A highly perceptual channel attention trajectory prediction method

Abstract

The application discloses a highly-perceptional channel attention trajectory prediction method, and relates to the technical field of aviation, comprising the following steps: S1, collecting historical trajectories and environmental information of an airplane; S2, defining trajectory prediction as constructing a prediction model by using an encoder module, a decoder module and a training module under the historical trajectories and the environmental information of the airplane; S3, generating a probability distribution of future trajectories by using the prediction model, defining model input by using a historical observation time window and a prediction time window, and generating a prediction result.The highly-perceptional channel attention trajectory prediction method can ensure that the model learns an effective latent space distribution in the training process by using a conditional variational autoencoder and an optimization loss function, so that diversified prediction of future trajectories is realized, the method is suitable for various aviation scenarios, can provide more accurate support for aviation traffic management and trajectory planning, and has a wide practical application prospect.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of aviation technology, in particular to a height-aware channel attention trajectory prediction method. BACKGROUND

[0002] Air traffic refers to the movement of aircraft between the air and the ground and its management, including civil aviation, air freight and military aviation, etc. The main purpose of air traffic is to ensure the safe, efficient and orderly operation of aircraft. Air traffic management is the core of air traffic, responsible for monitoring and guiding aircraft flight in the air.

[0003] With the increasing busy of air traffic, accurate prediction of aircraft trajectory becomes increasingly important. The existing technology usually relies on relatively simple time series models such as recurrent neural network (RNN) and long short-term memory (LSTM) network. There are problems of insufficient processing capacity for historical data and insufficient response to environmental factors. In addition, in the feature extraction and model training process of traditional methods, the height information of the trajectory cannot be effectively utilized, resulting in insufficient accuracy and reliability of the trajectory prediction.

[0004] Therefore, a height-aware channel attention trajectory prediction method is proposed to solve or alleviate the above problems. SUMMARY

[0005] The purpose of the present application is to provide a height-aware channel attention trajectory prediction method to solve the problems raised in the background.

[0006] To achieve the above purpose, the present application provides the following technical solution: a height-aware channel attention trajectory prediction method, comprising the following specific steps:

[0007] S1: Collecting the historical trajectory and environmental information of the aircraft to provide data for predicting the future trajectory of the aircraft;

[0008] S2: Defining trajectory prediction as constructing a prediction model with an encoder module, a decoder module and a training module under the historical trajectory and environmental information of the aircraft, comprising the following steps:

[0009] S2.1: Extracting three-dimensional historical trajectory and height features of the aircraft using a time domain convolution network, extracting wind information features using a convolutional neural network, and optimizing the spatial variation characteristics of wind;

[0010] S2.2: Fusing trajectory height features with trajectory representation, including high-dimensional trajectory representation channel longitudinal splicing, transverse splicing and height feature channel attention, selecting height feature channel attention mechanism, dynamically adjusting the channel contribution of trajectory feature representation, and enhancing the attention of the model to key features;

[0011] S2.3: Learning the probability distribution of the track representation using the conditional variational autoencoder, generating the Verlet acceleration, and deriving the predicted track coordinates through Verlet integration, inputting the integrated historical track and environmental information, and predicting the future track using the latent space distribution;

[0012] S2.4: Optimizing and training the prediction model using the trajectory prediction loss and the conditional variational autoencoder loss;

[0013] S3: Generating the probability distribution of the future track through the prediction model constructed above, defining the model input through the historical observation time window and the prediction time window, and generating the prediction result.

[0014] Preferably, the historical track in S1 refers to the specific path of the aircraft during flight, including takeoff point, flight route, flight altitude, and landing point information. The historical track collection sources include but are not limited to aircraft navigation system recorded data, black box stored data, ground radar and satellite monitoring system collected data. The environmental information includes terrain, map, weather and wind information.

[0015] Preferably, the process of extracting and encoding the three-dimensional historical trajectory data in S2.1 through the time domain convolution network is as follows:

[0016]

[0017]

[0018] wherein, and represent the hidden state of the three-dimensional historical track and the hidden state of the future track of the aircraft, represents the observed time, represents the time domain convolution network, and represent the historical track features and future track features extracted and encoded by the time domain convolution network. Preferably, the process of extracting wind information features by the convolutional neural network in S2.1 is as follows:

[0019]

[0020]

[0021] wherein, represents the extracted wind information features, represents the convolutional neural network, represents the feature extraction function.

[0022] ​Preferably, in S2.2, the vertical stitching of the high-dimensional track representation channel involves independently extracting features from the altitude data and fusing them into the track representation, thus integrating the extracted altitude representation... Directly spliced ​​to the flattened high-dimensional track representation Following this, the data is input into the post-processor model, where the height features are vertically stitched together with the high-dimensional track representation. Following the channel, the longitudinal stitching process of the high-dimensional track representation channel is shown in the following formula:

[0023]

[0024]

[0025] in, Indicates the first The aircraft's flight path altitude data, This indicates that the flight path is highly concealed. Indicates the highly extracted feature function. This is a flattening operation function. This is for splicing operations.

[0026] Preferably, the horizontal stitching of the high-dimensional track representation channel in S2.2 is to combine the height representation after feature extraction. spliced ​​into a high-dimensional trajectory representation At the end of each channel, that is, horizontally spliced ​​to the high-dimensional track representation. Behind each passage, we obtained The lateral stitching process of the high-dimensional track representation channel is shown in the following formula:

[0027]

[0028]

[0029] in, Indicates the first The aircraft's flight path altitude data, This indicates that the flight path is highly concealed. Indicates the highly extracted feature function. Indicates airplane The first in the high-dimensional representation space of the trajectory Channel vectors, For the stitching operation, the data is then processed by activation function and layer normalization. The introduction of activation function increases the nonlinearity of the track representation. After that, the data from each channel is flattened to obtain a new track representation for subsequent model processing. The process is shown in the following formula:

[0030]

[0031] wherein, is a flattening operation function, is a layer normalization operation, is an activation function, is a learnable weight matrix, is a learnable bias vector.

[0032] Preferably, the high-dimensional feature channel attention in S2.2 is to convert the dimensions of the extracted track height features through a fully connected layer, map them to the same channel dimension, and then use the Sigmoid function to obtain the channel attention weight of the track high-dimensional space representation. Then multiply these weights with the track high-dimensional representation space channel by channel to obtain a new track representation. Finally, the track representation channel data constructed is spliced to obtain the output of the channel attention module. The above process is shown in the following formula:

[0033]

[0034]

[0035] wherein, is an activation function, is a channel-wise multiplication, that is, each element in the weight vector is element-wise multiplied with the first row of the matrix .

[0036] Preferably, the generated Verlet acceleration in S2.3 is to integrate the historical track data and environmental information into , which is represented by the following formula:

[0037]

[0038] The conditional variational autoencoder CVAE is a generative model composed of an encoder and a decoder architecture. The encoder accepts two inputs: the first is a comprehensive representation vector , which contains historical track and environmental information, and the second is a real future trajectory data table vector . The task of the encoder is to map this information to a distribution in the latent space, which captures the relationship between the input track data and the latent space and learns the internal structure of the data. This process is represented by the following formula:

[0039]

[0040] wherein, ​​​represents an encoder, the input of the decoder is sampled from the latent space distribution learned from the encoder and encodes the historical input As conditional information, the future track information obtained by training the decoder together The process is represented by the following formula:

[0041]

[0042] represents the decoder;

[0043] And incorporate the motion state of the aircraft, using the dynamics formula to predict the Verlet acceleration, the output of the CVAE first needs to pass through an MLP layer to get the appropriate dimension Verlet acceleration tensor, represented by the following formula:

[0044]

[0045] wherein, is a multi-layer perceptron, according to the Verlet acceleration tensor, the Verlet integral deduction calculation to predict the track coordinates, represented by the following formula:

[0046]

[0047] wherein, represents the range of aircraft numbers, represents the time range, represents the track coordinate position at time step, represents the track coordinate position at the next time step, represents the track coordinate position at the previous time step, represents the acceleration at time step, represents the time step.

[0048] Preferably, the trajectory prediction loss in S2.4 is measured by calculating the mean square error MSE between the model predicted trajectory and the real trajectory, the mathematical formula of trajectory prediction loss is as follows:

[0049]

[0050] wherein, is the mean square error function, the CVAE loss includes two parts: one part is the reconstruction loss, and the other part is the KL divergence loss, which is used to measure the difference between the latent space distribution learned by the model and the pre-assumed standard normal distribution, the mathematical formula of CVAE loss is as follows:

[0051]

[0052] in, The latent spatial distribution learned by the CVAE encoder. It is a standard normal distribution. Let KL divergence be the denoting factor. The total loss of the prediction model is a weighted sum of the trajectory prediction loss and the CVAE loss, as shown in the following equation:

[0053]

[0054] During the training process of the prediction model, by minimizing the trajectory prediction loss and CVAE loss error, diverse future trajectories that match the real situation are generated.

[0055] Preferably, the prediction model constructed in S3 is based on the distribution formula for the future trajectory built in step S2, using... and They represent airplanes Location information and in Real-time environmental information, using and This represents all aircraft in the current scenario. During the observation period All trajectory and environmental information, the trajectory prediction task represents as historical trajectory and environmental information The following formula predicts the distribution of future flight paths:

[0056]

[0057] in, This indicates the size of the historical observation time window. The size of the observation time window is represented by the feature extraction of track data and environmental information. Simultaneously, the height features are encoded by TCN. Furthermore, the track encoding information is weighted and fused with historical track data through a channel attention mechanism. The conditional variational autoencoder uses the future track encoding features as a supervision condition, and the historical track encoding and environmental information are used as inputs to the CVAE to learn the spatial distribution of track data, predict the Verlet acceleration of the future track, and finally calculate the position of the predicted trajectory through the Verlet integral algorithm to obtain the prediction result.

[0058] The technical effects and advantages of this invention are as follows:

[0059] (1) The highly perceptual channel attention track prediction method is used, the conditional variational autoencoder and the optimization loss function are used, the effective potential space distribution learned by the model in the training process is ensured, diversified prediction of future tracks is realized, the method is suitable for various aviation scenes, more accurate support can be provided for aviation traffic management, track planning and the like, and has a wide practical application prospect.

[0060] (2) The highly perceptual channel attention track prediction method is used, time domain convolution network and convolutional neural network are used for feature extraction, the complex time sequence relationship of track data and wind information is effectively captured, the sensitivity of the model to key features is improved, the highly feature channel attention mechanism is introduced, the contribution of different feature channels to the final track feature representation is dynamically adjusted, the model pays more attention to the features that have a greater impact on the prediction result, and the learning ability and generalization ability of the model are improved. BRIEF DESCRIPTION OF DRAWINGS

[0061] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, used to explain the present application, and do not constitute a limitation on the present application, in the drawings:

[0062] Figure 1 The highly perceptual channel attention track prediction method flow chart of the present application is shown in the figure.

[0063] Figure 2 The prediction model construction flow chart of the present application is shown in the figure.

[0064] Figure 3 The high-dimensional track representation channel longitudinal splicing diagram of the present application is shown in the figure.

[0065] Figure 4 The high-dimensional track representation channel transverse splicing diagram of the present application is shown in the figure.

[0066] Figure 5 The highly feature channel attention diagram of the present application is shown in the figure. DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0068] The present application provides a highly perceptual channel attention track prediction method as shown in Figures 1-5 The specific steps are as follows:

[0069] S1: Collect the historical flight path and environmental information of the aircraft, which is used to provide data for predicting the future flight path of the aircraft;

[0070] S2: Define the flight path prediction as constructing a prediction model with an encoder module, a decoder module and a training module under the historical flight path and environmental information of the aircraft, including the following steps;

[0071] S2.1: Extract the three-dimensional historical flight path and height features of the aircraft using a time domain convolution network, extract the wind information features using a convolutional neural network, and optimize the spatial variation characteristics of the wind;

[0072] S2.2: Fuse the flight path height features with the flight path representation, including high-dimensional flight path representation channel longitudinal splicing, transverse splicing and height feature channel attention, select the height feature channel attention mechanism, dynamically adjust the channel contribution of the flight path feature representation, and enhance the attention of the model to key features;

[0073] S2.3: Use the conditional variational autoencoder to learn the probability distribution of the flight path representation, generate Verlet acceleration, Verlet is a numerical integration algorithm that predicts the position of the next time by using the position and acceleration of the particle, and deduces the predicted flight path coordinates by Verlet integration, inputs the integrated historical flight path and environmental information, and uses the latent space distribution to predict the future flight path;

[0074] S2.4: Use trajectory prediction loss and conditional variational autoencoder loss to optimize and train the prediction model;

[0075] S3: Generate the probability distribution of the future flight path by the prediction model constructed above, define the model input by the historical observation time window and the prediction time window, and generate the prediction result.

[0076] Specifically, the historical flight path in S1 refers to the specific path of the aircraft during flight, including takeoff point, flight route, flight altitude, landing point information, and the historical flight path collection source includes but is not limited to aircraft navigation system record data, black box storage data, ground radar and satellite monitoring system collected data, environmental information includes terrain, map, weather and wind information, and this method only considers the wind factor in the environmental information.

[0077] Further, for the feature extraction of the trajectory position information, compared with the traditional time sequence model based on recurrent neural network such as LSTM and GRU, the structure of the time domain convolutional neural network TCN is simpler and clearer, and its processing advantage over the former in the time sequence problem has been proved, so the model uses TCN to extract the features of the three-dimensional position information of the trajectory, and the ACTrajNet also uses TCN to realize the extraction of the single height feature, and benefits from the excellent properties of the space-time convolution, the dilated convolution and the residual connection in the structure of the time domain convolutional neural network, so that the TCN has better experimental effect in capturing the height characteristics of the trajectory compared with other networks, and the process of extracting and encoding the three-dimensional historical trajectory data in S2.1 by the time domain convolutional network is as follows:

[0078]

[0079]

[0080] wherein, and respectively represent the hidden state of the three-dimensional historical trajectory and the hidden state of the future trajectory of the aircraft, represents the observed time, represents the time domain convolutional network, and respectively represent the historical trajectory feature and the future trajectory feature extracted and encoded by the time domain convolutional network. Further, for the feature extraction of the environmental information, the model only considers the influence of the wind on the aircraft in the horizontal section, including the runway direction and the wind speed in the transverse direction of the runway, in order to fully utilize the spatial distribution characteristics contained in the wind data and extract useful information, the model uses a convolutional neural network (CNN) to extract the features of the wind variables, which is more suitable for data distribution with small data volume and strong global dependence than a multilayer perceptron (MLP), and does not explicitly consider the spatial relationship between elements in the data, CNN can better capture the local spatial mode of the dynamic continuity of the wind around the aircraft through the design of local connection and parameter sharing, which is helpful to understand the spatial variation characteristics of the wind, and the process of extracting the wind information features by the convolutional neural network in S2.1 is as follows:

[0081]

[0082]

[0083] wherein, represents the extracted wind information features, represents the convolutional neural network, represents a feature extraction function.

[0084] ​Furthermore, due to the diversity of general aviation flight missions and the flexibility of aircraft operation, general aviation tracks contain rich flight pattern information in their vertical sections. This model aims to design an effective input data fusion encoding method for track altitude information, enabling the model to fully utilize the potential flight pattern information in track altitude features to guide the model in better track prediction. Therefore, this section designs three fusion encoding methods for track altitude features and track representations, and ultimately selects the most effective altitude feature channel attention module. The three altitude combination mechanisms are: Vertical Channel Concatenation (VCC), Horizontal Channel Concatenation (HCC), and Altitude Feature Channel Attention (ACA). Since altitude information has certain application value in reflecting the flight pattern of tracks, this model specifically performs independent feature extraction on altitude data and integrates it into the track representation. In S2.2, the vertical concatenation of the high-dimensional track representation channel extracts independent features from altitude data and integrates them into the track representation, thus integrating the extracted altitude features into the track representation. Directly spliced ​​to the flattened high-dimensional track representation Following this, the data is input into the post-processor model, where the height features are vertically stitched together with the high-dimensional track representation. Following the channel, the longitudinal stitching process of the high-dimensional track representation channel is shown in the following formula:

[0085]

[0086]

[0087] in, Indicates the first The aircraft's flight path altitude data, This indicates that the flight path is highly concealed. This indicates a highly extracted feature function, corresponding to Figure 3 High-resolution feature extractor in the middle, This is a flattening operation function. This is for splicing operations.

[0088] Furthermore, the horizontal stitching of the high-dimensional track representation channel in S2.2 is to combine the height representation after feature extraction. spliced ​​into a high-dimensional trajectory representation At the end of each channel, that is, horizontally spliced ​​to the high-dimensional track representation. Behind each passage, we obtained The lateral stitching process of the high-dimensional track representation channel is shown in the following formula:

[0089]

[0090]

[0091] wherein, represents the first path height data of the aircraft, represents the path height hidden state, represents the height extraction feature function, represents the aircraft the first channel vector in the high-dimensional path representation space, is a splicing operation, the high-dimensional path representation channel transverse splicing structure diagram is shown in Figure 4 , this splicing method strengthens the information proportion of height features in each channel of the high-dimensional path representation, then, after adjusting the channel scale through the full connection layer, the information in the channel is further fused, then after the activation function and layer normalization processing, the introduction of the activation function improves the nonlinearity of the path representation, so that the model can learn more complex representations, at the same time, the information in the path representation is more sparse, which not only improves the computational efficiency of the model and reduces the risk of overfitting, but also enables the model to pay more attention to important features in the path, through the normalization processing, the internal covariate shift in the training process can be reduced, and the training efficiency and stability of the model can be improved, then the channel data is flattened to obtain a new path representation for subsequent model processing, the process is shown in the following formula:

[0092]

[0093] wherein, is a flattening operation function, is a layer normalization operation, is an activation function, is a learnable weight matrix, is a learnable bias vector.

[0094] Further, the last height information combination mechanism, height feature channel attention, is also the combination mechanism finally selected by ACTrajNet model, which plays an important role in the model, the module structure diagram is as shown in Figure 5As shown, the high feature channel attention in S2.2 is to convert the extracted height features of the track through a fully connected layer to the same channel dimension, then use the Sigmoid function to obtain the channel attention weight of the track high-dimensional space representation, and then multiply these weights with the track high-dimensional representation space channel by channel to obtain a new track representation. Finally, the newly constructed track representation channel data is spliced to obtain the output of the channel attention module. The above process is shown in the following formula:

[0095]

[0096]

[0097] wherein, is an activation function, is a channel-by-channel multiplication, that is, each element in the weight vector with a length of is multiplied element by element with the first row of the matrix with a row number of . The model designs three height information combination mechanisms with the original track, and the height feature channel attention combination mechanism has the optimal effect, so the height feature channel attention will be used as the final height combination module of the model. This method can dynamically adjust the contribution of different channels to the final track feature representation according to the vertical section characteristics of the historical aircraft track in the current time period, enhance the attention of the model to the key channel, and enable the subsequent backbone network to better mine the high semantic characteristics such as flight phase, flight mode and flight trend from the track, thereby providing support for better prediction.

[0098] Further, the generated Verlet acceleration in S2.3 is to integrate the historical track data and environmental information into , which is represented by the following formula:

[0099]

[0100] The conditional variational autoencoder CVAE is a generative model with encoder and decoder architecture. The encoder accepts two inputs: the comprehensive representation vector containing historical track and environmental information, and the real future trajectory data table vector . The task of the encoder is to map these information to a distribution in the latent space. Usually, this distribution is assumed to be a Gaussian distribution. The goal of this step is to capture the relationship between the input track data and the latent space and learn the internal structure of the data, which is represented by the following formula:

[0101]

[0102] where, represents the encoder, the input of the decoder is sampled from the latent space distribution learned from the encoder , and the historical input encoding As conditional information, the decoder is trained to obtain future track information The task of the decoder is to map the sampling points in the latent space back to the track data representation space, trying to reconstruct the future trajectory information matching the input condition, which is represented by the following formula:

[0103]

[0104] represents the decoder, it is worth noting that the probability distribution of the decoder will gradually approach the preset standard normal distribution during the training process, so as to guide the model to encode the track data into a latent space close to the standard normal distribution, so that in the test stage, the encoder of the CVAE can be omitted, and the CVAE is a conditional variational autoencoder, directly sampling from the standard normal distribution , combined with the historical input encoding, and input into the decoder of the CVAE to predict the future track, which not only simplifies the generation process, but also ensures that the generated samples can cover the latent space learned during training;

[0105] and the motion state of the aircraft, the Verlet acceleration is predicted by adopting the dynamic formula, which is a forward Verlet integral, which provides a constant speed model when the acceleration is zero, the CVAE does not directly predict the position of the track, but predicts the Verlet acceleration, and then calculates the absolute position of each agent, this method allows ACTrajNet not only to capture the probability distribution of the trajectory, but also to consider the physical dynamics of the aircraft, providing a more in-depth understanding and accuracy for trajectory prediction, the output of the CVAE first needs to pass through an MLP layer to obtain a Verlet acceleration tensor with appropriate dimensions, which is represented by the following formula:

[0106]

[0107] where, is a multi-layer perceptron, according to the Verlet acceleration tensor, the Verlet integral is deduced to calculate the predicted track coordinates, which is represented by the following formula:

[0108]

[0109] where, represents the range of aircraft numbers, represents the time range, represents the Time step trajectory coordinate position, represents the next time step trajectory coordinate position, represents the last time step trajectory coordinate position, represents the acceleration at the time step, represents the time step length.

[0110] In particular, the trajectory prediction loss in S2.4 is measured by calculating the mean square error (MSE) between the model predicted trajectory and the real trajectory, and part of the goal is to directly optimize the prediction accuracy of the model for the future trajectory, the trajectory prediction loss mathematical formula is as follows:

[0111]

[0112] wherein, is the mean square error function, and the CVAE loss includes two parts: one part is the reconstruction loss, which is also contained in the trajectory prediction loss, and the other part is the KL divergence loss, which is used to measure the difference between the latent space distribution learned by the model and the pre-assumed standard normal distribution, the goal of the KL divergence loss is to make the latent space distribution learned by the model close to the prior distribution, prevent the model from learning too much detail and noise, and thus improve the generalization ability of the model and avoid overfitting, the CVAE loss mathematical formula is as follows:

[0113]

[0114] wherein, is the latent space distribution learned by the CVAE encoder, is the standard normal distribution, represents the KL divergence, and the total loss of the prediction model is the weighted sum of the trajectory prediction loss and the CVAE loss, as shown in the following formula:

[0115]

[0116] During the training process of the prediction model, the trajectory prediction loss and the CVAE loss are minimized, and the model learns how to generate a diversified future trajectory consistent with the real situation under the condition of a given past trajectory.

[0117] In particular, the prediction model constructed in S3 is constructed by using the distribution formula of the future trajectory constructed in S2 step, and represent the position information of the aircraft and the The environmental information at any given time typically includes terrain, map, weather, and wind information. However, this method only considers the wind factor within the environmental information. and This represents all aircraft in the current scenario. During the observation period All trajectory and environmental information, the trajectory prediction task represents as historical trajectory and environmental information The following formula predicts the distribution of future flight paths:

[0118]

[0119] in, This indicates the size of the historical observation time window. The size of the observation time window is represented by the feature extraction of track data and environmental information. Simultaneously, the height features are encoded by TCN. Furthermore, the track encoding information is weighted and fused with historical track data through a channel attention mechanism. The conditional variational autoencoder uses the future track encoding features as a supervision condition, and the historical track encoding and environmental information are used as inputs to the CVAE to learn the spatial distribution of track data, predict the Verlet acceleration of the future track, and finally calculate the position of the predicted trajectory through the Verlet integral algorithm to obtain the prediction result.

[0120] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. 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 highly perceptual channel attention trajectory prediction method, characterized in that, The method comprises the following specific steps: S1: collecting the historical flight path and environmental information of the aircraft to provide data for predicting the future flight path of the aircraft; S2: defining the flight path prediction as constructing a prediction model by using an encoder module, a decoder module and a training module under the historical flight path and environmental information of the aircraft, comprising the following steps; S2.1: extracting the three-dimensional historical flight path and height features of the aircraft by using a time domain convolution network, and extracting the wind information features by using a convolutional neural network, and optimizing the change characteristics of the wind in space; S2.2: fusing the flight path height features and the flight path representation, including high-dimensional flight path representation channel longitudinal splicing, transverse splicing and height feature channel attention, selecting the height feature channel attention mechanism, dynamically adjusting the channel contribution of the flight path feature representation, and enhancing the attention of the model to the key features; S2.3: learning the probability distribution of the flight path representation by using a conditional variational autoencoder, generating a Verlet acceleration, and deducing the predicted flight path coordinates by using Verlet integration, inputting the integrated historical flight path and environmental information, and predicting the future flight path by using the latent space distribution; S2.4: optimize training of the prediction model using a trajectory prediction loss and a conditional variational autoencoder loss, the trajectory prediction loss is measured by computing the mean squared error, MSE, between the predicted trajectory by the model and the real trajectory, the trajectory prediction loss is mathematically formulated as follows: where, is the mean squared error function, the CVAE loss includes two parts: one part is the reconstruction loss, and the other part is the KL divergence loss, which is used to measure the difference between the latent space distribution learned by the model and the pre-assumed standard normal distribution. The mathematical formula of the CVAE loss is as follows: wherein, is the latent space distribution learned by the CVAE encoder, is a standard normal distribution, denotes the KL divergence, and the total loss of the prediction model is a weighted sum of the trajectory prediction loss and the CVAE loss, as shown in the following equation: In the training process of the prediction model, the trajectory prediction loss and the CVAE loss error are minimized to generate diversified future trajectories consistent with the actual situation; S3: generating the probability distribution of the future flight path by using the prediction model constructed above, defining the model input by using the historical observation time window and the prediction time window, and generating the prediction result.

2. The highly perceptual channel attention trajectory prediction method according to claim 1, wherein, The historical flight path in S1 refers to the specific path passed by the aircraft during flight, including takeoff point, flight route, flight height and landing point information, the historical flight path collection source includes aircraft navigation system record data, black box storage data and ground radar and satellite monitoring system collected data, and the environmental information includes terrain, map, weather and wind information.

3. The highly perceptual channel attention trajectory prediction method according to claim 2, wherein, The three-dimensional historical trajectory data in S2.1 is extracted by a time domain convolution network, as shown in the following formula: wherein, and denote the hidden state of the three-dimensional historical trajectory and the hidden state of the future trajectory of the airplane respectively, denotes the observed time, denotes a temporal convolutional network, and denote the encoded historical trajectory feature and the future trajectory feature extracted by the temporal convolutional network respectively.

4. The highly perceptual channel attention trajectory prediction method according to claim 3, wherein, The process of extracting wind information features by using a convolutional neural network in S2.1 is as shown in the following formula: wherein, represents the extracted wind information features, represents a convolutional neural network, represents a feature extraction function.

5. The highly perceptual channel attention trajectory prediction method according to claim 4, wherein, The vertical stitching of the high-dimensional track representation channel in S2.2 involves independently extracting features from altitude data and fusing them into the track representation, thus integrating the extracted altitude representation. Directly spliced ​​to the flattened high-dimensional track representation Following this, the data is input into the post-processor model, where the height features are vertically stitched together with the high-dimensional track representation. Following the channel, the longitudinal stitching process of the high-dimensional track representation channel is shown in the following formula: wherein, represents the represents the represents the represents the is a flattening operation function, is a stitching operation.

6. The highly perceptual channel attention trajectory prediction method according to claim 5, wherein, The horizontal stitching of the high-dimensional trajectory representation channel in S2.2 is to combine the height representation after feature extraction. spliced ​​into a high-dimensional trajectory representation At the end of each channel, that is, horizontally spliced ​​to the high-dimensional track representation. Behind each passage, we obtained The lateral stitching process of the high-dimensional track representation channel is shown in the following formula: wherein, represents the path height data of the aircraft, represents a path height hidden state, represents a height extraction feature function, represents an aircraft path height vector in the path height feature space, path height vector in the path height feature space, is a concatenation operation, followed by an activation function and layer normalization processing. The introduction of the activation function improves the nonlinearity of the path representation. Then, the channel data is flattened to obtain a new path representation for subsequent model processing. The process is shown in the following formula: wherein, is a flattening operation function, is a layer normalization operation, is an activation function, is a learnable weight matrix, is a learnable bias vector.

7. The highly perceptual channel attention trajectory prediction method according to claim 6, wherein, The height feature channel attention in S2.2 is to convert the dimension of the extracted flight path height features by using a full connection layer, map it to the same channel dimension, then use the Sigmoid function to obtain the channel attention weight of the flight path high-dimensional space representation, and then multiply these weights with the flight path high-dimensional representation space channel by channel to obtain new flight path representation, finally, the channel data of the newly constructed flight path representation is spliced to obtain the output of the channel attention module, and the above process is as shown in the following formula: wherein, is an activation function, is a channel-wise multiplication, i.e. each element of a weight vector is element-wise multiplied with the first row of a matrix of size . ​​ 8. The highly perceptual channel attention trajectory prediction method according to claim 7, wherein, The generating Verlet acceleration in S2.3 is to integrate the historical trajectory data and environmental information into This process is represented by the following equation: Conditional Variational Autoencoder (CVAE) is a generative model consisting of an encoder and a decoder architecture. The encoder receives two inputs: a synthesized representation vector and a decoder. It contains two types of data: historical flight paths and environmental information, and a vector table of real future flight path data. The encoder's task is to map this information to a distribution in the latent space. The above involves capturing the relationship between input track data and the potential space, and learning the inherent structure of the data. This process is represented by the following formula: where, represents an encoder, the input to the decoder is sampled from the latent space distribution learned from the encoder and encodes the historical inputs The decoder is trained jointly with the future trajectory information as conditional information This process is represented by the following equation: represents a decoder; And the motion state of the aircraft is adopted, the Verlet acceleration is predicted by using a dynamics formula, and the output of the CVAE needs to pass through an MLP layer first to obtain a dimensional Verlet acceleration tensor, which is represented by the following formula: wherein, is a multi-layer perceptron, the predicted trajectory coordinates are calculated from the Verlet acceleration tensor, Verlet integration, using the following equation: wherein, represents a range of aircraft numbers, represents a range of times, represents at a time step trajectory coordinate position, represents at a next time step trajectory coordinate position, represents at a previous time step trajectory coordinate position, represents at an acceleration at a time step, represents a time step length.

9. The highly perceptual channel attention trajectory prediction method according to claim 8, wherein, The prediction model constructed in S3 above is a distribution formula of future trajectories constructed by S2 step, using and represent the position information of the aircraft and the environmental information at the time , using and represent all the trajectories and environmental information of all aircrafts in the observation time under the current scene, the trajectory prediction task is represented as predicting the distribution of future trajectories under the historical trajectory and environmental information , as shown in the following formula: wherein, represents the size of the historical observation time window, represents the size of the observation time window, through feature extraction of track data and environmental information, and high-level features encoded by TCN, the historical track data is weighted and fused through the channel attention mechanism of the encoded information, the future track encoding features are taken as the supervision condition by the conditional variational autoencoder, and the historical track encoding and environmental information are taken as the input of the CVAE to learn the spatial distribution of the track data, the Verlet acceleration of the future track is predicted and generated, and finally the position of the predicted trajectory is calculated through the Verlet integral algorithm to obtain the prediction result.

Images