Multi-target interactive flight trajectory prediction method based on SCINet network

Abstract

A multi-target interactive flight trajectory prediction method based on SCINet network, belonging to the field of traffic trajectory prediction, includes the following steps: Step 1, extracting trajectory data to form a multi-target interactive scenario; Step 2, designing a model framework for multi-target interactive flight trajectory prediction based on SCINet network; the model captures the long-term temporal dependence, multi-level features, and multi-target spatial characteristics of the trajectory; the method used to capture the long-term temporal dependence and multi-level features of the trajectory is based on the SCIcross temporal feature extraction method; Step 3, inputting the observed trajectory information of the test into the trained model to obtain the predicted trajectory. This invention improves the accuracy of flight trajectory prediction in multi-target interactive scenarios within the same scene.

Description

Technical Field

[0001] This invention belongs to the field of traffic trajectory prediction and relates to a multi-target interactive flight trajectory prediction method based on SCINet network. Background Technology

[0002] Flight trajectory prediction is a crucial and highly challenging task in air traffic control, especially maneuver control. With the continued development of the global economy, the demand for air transport from various industries has increased significantly, resulting in rapid growth in air traffic volume and a corresponding increase in airspace complexity. Against this backdrop, accurate flight trajectory prediction has become a vital component of ensuring airspace safety. Precise trajectory prediction not only provides better decision support for air traffic control but also effectively reduces the risk of mid-air conflicts and collisions, avoiding potential traffic accidents and thus protecting the lives of pilots and passengers.

[0003] Trajectory prediction has long been a subject of research and interest for scholars. Traditional methods generally cannot effectively handle complex nonlinear dynamic changes, especially when dealing with multidimensional features and long-term series data, where their performance is limited. Furthermore, these methods often lack spatiotemporal interaction modeling for multiple target trajectories or suffer from insufficient performance. To address these issues, this invention proposes a multi-target interactive flight trajectory prediction method based on the SCINet (Sample Convolution and Interaction Network) network. By utilizing the unique sample convolution and cross-scale attention mechanisms of the SCI network and the efficient spatial relationship modeling of the GATv2 (GraphAttention Network v2) network, it can effectively capture long-term dependencies and spatial interaction features in the input sequence, thereby improving prediction accuracy.

[0004] To further enhance the model's performance, this method introduces a Conditional Variational Autoencoder (CVAE) to learn the latent distribution patterns of trajectories, thereby improving the ability to predict future trajectories. CVAE plays a crucial role in modeling uncertainty and capturing the diversity of trajectory data, contributing to improved model generalization ability and predictive stability, thus enhancing the accuracy and reliability of trajectory prediction in intelligent transportation systems for air transport. Summary of the Invention

[0005] To overcome the shortcomings of existing trajectory prediction techniques, such as difficulty in handling multidimensional features and long-term sequences, and lack of spatial interaction, this invention proposes a multi-target interactive flight trajectory prediction method based on the SCINet network. This method combines the SCINet network, capable of extracting multidimensional features, with the GATv2 network, capable of extracting complex spatial relationships and interactions, to achieve spatiotemporal modeling. Furthermore, a CVAE network is introduced to learn latent patterns in the airspace environment to reconstruct the trajectory. This effectively solves the problem of poor trajectory prediction accuracy, enables multi-target interaction within the same scene, and improves prediction performance.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] A method for predicting the interactive flight trajectory of multiple targets based on SCINet network includes the following steps:

[0008] Step 1: Obtain flight trajectories from the dataset and establish a multi-target interaction scenario;

[0009] Step 2: For multi-target interaction scenarios, design a model framework for multi-target interaction flight trajectory prediction based on SCINet network, and design the model training method and training objectives; the process is as follows: the model captures the long-term temporal dependence, multi-level features and multi-target spatial characteristics of the trajectory.

[0010] The method used to capture the long-term temporal dependence and multi-level features of trajectories is based on the SCIcross temporal feature extraction method, and the process is as follows:

[0011] SCIcross is composed of multiple SCI-Blocks stacked in a tree structure, performing dimensional interaction calculations on each decomposed sequence. This allows for increasing the stacking depth to enhance model capabilities. First, the input trajectory sequence is decomposed into odd and even sub-sequences, and convolutional filters are used to extract different temporal features from each sub-sequence. Then, the odd sequence is multiplied by the elements of the convolutional even sequence, and the even sequence is multiplied by the elements of the convolutional odd sequence, resulting in two sets of features with complete sequence views. Stacking SCI-Blocks further decomposes these two sets of features with complete sequence views, yielding more detailed information. Each decomposed sequence also requires attention calculations on dimensional features, considering inter-dimensional interaction attention to enhance prediction accuracy. Finally, the resulting features are recombine and rearranged to obtain the final features.

[0012] ;

[0013] ;

[0014] ;

[0015] ;

[0016] ;

[0017] ;

[0018] in, , These are the odd and even sequences in the original sequence. , , , It is a one-dimensional convolutional filter; This indicates multiplying the elements at corresponding positions of the two matrices; , These are the characteristics resulting from the interaction of odd and even sequences; , This represents the result of feature combination after interaction; Att is the attention calculation operation. , This is the result after dimensional interaction;

[0019] Step 3: Input the observed trajectory information used for testing into the trained model to obtain the predicted trajectory.

[0020] Furthermore, in step 1, a model for predicting multi-target interactive flight trajectories based on the SCINet network is designed. The process of establishing the multi-target interactive scenario is as follows:

[0021] Extract a specific time of day from the dataset. The flight trajectory, using a set The sequence number represents a different trajectory; based on the sequence number, all positions below it are retrieved, and each position contains... Information from three dimensions yields the corresponding trajectory information. ,in The divided observation trajectory is The predicted trajectory is , The specified time length is the time length of the observation trajectory; the specified time length is... The set of trajectories that exist below { This can be considered as a multi-objective interaction scenario.

[0022] A further preferred method for capturing multi-target interaction trajectory features is the spatial feature extraction method based on GATv2, the process of which is as follows:

[0023] The GATv2 network can calculate the correlation between neighboring points in an image and determine their relative importance to the central node, thereby allocating more resources to it. It treats trajectories within the same scene as nodes in the image and calculates the correlation coefficient between trajectories. Normalization yields the attention coefficients between nodes. Its multi-head attention mechanism further learns multiple nonlinear mappings to simulate dynamic spatial interactions influenced by various factors in different potential subspaces.

[0024] ;

[0025] ;

[0026] in, and The sequence number is considered here as the center point and adjacent points of the graph; and This represents the characteristics of the trajectory after time encoding corresponding to the sequence number; and It is a linear transformation matrix; For nodes and The correlation coefficient; Attention coefficient; This is a set of trajectory sequence numbers.

[0027] In step 2, a multi-target interactive flight trajectory prediction model based on the SCINet network is designed. The model adopts a trajectory spatiotemporal feature distribution law capture method based on CVAE, and the process is as follows:

[0028] CVAE networks introduce latent variables To model the relationship between observed trajectory features and predicted trajectory features, This represents features learned from historical trajectories that are related to the distribution of future trajectories and have randomness. First, the spatiotemporal features of the input trajectory are mapped to a Gaussian distribution with a mean by the encoder. and variance Then, samples were taken from a Gaussian distribution with the same statistical properties. Finally, the decoder parses... Spatiotemporal characteristics of the observed trajectory Generate future trajectory ;

[0029] ;

[0030] ;

[0031] ;

[0032] in, Predicted trajectory features after time modeling; This is the CVAE encoding part; The noise is a standard normally distributed Gaussian noise. This is the CVAE decoding part.

[0033] Preferably, in step 2, a multi-target interactive flight trajectory prediction model based on the SCINet network is designed, and the model training method and training objective process are as follows:

[0034] Step 2.1: Design the model loss function;

[0035] The model loss function consists of two parts: 2.1.1, the reconstruction loss represents the loss on the latent variables. The reconstruction loss is calculated by adjusting the mean squared error. A larger value indicates that the predicted trajectory generated by the model is closer to the true trajectory. 2.1.2 KL divergence represents the latent space learned by the model. The difference between the distribution pattern of the actual trajectory and the distribution pattern of the predicted trajectory is ultimately simplified to: Therefore, the model training objective is transformed into minimizing the loss function. ;

[0036] ;

[0037] ;

[0038] ;

[0039] in, This represents the total number of serial numbers. This is the total loss function; For serial number A reconstructed trajectory corresponding to the future trajectory;

[0040] Step 2.2: Model training;

[0041] During the training phase, the observed trajectory set is first processed using the SCITree encoding model. and predicted trajectory set Temporal modeling is performed to obtain a multi-level, long-term dependent set of temporal features. and ;Then, Spatial modeling using GATv2 yields spatiotemporal features of multi-objective interactions. Finally, and Feeding into the CVAE layer to generate latent variables And reconstruct future trajectory location Complete the training and save the model weights.

[0042] In step 3, a multi-target interactive flight trajectory prediction model based on the SCINet network is designed, and the testing process is as follows:

[0043] During the testing phase, the model only applies to the observed trajectory. Spatiotemporal modeling is performed, and its spatiotemporal characteristics are... With Gaussian noise Send them together into CVAE to decode their future trajectory Read the training weights, and feed the test data into the model for testing. The test is based on mean squared error and dynamic time warping.

[0044] Compared with existing technologies, the beneficial effects of this invention are mainly reflected in the following aspects: the proposed method can decompose time series according to different time scales and capture multi-level spatiotemporal information of trajectory data; the multi-layer recursive structure and cross-scale attention mechanism can also better model complex long-term and short-term dependencies, improving the accuracy of prediction; the proposed method can focus on the spatial relationship between trajectories and handle complex spatial interactions; the proposed method introduces randomness, which can maintain high robustness and prediction performance in prediction based on prior knowledge. Attached Figure Description

[0045] Figure 1 This is a flowchart of the multi-target interactive flight trajectory prediction based on the SCINet network according to an embodiment of the present invention;

[0046] Figure 2 This is a comparison diagram of multi-target interactive flight trajectory prediction algorithms based on SCINet network according to embodiments of the present invention. Detailed Implementation

[0047] The invention will now be further described with reference to the accompanying drawings.

[0048] Reference Figure 1 A multi-target interactive flight trajectory prediction method based on SCINet network includes the following steps:

[0049] Step 1: Obtain flight trajectories from the dataset and establish a multi-target interaction scenario;

[0050] In step 1, a model for predicting multi-target interactive flight trajectories based on the SCINet network is designed. The process of establishing the multi-target interactive scenario is as follows:

[0051] Extract a specific time of day from the dataset. The flight trajectory, using a set The sequence number represents a different trajectory; based on the sequence number, all positions below it are retrieved, and each position contains... Information from three dimensions yields the corresponding trajectory information. ,in The divided observation trajectory is The predicted trajectory is , The specified time length is the time length of the observation trajectory; the specified time length is... The set of trajectories that exist below { This can be considered as a multi-objective interaction scenario.

[0052] Step 2: For multi-target interaction scenarios, design a model framework for multi-target interaction flight trajectory prediction based on SCINet network, and design the model training method and training objectives.

[0053] In step 2, a multi-target interactive flight trajectory prediction model based on the SCINet network is designed, which captures the long-term temporal dependence, multi-level features and multi-target spatial characteristics of the trajectory.

[0054] The method used to capture the long-term temporal dependence and multi-level features of trajectories is based on the SCIcross temporal feature extraction method, and the process is as follows:

[0055] SCIcross is composed of multiple SCI-Blocks stacked in a tree structure, performing dimensional interaction calculations on each decomposed sequence. This allows for increasing the stacking depth to enhance model capabilities. First, the input trajectory sequence is decomposed into odd and even sub-sequences, and convolutional filters are used to extract different temporal features from each sub-sequence. Then, the odd sequence is multiplied by the elements of the convolutional even sequence, and the even sequence is multiplied by the elements of the convolutional odd sequence, resulting in two sets of features with complete sequence views. Stacking SCI-Blocks further decomposes these two sets of features with complete sequence views, yielding more detailed information. Each decomposed sequence also requires attention calculations on dimensional features, considering inter-dimensional interaction attention to enhance prediction accuracy. Finally, the resulting features are recombine and rearranged to obtain the final features.

[0056] ;

[0057] ;

[0058] ;

[0059] ;

[0060] ;

[0061] ;

[0062] in, , These are the odd and even sequences in the original sequence. , , , It is a one-dimensional convolutional filter; This indicates multiplying the elements at corresponding positions of the two matrices; , These are the characteristics resulting from the interaction of odd and even sequences; , This represents the result of feature combination after interaction; Att is the attention calculation operation. , This is the result after dimensional interaction.

[0063] The method described above for capturing multi-target interaction trajectory features is a spatial feature extraction method based on GATv2, and the process is as follows:

[0064] The GATv2 network can calculate the correlation between neighboring points in an image and determine their relative importance to the central node, thereby allocating more resources to it. It treats trajectories within the same scene as nodes in the image and calculates the correlation coefficient between trajectories. Normalization yields the attention coefficients between nodes. Its multi-head attention mechanism further learns multiple nonlinear mappings to simulate dynamic spatial interactions influenced by various factors in different potential subspaces.

[0065] ;

[0066] ;

[0067] in, and The sequence number is considered here as the center point and adjacent points of the graph; and This represents the characteristics of the trajectory after time encoding corresponding to the sequence number; and It is a linear transformation matrix; For nodes and The correlation coefficient; Attention coefficient; This is a set of trajectory sequence numbers.

[0068] In step 2, a multi-target interactive flight trajectory prediction model based on the SCINet network is designed. The model adopts a trajectory spatiotemporal feature distribution law capture method based on CVAE, and the process is as follows:

[0069] CVAE networks introduce latent variables To model the relationship between observed trajectory features and predicted trajectory features, This represents features learned from historical trajectories that are related to the distribution of future trajectories and have randomness. First, the spatiotemporal features of the input trajectory are mapped to a Gaussian distribution with a mean by the encoder. and variance Then, samples were taken from a Gaussian distribution with the same statistical properties. Finally, the decoder parses... Spatiotemporal characteristics of the observed trajectory Generate future trajectory ;

[0070] ;

[0071] ;

[0072] ;

[0073] in, Predicted trajectory features after time modeling; This is the CVAE encoding part; The noise is a standard normally distributed Gaussian noise. This is the CVAE decoding part.

[0074] In step 2, a multi-target interactive flight trajectory prediction model based on the SCINet network is designed. The model training method and training objective are as follows:

[0075] Step 2.1: Design the model loss function;

[0076] The model loss function consists of two parts: 2.1.1, the reconstruction loss represents the loss on the latent variables. The reconstruction loss is calculated by adjusting the mean squared error. A larger value indicates that the predicted trajectory generated by the model is closer to the true trajectory. 2.1.2 KL divergence represents the latent space learned by the model. The difference between the distribution pattern of the actual trajectory and the distribution pattern of the predicted trajectory is ultimately simplified to: Therefore, the model training objective is transformed into minimizing the loss function. ;

[0077] ;

[0078] ;

[0079] ;

[0080] in, This represents the total number of serial numbers. This is the total loss function; For serial number A reconstructed trajectory corresponding to the future trajectory;

[0081] Step 2.2: Model training;

[0082] During the training phase, the observed trajectory set is first processed using the SCITree encoding model. and predicted trajectory set Temporal modeling is performed to obtain a multi-level, long-term dependent set of temporal features. and ;Then, Spatial modeling using GATv2 yields spatiotemporal features of multi-objective interactions. Finally, and Feeding into the CVAE layer to generate latent variables And reconstruct future trajectory location Complete the training and save the model weights.

[0083] Step 3: Input the observed trajectory information used for testing into the trained model to obtain the predicted trajectory;

[0084] In step 3, a multi-target interactive flight trajectory prediction model based on the SCINet network is designed, and the testing process is as follows:

[0085] During the testing phase, the model only applies to the observed trajectory. Spatiotemporal modeling is performed, and its spatiotemporal characteristics are... With Gaussian noise Send them together into CVAE to decode their future trajectory Read the training weights, and feed the test data into the model for testing. The test is based on mean squared error and dynamic time warping.

[0086] In this embodiment, the designed trajectory pre-model, with its rich convolutional filters and excellent global perspective, enables the modeled trajectory features to exhibit multi-level and long-term dependent characteristics. Simultaneously, the Graph Attention Network (GATv2) provides a convenient mechanism for modeling spatial interactions between different trajectories, possessing not only good interpretability but also the ability to dynamically adjust attention weights, enhancing the model's understanding of different target trajectories. All of this provides richer information for the generative model to reconstruct trajectories, thereby improving the accuracy of multi-target trajectory prediction.

[0087] This embodiment focuses on flight trajectory prediction in a multi-objective interactive environment. For air traffic networks with increasing airspace complexity, especially in multi-objective environments, more accurate trajectory prediction performance can help reduce safety issues caused by the surge in air traffic volume and increased airspace complexity. More accurate trajectory prediction can provide stronger decision support for air traffic control, reducing potential flight conflicts and airspace safety problems.

[0088] To verify the feasibility of this invention, performance simulation tests were conducted on a general aviation trajectory dataset. This dataset was captured at a single-runway airport and contains 111 days of trajectory data. Performance testing used the average distance error (ADE) and final distance error (FDE) as metrics. One trajectory point was sampled per second, and an 80-second sequence was predicted using an 8-second sequence. The simulation results are as follows: Figure 2 As shown in the simulation results, compared with the traditional LSTM-based method, the Transformer method improves ADE by 8%, and its superior attention mechanism is better at capturing trends in long sequences. The newer Crossformer further adapts to temporal sequences and focuses on cross-dimensional correlations, further improving ADE by 9%. However, the traditional SCINet, due to its unique sequence decomposition and cross-attention mechanism, can notice deeper details, further improving ADE and FDE prediction performance by 4% and 8% respectively compared to the Crossformer model. Star and TrajAirNet introduce multi-objective interaction to increase the model's spatial modeling ability. Due to the mutual constraints of multiple objectives, more accurate endpoint prediction results are obtained compared with single temporal modeling, improving ADE and FDE by 1.8% and 4.7% respectively compared to Crossformer. This result proves that spatial modeling of multiple objectives can effectively constrain endpoint distance error, while improving long-term modeling ability further reduces the errors of both; combining the characteristics of both, the performance of the method (Our) of this invention is further enhanced.

[0089] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-target interactive flight trajectory prediction method based on SCINet network, characterized in that, The method includes the following steps: Step 1: Obtain flight trajectories from the dataset and establish a multi-target interaction scenario; Step 2: For multi-target interaction scenarios, a model framework for multi-target interaction flight trajectory prediction based on the SCINet network is designed, along with the model training method and training objectives. The model captures the long-term temporal dependence, multi-level features, and multi-target spatial characteristics of the trajectory. The method used to capture the long-term temporal dependence and multi-level features of the trajectory is based on the SCIcross temporal feature extraction method, as follows: SCIcross is composed of multiple SCI-Blocks stacked in a tree structure, performing dimensional interaction calculations on each decomposed sequence. This allows for increasing the stacking depth to enhance model capabilities. First, the input trajectory sequence is decomposed into odd and even sub-sequences, and convolutional filters are used to extract different temporal features from each sub-sequence. Then, the odd sequence is multiplied by the elements of the convolutional even sequence, and the even sequence is multiplied by the elements of the convolutional odd sequence, resulting in two sets of features with complete sequence views. Stacking SCI-Blocks further decomposes these two sets of features with complete sequence views, yielding more detailed information. Each decomposed sequence also requires attention calculations on dimensional features, considering inter-dimensional interaction attention to enhance prediction accuracy. Finally, the resulting features are recombine and rearranged to obtain the final features. F″ odd =That(F′ odd (x),F′ odd (y),F′ odd (z)) F″ even =That(F′ even (x),F′ even (y),F′ even (z)) Among them, F odd F even These are the odd and even sequences in the original sequence. ρ and η are one-dimensional convolution filters; ⊙ represents multiplying corresponding elements of two matrices; Features resulting from the interaction of odd and even sequences; F′ odd F′ even The result is the combination of features after the interaction; Att is the attention computation operation; F″ odd 、F″ even This is the result after dimensional interaction; Step 3: Input the observed trajectory information used for testing into the trained model to obtain the predicted trajectory; In step 1, the scene segmentation method proceeds as follows: To obtain the flight trajectory for a specific time period T during a day from the dataset, use the set A = {a1, a2, ..., a...}. N } represents the sequence number of different trajectories; based on the sequence number, all positions under it are obtained, and each position contains information in three dimensions {x, y, z}, thus obtaining the corresponding trajectory information. Where a i ∈A, the divided observation trajectory is Predicted trajectory t represents the time length of the observed trajectory; the set of trajectories existing at the specified time length T. Consider it as a multi-objective interaction scenario; The method used to capture the interactive trajectory features of multiple targets is a spatial feature extraction method based on GATv2, and the process is as follows: The GATv2 network can calculate the correlation between neighboring points in an image and determine their relative importance to the central node, thereby allocating more resources to it. It treats trajectories within the same scene as nodes in the image and calculates the correlation coefficient between trajectories. Normalization yields the attention coefficients between nodes. Its multi-head attention mechanism further learns multiple nonlinear mappings to simulate dynamic spatial interactions influenced by various factors in different potential subspaces; Among them, a i and a j The sequence number is considered here as the center point and adjacent points of the graph; and This represents the time-encoded features of the trajectory corresponding to the sequence number; W and a are linear transformation matrices; e is the node a. i and a j The correlation coefficient; att is the attention coefficient; A is the set of trajectory sequence numbers.

2. The method for multi-target interactive flight trajectory prediction based on SCINet network according to claim 1, characterized in that, In step 2, the model employs a CVAE-based method for capturing the spatiotemporal feature distribution patterns of trajectories, as follows: The CVAE network models the relationship between observed trajectory features and predicted trajectory features by introducing a latent variable z, which represents a feature learned from historical trajectories that is related to the future trajectory distribution and has randomness. First, the input trajectory spatiotemporal features are mapped to a Gaussian distribution with mean μ and variance σ by the encoder. Then, z is sampled from a Gaussian distribution with similar statistical properties. Finally, the decoder parses z and the observed trajectory spatiotemporal features g. obs Generate future trajectory μ,logσ 2 =E(g obs ,h pred ); z=μ+N noisy ·s; Among them, h pred The predicted trajectory features are modeled over time; E is the CVAE encoded part; N noisy is standard normally distributed Gaussian noise; D is the CVAE decoding part.

3. The method for multi-target interactive flight trajectory prediction based on SCINet network according to claim 1, characterized in that, In step 2, the model training method and training objective are as follows: Step 2.1: Design the model loss function; The model loss function consists of two parts: 2.1.1 Reconstruction loss represents the reconstruction loss for the latent variable z. A larger value indicates that the predicted trajectory generated by the model is closer to the true trajectory. Mean squared error is used to measure this. 2.1.2 KL divergence represents the difference between the distribution of the latent variable z learned by the model and the distribution of the predicted trajectory, which is ultimately simplified to... Therefore, the model training objective is transformed into minimizing the loss function. in, N is the total number of serial numbers; This is the total loss function; For serial number a i A reconstructed trajectory corresponding to the future trajectory; Step 2.2: Model training; During the training phase, the observed trajectory set is first processed using the SCITree encoding model. and predicted trajectory set Temporal modeling is performed to obtain a multi-level, long-term dependent set of temporal features. and Then, h obs Spatial modeling using GATv2 yields spatiotemporal features of multi-objective interactions. Finally, g obs with h pred The data is fed into a CVAE layer to generate latent variable z and reconstruct the future trajectory position. Complete training and save the model weights.

4. The method for multi-target interactive flight trajectory prediction based on SCINet network according to claim 1, characterized in that, In step 3, the testing process is as follows: During the testing phase, the model only applies to the observed trajectory. Perform spatiotemporal modeling, its spatiotemporal characteristics The future trajectory is decoded by feeding it along with Gaussian noise β. Read the training weights and feed the test data into the model for testing. The test is based on mean squared error and dynamic time warping.

Images