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

Abstract

A multi-target interaction flight trajectory prediction method based on an SCINet network belongs to the field of traffic trajectory prediction, and comprises the following steps: step 1, extracting trajectory data to form a multi-target interaction scene; step 2, designing a model framework of multi-target interactive flight trajectory prediction based on an SCINet network; the model captures long-term time dependence, multi-level characteristics and multi-target space characteristics of the trajectory; the method for capturing the long-term time dependence of the trajectory and the multilevel features is a time feature extraction method based on SCIcross; and step 3, inputting the tested observation trajectory information into the trained model to obtain a prediction trajectory. According to the invention, the flight path prediction accuracy under the condition of multi-target interaction of the same scene is improved.

Description

Technical Field

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

[0002] Flight trajectory prediction is a crucial and challenging task in air traffic control, especially in maneuvering control. With the continuous development of the global economy, the demand for air transportation in various industries has increased significantly, the volume of flight traffic has shown a rapid growth trend, and the complexity of airspace has also increased. In this context, the ability to accurately predict flight trajectories has become one of the important links in ensuring airspace safety. Accurate trajectory prediction can not only provide better decision support for air traffic control, but also effectively reduce the risk of air conflicts and collisions, avoid potential traffic accidents, and thus protect the lives of pilots and passengers.

[0003] Trajectory prediction has always been a research topic of concern to many scholars. Traditional methods are generally unable to effectively handle complex nonlinear dynamic changes, especially when faced with multi-dimensional features and long time series data. In addition, these methods often lack the ability to model the spatiotemporal interactions between multi-target trajectories or have insufficient performance. In order to solve these problems, the present 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 mechanism of the SCI network and the efficient spatial relationship modeling of the GATv2 (GraphAttention Network v2) network, it can effectively capture the long-term dependencies and spatial interaction features in the input sequence, thereby improving the accuracy of the prediction.

[0004] To further enhance the performance of the model, the method introduces a conditional variational autoencoder to learn the potential distribution of trajectories, thereby improving the ability to predict future trajectories. CVAE plays an important role in modeling uncertainty and capturing the diversity of trajectory data, which helps to improve the generalization ability and prediction stability of the model, thereby improving the accuracy and reliability of trajectory prediction in the aviation transportation intelligent transportation system. Summary of the invention

[0005] In order to overcome the shortcomings of existing trajectory prediction technology, which is difficult to handle multi-dimensional features and long time series and lacks spatial interaction, this paper proposes a multi-target interactive flight trajectory prediction method based on the SCINet network, which combines the SCINet network that can extract multi-dimensional features with the GATv2 network that can extract complex spatial relationship interactions to achieve spatiotemporal modeling, and introduces the CVAE network to learn the potential laws in the airspace environment to reconstruct the trajectory. It effectively solves the problem of poor trajectory prediction accuracy, realizes multi-target interaction in the same scene, and improves the prediction performance.

[0006] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is specifically as follows: A multi-target interactive flight trajectory prediction method based on SCINet network includes the following steps: Step 1: Obtain flight trajectories from the dataset and establish a multi-target interaction scenario; Step 2: Design a model framework for multi-target interactive flight trajectory prediction based on the SCINet network for multi-target interactive scenarios, and design the model training method and training objectives; the process is as follows: the model captures the long-term temporal dependency, multi-level features, and multi-target spatial characteristics of the trajectory; The method used to capture the long-term temporal dependency and multi-level features of trajectories is based on the temporal feature extraction method of SCIcross. The process is as follows: SCIcross is composed of multiple SCI-Blocks stacked in a tree structure, and each decomposed sequence is interactively calculated in dimensions, allowing the stacking depth to be deepened to improve the model capability. First, the input trajectory sequence is decomposed into odd and even subsequences, and different temporal features are extracted from each subsequence using a convolution filter. Then, the odd sequence is multiplied with the even sequence elements after convolution, and the even sequence is multiplied with the odd sequence elements after convolution, thereby obtaining two sets of features with a complete sequence view. Stacking SCI-Blocks can further decompose the two sets of features with a complete sequence view to obtain more detailed information. Among them, the sequence after each decomposition process also needs to perform attention calculation on the dimensional features, considering the interactive attention between dimensions to enhance the accuracy of the prediction. Finally, the obtained features are recombined and arranged to obtain the final features. ; ; ; ; ; ; in, , is the odd sequence and even sequence in the original sequence; , , , is a one-dimensional convolution filter; It means multiplying the elements of the corresponding positions of the two matrices; , is the characteristic of the interaction between odd and even sequences; , is the result of feature combination after interaction; Att is the attention calculation operation, , is the result after dimension interaction; Step 3: Input the observed trajectory information for testing into the trained model to obtain the predicted trajectory.

[0007] Furthermore, in step 1, a model for multi-target interactive flight trajectory prediction based on the SCINet network is designed, and the process of establishing a multi-target interactive scenario is as follows: Get the time of day from the dataset The flight trajectory of The serial numbers representing different tracks; get all the positions under the serial number, each position contains Information in three dimensions, to obtain corresponding trajectory information ,in , the observation trajectory divided into The predicted trajectory is , The time length of the observation track is set; the time length will be specified There exists a set of trajectories under { } is regarded as a multi-target interaction scenario.

[0008] More preferably, the method for capturing the multi-target interaction trajectory features is a spatial feature extraction method based on GATv2, and the process is as follows: The GATv2 network can calculate the correlation between adjacent points in the image and calculate their importance ratio to the central node, thereby allocating more resources to it; it treats the trajectory in the same scene as a node in the image and calculates the correlation coefficient between trajectories. , normalized to get the attention coefficient between nodes The multi-head attention mechanism contained in it further learns multiple nonlinear mappings to simulate dynamic spatial interactions affected by various factors in different latent subspaces; ; ; in, and is the serial number, which is regarded as the center point and adjacent points of the graph; and Indicates the characteristics of the trajectory time encoding corresponding to the sequence number; and is the linear transformation matrix; For Node and The correlation coefficient of is the attention coefficient; is a set of trajectory sequence numbers.

[0009] In step 2, a model for multi-target interactive flight trajectory prediction based on the SCINet network is designed. The model adopts a method for capturing the spatiotemporal feature distribution law of trajectories based on CVAE. The process is as follows: The CVAE network introduces hidden variables To model the relationship between observed trajectory features and predicted trajectory features, Represents the features related to the future trajectory distribution learned from the historical trajectory, which is random; first, the trajectory spatiotemporal features are input, and the encoder maps the features to a Gaussian distribution with mean and variance ; Then, we sample from a Gaussian distribution with the same statistical properties ; Finally, the decoder parses Spatial and temporal characteristics of the observed trajectory , generating future trajectories ; ; ; ; in, Predicted trajectory features after modeling time; Encoding part for CVAE; is the standard normally distributed Gaussian noise; For the CVAE decoding part.

[0010] Preferably, in step 2, a model for multi-target interactive flight trajectory prediction based on the SCINet network is designed, and the model training method and training target process are as follows: Step 2.1, design model loss function; The model loss function consists of two parts: 2.1.1. Reconstruction loss represents the loss of latent variables The reconstruction loss under , the larger its value, the closer the predicted trajectory generated by the model is to the true trajectory, and the mean square error is used to measure it, that is ; 2.1.2, KL divergence represents the potential space learned by the model The difference between the distribution law of the predicted trajectory and the distribution law of the predicted trajectory is finally simplified to ; Therefore, the model training objective is transformed into minimizing the loss function ; ; ; ; in, is the total number of serial numbers; is the total loss function; The serial number is Reconstructed trajectories corresponding to future trajectories; Step 2.2: Model training; In the training phase, the observation trajectory set is first encoded by the SCITree model. And the predicted trajectory set Perform temporal modeling to obtain a multi-level, long-term dependent temporal feature set and ;Then, Through GATv2 spatial modeling, the spatiotemporal characteristics of multi-target interactions are obtained ; Finally, and Send to CVAE layer to generate hidden variables , and reconstruct the future trajectory position , complete the training and save the model weights.

[0011] In step 3, a model for multi-target interactive flight trajectory prediction based on the SCINet network is designed, and the test process is as follows: During the testing phase, the model only observes the trajectory Perform spatiotemporal modeling, whose spatiotemporal characteristics With Gaussian noise Send them to CVAE to decode the future trajectory ; Read the training weights and send the test data into the model for testing. The test uses mean square error and dynamic time warping as indicators.

[0012] Compared with the prior art, the beneficial effects of the present invention are mainly manifested in: the proposed method can decompose the time series according to different time scales and capture the multi-level spatiotemporal information of the trajectory data; the multi-layer recursive structure and cross-scale attention mechanism can also better model complex long-term and short-term dependencies and improve the prediction accuracy; the proposed method can focus on the spatial relationship between trajectories and handle complex spatial interactions; the proposed method introduces randomness and can maintain high robustness and prediction performance in predictions based on prior knowledge. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of multi-target interactive flight trajectory prediction based on the SCINet network in an embodiment of the present invention; Figure 2 It is a comparison diagram of the multi-target interactive flight trajectory prediction algorithm based on the SCINet network in an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The present invention will be further described below in conjunction with the accompanying drawings.

[0015] Reference Figure 1 , a multi-target interactive flight trajectory prediction method based on SCINet network, comprising the following steps: Step 1: Obtain flight trajectories from the dataset and establish a multi-target interaction scenario; In step 1, a model for multi-target interactive flight trajectory prediction based on the SCINet network is designed, and the process of establishing a multi-target interactive scenario is as follows: Get the time of day from the dataset The flight trajectory of The serial numbers representing different tracks; get all the positions under the serial number, each position contains Information in three dimensions, to obtain corresponding trajectory information ,in , the observation trajectory divided into The predicted trajectory is , The time length of the observation track is set; the time length will be specified There exists a set of trajectories under { } is regarded as a multi-target interaction scenario.

[0016] Step 2: Design a model framework for multi-target interactive flight trajectory prediction based on the SCINet network for multi-target interactive scenarios, and design the model training method and training objectives; In the step 2, a model for multi-target interactive flight trajectory prediction based on the SCINet network is designed to capture the long-term temporal dependency, multi-level features and multi-target spatial characteristics of the trajectory.

[0017] The method used to capture the long-term temporal dependency and multi-level features of trajectories is based on the temporal feature extraction method of SCIcross. The process is as follows: SCIcross is composed of multiple SCI-Blocks stacked in a tree structure, and each decomposed sequence is interactively calculated in dimensions, allowing the stacking depth to be deepened to improve the model capability. First, the input trajectory sequence is decomposed into odd and even subsequences, and different temporal features are extracted from each subsequence using a convolution filter. Then, the odd sequence is multiplied with the even sequence elements after convolution, and the even sequence is multiplied with the odd sequence elements after convolution, thereby obtaining two sets of features with a complete sequence view. Stacking SCI-Blocks can further decompose the two sets of features with a complete sequence view to obtain more detailed information. Among them, the sequence after each decomposition process also needs to perform attention calculation on the dimensional features, considering the interactive attention between dimensions to enhance the accuracy of the prediction. Finally, the obtained features are recombined and arranged to obtain the final features. ; ; ; ; ; ; in, , is the odd sequence and even sequence in the original sequence; , , , is a one-dimensional convolution filter; It means multiplying the elements of the corresponding positions of the two matrices; , is the characteristic of the interaction between odd and even sequences; , is the result of feature combination after interaction; Att is the attention calculation operation, , is the result of dimension interaction.

[0018] The method for capturing multi-target interactive trajectory features is a spatial feature extraction method based on GATv2, and the process is as follows: The GATv2 network can calculate the correlation between adjacent points in the image and calculate their importance ratio to the central node, thereby allocating more resources to it; it treats the trajectory in the same scene as a node in the image and calculates the correlation coefficient between trajectories. , normalized to get the attention coefficient between nodes The multi-head attention mechanism contained in it further learns multiple nonlinear mappings to simulate dynamic spatial interactions affected by various factors in different latent subspaces; ; ; in, and is the serial number, which is regarded as the center point and adjacent points of the graph; and Indicates the characteristics of the trajectory time encoding corresponding to the sequence number; and is the linear transformation matrix; For Node and The correlation coefficient of is the attention coefficient; is a set of trajectory sequence numbers.

[0019] In step 2, a model for multi-target interactive flight trajectory prediction based on the SCINet network is designed. The model adopts a method for capturing the spatiotemporal feature distribution law of trajectories based on CVAE. The process is as follows: The CVAE network introduces hidden variables To model the relationship between observed trajectory features and predicted trajectory features, Represents the features related to the future trajectory distribution learned from the historical trajectory, which is random; first, the trajectory spatiotemporal features are input, and the encoder maps the features to a Gaussian distribution with mean and variance ; Then, we sample from a Gaussian distribution with the same statistical properties ; Finally, the decoder parses Spatial and temporal characteristics of the observed trajectory , generating future trajectories ; ; ; ; in, Predicted trajectory features after modeling time; Encoding part for CVAE; is the standard normally distributed Gaussian noise; For the CVAE decoding part.

[0020] In step 2, a model for multi-target interactive flight trajectory prediction based on the SCINet network is designed, and the model training method and training target process are as follows: Step 2.1, design model loss function; The model loss function consists of two parts: 2.1.1. Reconstruction loss represents the loss of latent variables The reconstruction loss under , the larger its value, the closer the predicted trajectory generated by the model is to the true trajectory, and the mean square error is used to measure it, that is ; 2.1.2, KL divergence represents the potential space learned by the model The difference between the distribution law of the predicted trajectory and the distribution law of the predicted trajectory is finally simplified to ; Therefore, the model training objective is transformed into minimizing the loss function ; ; ; ; in, is the total number of serial numbers; is the total loss function; The serial number is Reconstructed trajectories corresponding to future trajectories; Step 2.2: Model training; In the training phase, the observation trajectory set is first encoded by the SCITree model. And the predicted trajectory set Perform temporal modeling to obtain a multi-level, long-term dependent temporal feature set and ;Then, Through GATv2 spatial modeling, the spatiotemporal characteristics of multi-target interactions are obtained ; Finally, and Send to CVAE layer to generate hidden variables , and reconstruct the future trajectory position , complete the training and save the model weights.

[0021] Step 3: Input the observed trajectory information for testing into the trained model to obtain the predicted trajectory; In step 3, a model for multi-target interactive flight trajectory prediction based on the SCINet network is designed, and the test process is as follows: During the testing phase, the model only observes the trajectory Perform spatiotemporal modeling, whose spatiotemporal characteristics With Gaussian noise Send them to CVAE to decode the future trajectory ; Read the training weights and send the test data into the model for testing. The test uses mean square error and dynamic time warping as indicators.

[0022] In this embodiment, the designed trajectory pre-model, with its rich convolutional filters and good global vision, makes the modeled trajectory features have multi-level and long-term dependency characteristics. At the same time, the graph attention network (GATv2) provides a convenient mechanism for modeling spatial interactions between different trajectories. It not only has good interpretability, but also can dynamically adjust the attention weights to enhance the model's understanding of different target trajectories. This provides richer information for the generative model to reconstruct the trajectory, thereby improving the accuracy of multi-target trajectory prediction.

[0023] This embodiment focuses on flight trajectory prediction in a multi-target interactive environment. For a flight traffic network with increasing airspace complexity, especially in a multi-target environment, more accurate trajectory prediction performance can help reduce safety issues caused by the surge in flight traffic and increased airspace complexity. More accurate trajectory prediction can provide stronger decision support for air traffic control and reduce potential flight conflicts and airspace safety issues.

[0024] To verify the feasibility of the present invention, this embodiment performs a performance simulation test on a general aviation trajectory dataset, which is captured at an airport with a single runway and contains 111 days of data trajectory data. The performance test uses the average distance error (ADE) and the terminal distance error (FDE) as indicators, samples one trajectory point every 1s, and predicts an 80s sequence with an 8s sequence. The simulation results are shown in Figure 2. Figure 2 As shown. From the simulation results, it can be seen that compared with the traditional Lstm-based method, the Transformer method has improved ADE by 8%, and its superior attention mechanism is better at grasping the trend in long sequences. The newer Crossformer has further adapted the time series, focusing on the correlation across dimensions, and further improved ADE by 9%. However, the traditional SCINet, due to its unique sequence decomposition and cross-attention mechanism, can pay attention to deeper details, and further improves the ADE and FDE prediction performance by 4% and 8% on the basis of the Crossformer model. Star and TrajAirNet introduce multi-target interaction to increase the spatial modeling ability of the model. Due to the mutual constraints of multiple targets, more accurate endpoint prediction results are obtained compared with single time modeling, which are 1.8% and 4.7% higher than Crossformer in ADE and FDE respectively. This result proves that spatial modeling of multiple targets can effectively constrain the endpoint distance error, and improving the long-term modeling capability can further reduce the errors of the two; combining the characteristics of the two, the performance of the method (Our) of the present invention is further enhanced.

[0025] 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 principle of the present invention should be included in 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 comprises 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 interactive flight trajectory prediction based on the SCINet network is designed, and the model training method and training objectives are designed; the model captures the long-term temporal dependency, multi-level features and multi-target spatial characteristics of the trajectory; the method used to capture the long-term temporal dependency and multi-level features of the trajectory is based on the time feature extraction method of SCIcross, and the process is as follows: SCIcross is composed of multiple SCI-Blocks stacked in a tree structure, and each decomposed sequence is interactively calculated in dimensions, allowing the stacking depth to be deepened to improve the model capability. First, the input trajectory sequence is decomposed into odd and even subsequences, and different temporal features are extracted from each subsequence using a convolution filter. Then, the odd sequence is multiplied with the even sequence elements after convolution, and the even sequence is multiplied with the odd sequence elements after convolution, thereby obtaining two sets of features with a complete sequence view. Stacking SCI-Blocks can further decompose the two sets of features with a complete sequence view to obtain more detailed information. Among them, the sequence after each decomposition process also needs to perform attention calculation on the dimensional features, considering the interactive attention between dimensions to enhance the accuracy of the prediction. Finally, the obtained features are recombined and arranged to obtain the final features. ； ； ； ； ； ； in, , is the odd sequence and even sequence in the original sequence; , , , is a one-dimensional convolution filter; It means multiplying the elements of the corresponding positions of the two matrices; , is the characteristic of the interaction between odd and even sequences; , is the result of feature combination after interaction; Att is the attention calculation operation, , is the result after dimension interaction; Step 3: Input the observed trajectory information for testing into the trained model to obtain the predicted trajectory.

2. The multi-target interactive flight trajectory prediction method based on SCINet network according to claim 1 is characterized in that: In step 1, the process of the scene division method is as follows: Get the time of day from the dataset The flight trajectory of The serial numbers representing different tracks; get all the positions under the serial number, each position contains Information in three dimensions, to obtain corresponding trajectory information ,in , the observation trajectory divided into The predicted trajectory is , The time length of the observation track is set; the time length will be specified There exists a set of trajectories under { } is regarded as a multi-target interaction scenario.

3. The multi-target interactive flight trajectory prediction method based on SCINet network according to claim 1 is characterized in that: The method used to capture the characteristics of multi-target interaction trajectories is based on the spatial feature extraction method of GATv2. The process is as follows: The GATv2 network can calculate the correlation between adjacent points in the image and calculate their importance ratio to the central node, thereby allocating more resources to it; it treats the trajectory in the same scene as a node in the image and calculates the correlation coefficient between trajectories. , normalized to get the attention coefficient between nodes The multi-head attention mechanism contained in it further learns multiple nonlinear mappings to simulate dynamic spatial interactions affected by various factors in different latent subspaces; ； ； in, and is the serial number, which is regarded as the center point and adjacent points of the graph; and Indicates the characteristics of the trajectory time encoding corresponding to the sequence number; and is the linear transformation matrix; For Node and The correlation coefficient of is the attention coefficient; is a set of trajectory sequence numbers.

4. The multi-target interactive flight trajectory prediction method based on the SCINet network according to claim 1 or 2, characterized in that: In step 2, the model adopts a method for capturing the temporal and spatial distribution law of trajectory features based on CVAE, and the process is as follows: The CVAE network introduces hidden variables To model the relationship between observed trajectory features and predicted trajectory features, Represents the features related to the future trajectory distribution learned from the historical trajectory, which is random; first, the trajectory spatiotemporal features are input, and the encoder maps the features to a Gaussian distribution with mean and variance ; Then, we sample from a Gaussian distribution with the same statistical properties ; Finally, the decoder parses Spatial and temporal characteristics of the observed trajectory , generating future trajectories ; ； ； ； in, Predicted trajectory features after modeling time; Encoding part for CVAE; is the standard normally distributed Gaussian noise; For the CVAE decoding part.

5. The multi-target interactive flight trajectory prediction method based on the SCINet network according to claim 1 or 2, characterized in that: In step 2, the model training method and training target process are as follows: Step 2.1, design model loss function; The model loss function consists of two parts: 2.1.1 Reconstruction loss represents the latent variables The reconstruction loss under , the larger its value, the closer the predicted trajectory generated by the model is to the true trajectory, and the mean square error is used to measure it, that is ; 2.1.2 KL divergence represents the potential space learned by the model The difference between the distribution law of the predicted trajectory and the distribution law of the predicted trajectory is finally simplified to ; Therefore, the model training objective is transformed into minimizing the loss function ; ； ； ； in, is the total number of serial numbers; is the total loss function; The serial number is Reconstructed trajectories corresponding to future trajectories; Step 2.2: Model training; In the training phase, the observation trajectory set is first encoded by the SCITree model. And the predicted trajectory set Perform temporal modeling to obtain a multi-level, long-term dependent temporal feature set and ;Then, Through GATv2 spatial modeling, the spatiotemporal characteristics of multi-target interactions are obtained ; Finally, and Send to CVAE layer to generate hidden variables , and reconstruct the future trajectory position , complete the training and save the model weights.

6. The multi-target interactive flight trajectory prediction method based on SCINet network according to claim 1 or 2, characterized in that: In step 3, the testing process is as follows: During the testing phase, the model only observes the trajectory Perform spatiotemporal modeling, whose spatiotemporal characteristics With Gaussian noise Send them to CVAE to decode the future trajectory ; Read the training weights and feed the test data into the model for testing. The test uses mean square error and dynamic time warping as indicators.

Images