts the encoder network architecture, Z mask First, the trajectory position encoding and data encoding are embedded into the feature space through the embedding layer:
[0057]
[0058] Afterwards Processed through N consecutive coding layers, the output coding matrix
[0059] In the reconstruction task, the model transforms Restore to reconstruct trajectory data The model reconstructs Z by interpolation mask The missing mask data in is filled and the trajectory is reconstructed, and the reconstruction performance of the model is optimized through the reconstruction loss:
[0060]
[0061] in: is the reconstruction error ( refers to N samples), is the interpolation error, and the parameter α weighs the reconstruction error and interpolation error. The idea behind the self-supervised learning task is to train the model to grasp the intrinsic representation and structure of the trajectory, thereby improving the performance in prediction.
[0062] In the trajectory prediction fine-tuning stage, the trajectory data Z Part of the trajectory in is divided into input trajectory X X and Y are used as the training data set for this stage. X is used as the model input and Y is used as the true value label to help optimize the model. X Generate prediction trajectories through the encoder network And optimize the model using MSE loss:
[0063]
[0064] Since the model has learned some representation information of the trajectory data in the self-supervised pre-training stage, its weights can be used as the initialization weight parameters in the prediction fine-tuning stage (using the pre-trained model parameters as the initialization parameters of the prediction model training parameters), thereby accelerating the model convergence and improving the prediction accuracy in this stage.
[0065] Example 2
[0066] like Figures 1 to 3 As shown, based on Example 1, this example provides a more detailed implementation method.
[0067] The experimental data is composed of 10 sets of civil aviation ADS-B data. The training set and the test set are divided into 7:3 based on a single trajectory, and the trajectory data is intercepted using a sliding window of 160 data points. When making predictions, 60 points are used as input trajectories and 100 points are predicted.
[0068] Table 1 compares the prediction performance of the method of the present invention (trajectory representation model) and the Transformer model. It can be seen that the present invention can improve the trajectory prediction accuracy.
[0069] Table 1 Prediction performance comparison table
[0070]
[0071] Figure 2 , Figure 3 The specific prediction results of the two methods are shown: TRL is the method of the present invention, Transformer is the method of the Transformer model, input is the input trajectory, and ground_truth is the real future trajectory.
[0072] As described above, the present invention can be preferably implemented.
[0073] All features disclosed in all embodiments in this specification, or steps in all methods or processes implicitly disclosed, except for mutually exclusive features and / or steps, can be combined and / or expanded or replaced in any manner.
[0074] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. According to the technical essence of the present invention, within the spirit and principles of the present invention, any simple modification, equivalent replacement and improvement made to the above embodiment still falls within the protection scope of the technical solution of the present invention.
Claims
1. A method for predicting aircraft trajectories based on representation learning, characterized in that: The aircraft trajectory data is predicted based on the self-supervised learning method of the aircraft trajectory data.
2. The method for predicting aircraft trajectories based on representation learning according to claim 1, characterized in that: The following steps are involved: A. Self-supervised pre-training: Use self-supervised learning methods to learn the representation of trajectory data and reconstruct trajectory data; B, Prediction task fine-tuning: perform trajectory prediction on the reconstructed trajectory data.
3. The method for predicting aircraft trajectory based on representation learning according to claim 2, characterized in that: Step A includes the following steps: A1, trajectory data mask processing: Get input trajectory data Z based on trajectory dataset Z mask , randomly add masks of different lengths to the trajectory to realize trajectory data masking. The formula is: Where i represents the number of the trajectory point in the trajectory set Z, represents the trajectory point data after mask processing of the ith trajectory set Z, z i represents the i-th original trajectory point data in the trajectory set Z, and Mask represents the randomly generated mask position sequence; A2, trajectory interpolation reconstruction: based on Z and Z mask Perform trajectory interpolation reconstruction.
4. The method for predicting aircraft trajectory based on representation learning according to claim 3, characterized in that: Step A2 includes the following steps: A21, Z mask With PE mask Embedded into feature space: Where T represents the trajectory sequence length, d embed represents the embedding dimension, represents the embedding vector representation, Embedding(·) represents the embedding operation, PE represents the position encoding, PE mask represents a positional encoding with a mask; A22, will Perform encoding processing and output the encoding matrix Where N represents the number of layers of the embedding network, Represents the encoding matrix of the Nth layer embedding network output, EncodingLayer N (·) represents the Nth layer embedding operation; A23, Restore to reconstruct trajectory data Reconstruct Z by interpolation mask The missing mask data in is filled and the trajectory is reconstructed.
5. The method for predicting aircraft trajectory based on representation learning according to claim 4, characterized in that: In step A2, an encoder network is used to perform trajectory interpolation reconstruction, and the encoder network includes an embedding layer, N encoding layers, and a linear projection layer that are sequentially connected in communication; wherein N ≥ 1 and N is an integer; In step A21, the trajectory position code and the data code are embedded into the feature space through an embedding layer; In step A22, the N-layer coding layer is used to to process; In step A23, the linear projection layer is used to transform Restore to reconstruct trajectory data 6. The method for predicting aircraft trajectory based on representation learning according to claim 3, characterized in that: In step A2, the trajectory data learned by training representation is calculated by the formula of training reconstruction loss: Where n represents the number of trajectory points, α represents the weight coefficient, i and j represent the numbers of trajectory points, express The i-th reconstruction trajectory in express The jth reconstruction trajectory in .
7. The method for predicting aircraft trajectory based on representation learning according to claim 2, characterized in that: In step B, the input trajectory X and the predicted trajectory Y are divided from the trajectory data set Z. X and Y are used as training data sets. The reconstructed trajectory data is trained and the trained predicted trajectory is generated based on X.
8. The method for predicting aircraft trajectory based on representation learning according to claim 7, characterized in that: In step B, the loss function for training the reconstructed trajectory data is calculated as: Among them, L MSE represents the MSE loss, i represents the i-th sample number, n represents the total number of samples, and y i represents the i-th element in Y, express The i-th element in .
9. The method for predicting aircraft trajectory based on representation learning according to any one of claims 2 to 8, characterized in that: In step A, before using the self-supervised learning method to learn the representation of the trajectory data, the trajectory data is subjected to sliding window processing.
10. An aircraft trajectory prediction system based on representation learning, characterized in that: A method for predicting aircraft trajectories based on representation learning, used to implement any one of claims 1 to 9, comprising the following modules that are sequentially communicated with each other: Self-supervised pre-training module: used to learn the representation of trajectory data using self-supervised learning methods and reconstruct trajectory data; Prediction task fine-tuning module: used to perform trajectory prediction on the reconstructed trajectory data.