Track prediction method and device based on sequence decomposition

Abstract

Along with the increase of airspace traffic complexity, accurate track prediction is crucial to an air traffic control system. Aiming at the problem that long-term trend and short-term fluctuation are difficult to distinguish due to unified modeling of track data in a traditional method, the invention provides a track prediction method and device based on sequence decomposition. According to the method, track data is decomposed into a trend term (reflecting long-term stable features) and a deviation term (representing short-term local fluctuation) through a sequence decomposition module, multi-level spatial-temporal features of the trend term are extracted in combination with a multi-scale feature self-attention module, and dynamic fusion of the two types of features is achieved through a trend-deviation feature interaction module. Specifically, track data collected by an ADS-B system is adopted to construct a data set, a mean square error loss function optimization parameter is introduced in model training, the accuracy and robustness of track prediction are effectively improved through a sequence decomposition and feature interaction mechanism, the flight law of an aircraft is captured more accurately, and reliable decision support is provided for real-time airspace management.

Description

Technical Field

[0001] The present invention relates to the technical field of air traffic management, and particularly to a track prediction method and device based on sequence decomposition. Background Art

[0002] With the increasing complexity of air traffic, accurate track prediction plays a crucial role in the air traffic control (ATC) system. Track prediction not only helps with tasks such as air traffic flow prediction, track planning, and conflict detection, but also effectively assists air traffic controllers in real-time airspace management. Accurate track prediction can improve the decision-making efficiency of air traffic controllers and ensure the safe operation of aircraft. However, traditional track prediction methods face many challenges in practical applications, especially when dealing with complex and variable track data, often having significant limitations. Most existing track prediction methods rely on regression models based on time-series data. Such methods model track data as a unified time series and predict the positions of future track points through regression methods. Although this method can complete the track prediction task to a certain extent, it ignores two different characteristics of the long-term stable trend and short-term local fluctuations of the aircraft during flight. Traditional methods mix these characteristics in a unified time-series model for modeling, making it difficult to fully distinguish the different effects of long-term trends and short-term fluctuations on track prediction, and not considering the impact of short-term fluctuations on long-term trends, thus unable to accurately model track data. Summary of the Invention

[0003] To address the above problems, the present invention provides a track prediction method and device based on sequence decomposition. This method effectively separates the long-term stable trend and short-term local fluctuations of the aircraft by decomposing track data into a trend term and a deviation term, thereby achieving accurate modeling of track data and further improving the accuracy and robustness of track prediction.

[0004] The present invention adopts the following technical solutions:

[0005] A track prediction method based on sequence decomposition, comprising the following steps:

[0006] S1: Collect the original track data collected by the ADS-B system, decode and extract the core information of the original track data, perform data preprocessing on the core information to construct a data set, and divide it into a training set, a validation set, and a test set;

[0007] S2: Construct a track prediction model, including a sequence decomposition module, a multi-scale feature self-attention module, a trend-deviation feature interaction module, and a prediction module;

[0008] The sequence decomposition module decomposes the input track data into a trend term and a deviation term; the multi-scale feature self-attention module includes a multi-scale feature extraction module and a time-series feature modeling module. The multi-scale feature extraction module extracts features from the trend term, and the time-series feature modeling module performs time-series dependence modeling on the output features of the multi-scale feature extraction module; the trend-deviation feature interaction module fuses the deviation term and the output of the multi-scale feature self-attention module.

[0009] S3: Construct a mean square error loss function, use the training set and the validation set to train and validate the track prediction model, fine-tune the hyperparameters according to the validation results, and finally obtain an optimized track prediction model;

[0010] S4: Use the test set to test the optimized track prediction model, and evaluate the model performance according to the evaluation index;

[0011] S5: Decode the real-time track data collected by the ADS-B system, and input it into the optimized track prediction model to generate a track prediction result.

[0012] Preferably, the S1 is specifically:

[0013] S11: Collect the original track data collected by the ADS-B system, decode the original track data, and extract core information such as track number, timestamp, longitude, latitude, altitude, longitude-direction speed, latitude-direction speed, and vertical-direction speed;

[0014] S12: Use the method of linear interpolation to fill in the missing values in the track data; perform normalization processing on the filled track data, and use the preprocessed data as the data set. The normalization processing adopts the Min-Max normalization method:

[0015]

[0016] where X is the unnormalized track data, X * is the normalized track data, X max is the maximum value in the track data, X min is the minimum value in the track data.

[0017] S13: Based on the preprocessed track point data, use a sliding window with a window size of m, with a step size of 1 track point, and sequentially intercept data pairs from the track point data. Specifically, each data pair consists of the observed values of the first m - 1 track points and the target value of the last 1 track point. The first m - 1 track points are used to provide the historical information of the track, and the last 1 track point is used to represent the future position to be predicted by the track prediction model;

[0018] S14: Divide all data pairs into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0019] Preferably, in S2, the sequence decomposition module is:

[0020] The role of the sequence decomposition module is to separate the global trend and local fluctuations in the input track data. The sequence decomposition module consists of a one-dimensional average pooling layer, which decomposes the input track data into a trend term and a deviation term, specifically including:

[0021] S21: Symmetrically pad the input track data at both ends, and then smooth the local fluctuations through a one-dimensional average pooling operation to extract the trend term, representing the global trend of the input track data.

[0022] S22: Subtract the trend term from the input track data to calculate the deviation term, representing the local fluctuations of the input track data. The specific process is as follows:

[0023] X trend = AvgPool(Padding(X))

[0024] X divation = X - X trend

[0025] In the formula, X ∈ R T×D is the input track data, X trend ∈ R T×D is the trend term, X divation ∈ R T×D is the deviation term, where T is the time step, D is the feature dimension, AVgPool(·) represents the one-dimensional average pooling operation, and Padding(·) represents the padding operation.

[0026] Preferably, in S2, the multi-scale feature extraction module is

[0027] The multi-scale feature extraction module aims to extract multi-level feature correlations from track data. The multi-scale feature extraction module is composed of a multi-scale convolutional neural network (MSCNN - Multi-Scale Convolutional Neural Network) and an adaptive weighted feature fusion network (AWFFN - Adaptive Weighted Feature Fusion Network) connected in series.

[0028] Specifically, the multi-scale feature extraction module takes the trend term output by the sequence decomposition module as input, performs convolution operations in the feature dimension using convolutional kernels of different sizes in MSCNN, and outputs multiple sequences containing feature correlation information, aiming to capture the local complex dynamic changes in the trajectory data and extract the correlation between features in the trajectory data. The specific process is expressed as:

[0029] H i =Conv i (Padding(X trend , p i ), i = 1, 2,..., n

[0030] where X trend ∈R T×D represents the input trend term, T is the time step, and D is the feature dimension.

[0031] Padding(·) represents padding the input trend term, and the padding size is p i =(k i -1) / 2, k i is the size of the i-th convolutional kernel, Conv i (·) represents convolving using the i-th convolutional kernel, and H i ∈R T×D is the output sequence extracted by the i-th convolutional kernel, and n is the number of convolutional kernels.

[0032] AWFFN receives multiple sequences output by MSCNN, and adaptively strengthens the correlation between important features through learnable weight linear weighting, and outputs a sequence containing multi-feature information fusion. The specific process is expressed as:

[0033]

[0034] where w i ∈R D is the normalized weight corresponding to H i , and Z ∈ R T×D is the fused output sequence.

[0035] Preferably, in S2, the timing feature modeling module is:

[0036] The timing feature modeling module aims to model the global dependence between time steps of the trajectory data. The timing feature modeling module consists of a feedforward neural network (FNN—Feedforward Neural Network) and a self-attention module (SAM—Self-Attention Module).

[0037] The temporal feature modeling module takes the output of the multi-scale feature extraction module as input, further enhances the features through FFN. SAM receives the sequence output by FNN, establishes global dependencies in the time dimension through the self-attention mechanism, dynamically adjusts the attention between time steps, extracts the correlation between time steps, and captures the long-term dependencies and complex temporal dynamic changes in the trajectory data. The specific process is expressed as:

[0038] Z enhanced = FFN(Z)

[0039] Q = Z enhanced W Q

[0040] K = Z enhanced W K

[0041] V = Z enhanced W V

[0042]

[0043] where Z ∈ R T×D is the input sequence, Z enhanced ∈ R T×D is the output sequence after feature enhancement by FFN, T is the time step, D is the feature dimension, is the learnable weight matrix, D k is the dimension of the query, key, and value, are the query, key, and value matrices respectively, is the output sequence.

[0044] Preferably, in S2, the trend-deviation feature interaction module is:

[0045] The trend-deviation feature interaction module aims to model the dependency between the trend term and the deviation term, and capture the impact of local fluctuations on the global trend in the trajectory data. The trend-deviation feature interaction module is composed of a feature reconstruction module (FRM—Feature Reconstruction Module) and a cross-attention feature fusion module (CAFFM—Cross-Attention Feature Fusion Module) in series.

[0046] Specifically, the trend-deviation feature interaction module takes the deviation term output by the sequence decomposition module and the trend term output by the multi-scale feature self-attention module as input. The role of FRM is to perform feature reconstruction on the input sequence, map the input sequence to a high-dimensional space through the embedding layer, and then, through the reshaping operation, rearrange the dimensions of the features. The specific process is expressed as:

[0047] Strend =transpose(reshape(Embedding(X trend ′)))

[0048] S divation =transpose(reshape(Embedding(X divation )))

[0049] In the formula, is the trend term output by the multi-scale feature self-attention module, and X divation ∈R T×D is the deviation term output by the sequence decomposition module, Embedding(·) is a linear transformation, reshape(·) is a reshaping operation, and transpose(·) is a transpose operation. is the trend term after feature reconstruction, is the deviation term after feature reconstruction, and D1×D2 = D.

[0050] CAFFM takes the trend term and deviation term after feature reconstruction output by FRM as inputs, and uses the cross-attention mechanism to capture the influence of local fluctuations in the track data on the global trend. The specific process is expressed as:

[0051]

[0052]

[0053]

[0054] In the formula, b, c, d, e are the indices of the Einstein summation convention. is the transpose of Sdivation. is the cross-attention weight. is S trend 's transpose, and Y output ∈R T×D is the output sequence.

[0055] Preferably, in S2, the prediction module is:

[0056] The prediction module receives the output features from the trend-deviation feature interaction module and makes predictions for future track points based on these multi-dimensional features. Specifically, the prediction module consists of 6 multi-layer perceptrons (MLPs), which process the multi-dimensional features input to the prediction module channel by channel. After feature mapping processing, the features of each channel will be predicted independently. Specifically, the MLP generates a set of prediction outputs for each channel, and these outputs will reflect the track state of that channel at future times, representing the prediction results of different attributes of the track.

[0057] Preferably, in S4, the calculation formula of the mean square error loss function is as follows:

[0058]

[0059] where N is the number of samples in a batch, K = 6 is the number of attributes, and are the true value and the predicted value of attribute a k respectively.

[0060] Preferably, a trajectory prediction method and device based on sequence decomposition, characterized in that it includes at least one processor and one memory; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a trajectory prediction method according to any one of claims 1 to 8.

[0061] The beneficial effects of the present invention are:

[0062] 1. The present invention proposes a trajectory prediction method based on sequence decomposition, which can effectively separate the long-term trend and short-term deviation in trajectory data, and significantly improve the accuracy and robustness of trajectory prediction. By decomposing the trajectory data into a trend term and a deviation term, it can fully distinguish the different effects of long-term trends and short-term fluctuations on trajectory prediction, and avoid the limitations of traditional methods in mixed feature modeling.

[0063] 2. The present invention proposes a sequence decomposition module, which decomposes the input trajectory data into a trend term and a deviation term through a one-dimensional average pooling layer. This module can effectively separate the global trend and local fluctuations in trajectory data, thus providing clearer feature information for subsequent feature extraction and prediction.

[0064] 3. The present invention proposes a multi-scale feature self-attention module, which combines the multi-scale convolution and the adaptive weighting mechanism with the self-attention mechanism. It extracts the correlation between trajectory data features in the feature dimension by using multi-scale convolution, adaptively weights and fuses features of different scales, and extracts the correlation between time steps in the time dimension by using the self-attention mechanism. This module can capture trajectory features of different granularities, extract the correlation between features in trajectory data, model the global dependence between time steps of trajectory data, and capture the long-term dependence and complex temporal dynamic changes in trajectory data.

[0065] 4. The present invention proposes a trend-deviation feature interaction module, which reconstructs features through trend terms and deviation terms, and uses the cross-attention mechanism for the trend terms and deviation terms after feature reconstruction to model the mutual dependence relationship between the trend terms and the deviation terms. Through the cross-attention mechanism, this module captures the influence of local fluctuations on the global trend in the track data, further improving the prediction accuracy and the expressive ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In the following, the present invention will be described in more detail based on embodiments with reference to the drawings. Among them:

[0067] Figure 1 is a flowchart of the method proposed by the present invention;

[0068] Figure 2 is a model structure diagram of the method proposed by the present invention;

[0069] Figure 3 is a structure diagram of the multi-scale feature self-attention module proposed by the present invention;

[0070] Figure 4 is a structure diagram of the trend-deviation feature interaction module proposed by the present invention;

[0071] Figure 5 is a schematic diagram of the track prediction method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the described embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0073] The present invention will be further described below with reference to the drawings and embodiments.

[0074] As Figure 1 、 Figure 2 shown, a track prediction method based on sequence decomposition includes the following steps:

[0075] S1: Collect the original track data collected by the ADS-B system, decode the original track data and extract the core information, perform data preprocessing on the core information to construct a data set, and divide it into a training set, a validation set, and a test set;

[0076] 1) Collecting the raw track data collected by the ATC system. To obtain the plaintext information, the data must first be decoded. Next, data unrelated to the aircraft's track, such as warning information, should be eliminated to extract the key track information.

[0077] 2) Organize key track information for each flight into a TXT text file, covering the background information and spatiotemporal information for each track point. Background information includes the flight number and track number; spatiotemporal information details the aircraft's various motion trends, such as climbs, descents, turns, and holds. Specifically, this spatiotemporal information includes timestamp, longitude, latitude, altitude, longitudinal speed, latitudinal speed, and vertical speed. The units of longitude and latitude are degrees, the unit of altitude is meters, and the units of speed in all three directions are kilometers per hour.

[0078] 3) Read the spatiotemporal information from the TXT text file and preprocess all data except the timestamp. Based on the time intervals between track points, ensure that the spatiotemporal information is evenly spaced in the time dimension. For track segments with missing values, linear interpolation is used if the number of missing values is small; if the number of missing values is large, the segment is discarded. After preprocessing, a total of 144,605 route data points were obtained, which were used to construct the dataset.

[0079] 4) Based on the preprocessed track point data, a sliding window with a window size of 10 and a step size of 1 track point is used to sequentially intercept data pairs from the track point data. Specifically, each data pair consists of the observation values of the first 9 track points and the target value of the last track point. The first 9 track points are used to provide historical track information, and the last track point is used to represent the future position to be predicted by the track prediction model.

[0080] 5) All data pairs are divided into training set, validation set and test set in the ratio of 8:1:1.

[0081] S2: Construct a track prediction model, including a sequence decomposition module, a multi-scale feature self-attention module, a trend-deviation feature interaction module, and a prediction module;

[0082] like Figure 3 、 Figure 4 、 Figure 5 As shown, the sequence decomposition module decomposes the input track data into trend terms and deviation terms; the multi-scale feature self-attention module includes a multi-scale feature extraction module and a temporal feature modeling module, the multi-scale feature extraction module extracts features from the trend terms, and the temporal feature modeling module performs temporal dependency modeling on the output features of the multi-scale feature extraction module; the trend-deviation feature interaction module fuses the output of the deviation term and the multi-scale feature self-attention module.

[0083] Specifically, the sequence decomposition module symmetrically pads the input track data from beginning to end, then smoothes local fluctuations through a one-dimensional average pooling operation to extract a trend term, which represents the global trend of the input track data. The trend term is subtracted from the input track data to calculate the deviation term, which represents the local fluctuation of the input track data. The specific process is as follows:

[0084] X trend =AvgPool(Padding(X))

[0085] X divation =XX trend

[0086] Where X∈R T×D To input track data, X trend ∈R T×D is the trend term, X divation ∈R T×D is the deviation term, where T is the time step, D is the feature dimension, AvgPool(·) represents the one-dimensional average pooling operation, and Padding(·) represents the padding operation.

[0087] The multi-scale feature self-attention module consists of a multi-scale feature extraction module and a temporal feature modeling module. The multi-scale feature extraction module is designed to extract multi-level feature correlations from track data. It consists of a multi-scale convolutional neural network (MSCNN) and an adaptive weighted feature fusion network (AWFFN) connected in series.

[0088] Specifically, the multi-scale feature extraction module takes the trend item output by the sequence decomposition module as input and performs convolution operations on the feature dimension using one-dimensional convolution kernels of sizes 3, 5, 7, 9, and 11 in the MSCNN. It outputs multiple sequences containing feature correlation information, aiming to capture the complex local dynamic changes in the track data and extract the correlation between features in the track data. The specific process is expressed as follows:

[0089] H i =Conv i (Padding(X trend , p i )),i=1,2,3,4,5

[0090] Where, X trend ∈R T×DDenote the input trend item, T is the time step, and D is the feature dimension. Padding(·) denotes padding the input trend item, and the padding size is p i =(k i -1) / 2, where k i is the size of the i-th convolutional kernel, and Conv i (·) denotes convolving using the i-th convolutional kernel, and H i ∈R T×D is the output sequence extracted by the i-th convolutional kernel.

[0091] AWFFN receives multiple sequences output by MSCNN and adaptively strengthens the correlation between important features through linearly weighting with learnable weights, and outputs a sequence containing multi-feature information fusion. The specific process is expressed as:

[0092]

[0093] where, w i ∈R D is the normalized weight corresponding to H i , and Z ∈ R T×D is the fused output sequence.

[0094] The temporal feature modeling module aims to model the global dependencies between time steps of the track data. The temporal feature modeling module consists of a feedforward neural network (FNN—Feedforward Neural Network) and a self-attention module (SAM—Self-Attention Module).

[0095] Specifically, the temporal feature modeling module takes the output of the multi-scale feature extraction module as input, further enhances the features through FFN, and SAM receives the sequence output by FNN, establishes global dependencies in the time dimension through the self-attention mechanism, dynamically adjusts the attention between time steps, extracts the correlation between time steps, and captures the long-term dependencies and complex temporal dynamic changes in the track data. The specific process is expressed as:

[0096] Z enhanced =FFN(Z)

[0097] Q = Z enhanced W Q

[0098] K = Z enhanced W K

[0099] V = Z enhanced W V

[0100]

[0101] where \(Z\in R\) T×D is the input sequence, \(Z\) enhanced \(\in R\) T×D is the output sequence after FFN feature enhancement, \(T\) is the time step, \(D\) is the feature dimension, is the learnable weight matrix, \(D\) k is the dimension of query, key and value, are the query, key and value matrices respectively, is the output sequence.

[0102] The trend-deviation feature interaction module aims to model the dependence between the trend term and the deviation term, and capture the impact of local fluctuations on the global trend in the track data. The trend-deviation feature interaction module is composed of a feature reconstruction module (FRM - Feature Reconstruction Module) and a cross-attention feature fusion module (CAFFM - Cross-Attention Feature Fusion Module) in series.

[0103] Specifically, the trend-deviation feature interaction module takes the deviation term output by the sequence decomposition module and the trend term output by the multi-scale feature self-attention module as inputs. The role of FRM is to reconstruct the features of the input sequence. The input sequence is mapped to a high-dimensional space through the embedding layer. Next, the dimensions of the features are rearranged through the reshape operation. The specific process is expressed as:

[0104] \(S\) trend =transpose(reshape(Embedding(X trend ')))

[0105] \(S\) divation =transpose(reshape(Embedding(X divaion )))

[0106] where is the trend term output by the multi-scale feature self-attention module, \(X\) divation \(\in R\) T×D is the deviation term output by the sequence decomposition module, Embedding(·) is the linear transformation, reshape(·) is the reshape operation, transpose(·) is the transpose operation, is the trend term after feature reconstruction, is the deviation term after feature reconstruction, \(D1\times D2 = D\).

[0107] CAFFM takes the trend term and deviation term after reconstructing the features output by FRM as inputs, and uses the cross-attention mechanism to capture the impact of local fluctuations on the global trend in the track data. The specific process is expressed as follows:

[0108]

[0109]

[0110]

[0111] In the formula, b, c, d, e are the indices of the Einstein summation convention. is the transpose of S divation , is the cross-attention weight, is the transpose of S trend , Y output ∈R T×D is the output sequence.

[0112] The prediction module receives the output features from the trend-deviation feature interaction module and makes predictions of future track points based on these multi-dimensional features. Specifically, the prediction module consists of 6 multi-layer perceptrons (MLPs), which process the multi-dimensional features input to the prediction module channel by channel. After feature mapping processing, the features of each channel are predicted independently. Specifically, the MLP generates a set of prediction outputs for each channel, and these outputs will reflect the track state of that channel at future times, representing the prediction results of different attributes of the track.

[0113] S3: Construct a mean square error loss function, use the training set and validation set to train and validate the track prediction model, and fine-tune the hyperparameters according to the validation results to finally obtain an optimized track prediction model;

[0114]

[0115] Among them, N = 2048 is the number of samples in a batch, K = 6 is the number of attributes, and are the true value and predicted value of attribute a k respectively.

[0116] S4: Use the test set to test the optimized track prediction model and evaluate the model performance according to the evaluation metrics;

[0117] S5: Real-time track data is collected from the ADS-B system. After the original data is decoded, historical flight track segments in a specific period are input into the trained track prediction model to generate position prediction results for future periods, providing decision-making support for the aviation operation management system. This predicted data can serve multiple types of air traffic control scenarios, including data support for intelligent air traffic control functional modules such as airspace traffic situation deduction, collision risk probability assessment, and aircraft separation maintenance monitoring.

Claims

1. A track prediction method based on sequence decomposition, characterized in that It includes the following steps: S1: Collect the original track data collected by the ADS-B system, decode the original track data and extract the core information, perform data preprocessing on the core information to construct a data set, and divide it into a training set, a validation set, and a test set; S2: Construct a track prediction model, including a sequence decomposition module, a multi-scale feature self-attention module, a trend-deviation feature interaction module, and a prediction module; The sequence decomposition module decomposes the input track data into a trend term and a deviation term; the multi-scale feature self-attention module includes a multi-scale feature extraction module and a time series feature modeling module. The multi-scale feature extraction module extracts features from the trend term, and the time series feature modeling module performs time series dependence modeling on the output features of the multi-scale feature extraction module; The trend-deviation feature interaction module fuses the deviation term and the output of the multi-scale feature self-attention module; S3: Construct a mean square error loss function, use the training set and the validation set to train and validate the track prediction model, fine-tune the hyperparameters according to the validation results, and finally obtain an optimized track prediction model; S4: Use the test set to test the optimized track prediction model, and evaluate the model performance according to the evaluation index; S5: Decode the real-time track data collected by the ADS-B system, and input it into the optimized track prediction model to generate a track prediction result.

2. The method for predicting a track based on sequence decomposition according to claim 1, wherein Specifically, S1 is as follows: S11: Collect the original track data collected by the ADS-B system, decode the original track data, and extract core information such as track number, timestamp, longitude, latitude, altitude, longitude-direction speed, latitude-direction speed, and vertical-direction speed; S12: Use the method of linear interpolation to fill in the missing values in the track data; perform standardization processing on the filled track data, and use the preprocessed data as the data set; S13: Based on the preprocessed track point data, use a sliding window with a window size of m, with a step size of 1 track point, and sequentially intercept data pairs from the track point data. Specifically, each data pair is composed of the observed values of the first m - 1 track points and the target value of the last 1 track point. The first m - 1 track points are used to provide the historical information of the track, and the last 1 track point is used to represent the future position to be predicted by the track prediction model; S14: Divide all data pairs into a training set, a validation set, and a test set according to the ratio of 8:1:

1.

3. The track prediction method based on sequence decomposition according to claim 1, wherein In step S2, the sequence decomposition module is specifically: The function of the sequence decomposition module is to separate the global trend and local fluctuations in the input track data. The sequence decomposition process decomposes the input track data into a trend term and a deviation term, specifically including: S21: Symmetrically pad the head and tail of the input track data, and then smooth the local fluctuations through a one-dimensional average pooling operation to extract the trend term, representing the global trend of the input track data. S22: Subtract the trend term from the input track data to calculate the deviation term, representing the local fluctuations of the input track data.

4. The method for predicting a track based on sequence decomposition according to claim 1, wherein In step S2, the multi-scale feature extraction module is composed of a multi-scale convolutional neural network and an adaptive weighted feature fusion network connected in series. The multi-scale feature extraction module takes the trend term output by the sequence decomposition module as input, and uses convolutional kernels of different sizes in the multi-scale convolutional neural network to perform convolutional operations in the feature dimension, outputting multiple sequences containing feature correlation information.

5. The track prediction method based on sequence decomposition according to claim 1, characterized in that In step S2, the temporal feature modeling module consists of a feed-forward neural network and a self-attention module. The temporal feature modeling module takes the output of the multi-scale feature extraction module as input, further enhances the features through the feed-forward neural network. The self-attention module receives the sequence output by the feed-forward neural network, establishes global dependencies in the time dimension through the self-attention mechanism, dynamically adjusts the attention between time steps, extracts the correlation between time steps, and captures the long-term dependencies and complex temporal dynamic changes in the trajectory data.

6. The method for predicting a track based on sequence decomposition according to claim 1, wherein In step S2, the trend-deviation feature interaction module is composed of a feature reconstruction module and a cross-attention feature fusion module connected in series. The trend-deviation feature interaction module takes the deviation term output by the sequence decomposition module and the trend term output by the multi-scale feature self-attention module as input. The feature reconstruction module performs feature reconstruction on the input sequence, maps the input sequence to a high-dimensional space through an embedding layer, and rearranges the dimensions of the features through a reshaping operation.

7. The method for predicting a track based on sequence decomposition according to claim 1, wherein In step S2, the prediction module receives the output features from the trend-deviation feature interaction module and predicts future trajectory points based on this.

8. The method for track prediction based on sequence decomposition according to claim 7, wherein The prediction module includes six multi-layer perceptrons, processes the multi-dimensional features input to the prediction module channel by channel. After feature mapping processing, the features of each channel are predicted independently.

9. A track prediction device based on sequence decomposition, characterized in that, It includes at least one processor and a memory. The memory stores instructions executable by the at least one processor. The instructions are executed by the at least one processor so that the at least one processor can execute the trajectory prediction method based on trend decomposition according to any one of claims 1 to 8.

Images