High-speed target sparse flight trajectory data prediction method and system

Abstract

The invention relates to the technical field of air traffic management, in particular to a high-speed target sparse flight trajectory data prediction method and system, and the method comprises the steps: obtaining flight trajectory sample data which comprises complete flight trajectory sample data, point-shaped missing flight trajectory sample data and segment-shaped missing flight trajectory sample data; the point-shaped missing flight path sample data and the segment-shaped missing flight path sample data are obtained by performing random span masking on complete flight path sample data by using a span covering method; constructing a trajectory prediction model based on a Transform encoder, and training the trajectory prediction model by using the flight trajectory sample data; and inputting the flight trajectory data of the to-be-predicted target into the trajectory prediction target model to obtain complete flight trajectory data of the to-be-predicted target. The method has a good prediction effect on the missing flight path data at different moments and continuous moments, and can be applied to scenes such as exploration, search and rescue, fire extinguishing and the like with unlimited channels.

Description

Technical Field

[0001] The present invention relates to the technical field of air traffic management, and in particular to a method and system for predicting sparse flight trajectory data of a high-speed target. Background Art

[0002] Flight trajectory data is a time series data that records the motion state of an aircraft. It is collected by monitoring equipment and is an important data of the air traffic control system (ATCS). It records the spatiotemporal state change process of a moving target and has dynamic laws and spatiotemporal characteristics. By mining and analyzing the motion trajectory, the motion law and behavior pattern of the moving target can be analyzed, and even the goal and motivation of the trajectory action can be obtained. However, in practical applications, due to various reasons such as limited coverage of positioning equipment, equipment failure, poor signal, and transmission loss, the flight trajectory collected by the monitoring equipment usually loses data or has invalid data at certain different or continuous moments, and presents sparse characteristics as a whole. Missing data at different moments appear as point-like missing data, and missing data at continuous moments appear as segment-like missing data. In the clustering study of missing flight trajectory data, it was found that the feature extraction of flight trajectories with continuous long missing segments is more difficult than that of flight trajectories with scattered small missing segments. Missing flight trajectories not only reduce the accuracy and robustness of real-time traffic condition monitoring, but also hinder downstream applications such as air traffic flow prediction, conflict detection, air traffic management, and accident analysis. Missing flight trajectory prediction, as the primary task of data cleaning, has attracted more and more attention from air traffic-related fields around the world. By estimating the missing positions in the flight trajectory, the integrity of the data can be maintained and the development of downstream applications can be better driven.

[0003] The technical routes for missing flight trajectory prediction can be summarized into two types. One is to estimate the missing position of the flight trajectory based on the flight plan and its historical path (early trajectories with the same flight number), such as the air traffic control trajectory interpolation method based on tensor completion. The flight trajectories of the same flight number are organized into a three-dimensional tensor by using the dependency between trajectories, and the interpolation task is achieved through the tensor completion method that minimizes the trace norm. The other is to establish a motion model based on the motion law of the target to estimate the missing flight trajectory, such as the method of interpolating the missing flight trajectory based on the trajectory pre-training model of Bert, randomly cover part of the flight trajectory, and predict the flight trajectory of the covered position based on the uncovered flight trajectory. Through this process, the motion law of the target flight trajectory is learned to achieve the interpolation of the missing flight trajectory. However, the interpolation method for estimating the missing position of the flight trajectory based on the historical path has a strong dependence on historical data. The interpolated target flight trajectory is restricted by the channel, and the spatial characteristics are strictly related to the geographical location. However, in applications such as exploration, search and rescue or fire fighting, the flight trajectory data is not restricted by the channel, the spatial characteristics are not related to the geographical location, and the absolute position of the historical path has no guiding role in the interpolation of the current missing flight trajectory. For example, the same type of target may perform different tasks, and the flight trajectory may be in different longitude and latitude ranges, and there may be transformations such as rotation and scaling. Therefore, a missing flight trajectory prediction method is urgently needed to meet the application of exploration, search and rescue or fire fighting scenarios that are not restricted by the route. Summary of the invention

[0004] To this end, the present invention provides a method and system for predicting sparse flight trajectory data of high-speed targets, which solves the problem that the existing flight trajectory prediction capability of missing continuous moments is insufficient and cannot meet the application requirements of scenarios with restricted routes.

[0005] According to the design scheme provided by the present invention, on the one hand, a method for predicting sparse flight trajectory data of a high-speed target is provided, comprising:

[0006] Acquire flight trajectory sample data, the flight trajectory sample data including complete flight trajectory sample data, point-like missing flight trajectory sample data and segment-like missing flight trajectory sample data, the point-like missing flight trajectory sample data and segment-like missing flight trajectory sample data are obtained by performing random span masking on the complete flight trajectory sample data using a span masking method, the complete flight trajectory is a trajectory with continuous latitude and longitude coordinates and timestamps, the point-like missing means that there is a discontinuous timestamp in the flight trajectory that is missing or invalid, and the segment-like missing means that there is a continuous timestamp in the flight trajectory that is missing or invalid;

[0007] A trajectory prediction model is constructed based on a Transformer encoder, and the trajectory prediction model is trained using the flight trajectory sample data to obtain a trajectory prediction target model, wherein the trajectory prediction model uses position embedding and dimension embedding to extract temporal features and spatial features in the input flight trajectory data, and uses an attention mechanism to learn the correlation between the attributes of the flight trajectory points in the flight trajectory sample data and the reasoning relationship between the flight trajectory points;

[0008] The flight trajectory data of the target to be predicted is input into the trajectory prediction target model, and the trajectory prediction target model is used to obtain the complete flight trajectory data of the target to be predicted.

[0009] As a high-speed target sparse flight trajectory data prediction method of the present invention, further, a span masking method is used to perform random span masking on the complete flight trajectory sample data, including:

[0010] Get complete flight trajectory sample data;

[0011] A span of random length is obtained by randomly sampling a specified geometric distribution, wherein the span length includes but is not limited to 1 under a specified measurement unit; and a span starting point is obtained by sampling from the complete flight trajectory sample data using a uniform distribution based on the span, and a span mask is performed on the corresponding position of the flight trajectory using the span starting point;

[0012] Return to obtain a span of random length and iterate until the flight trajectory meets the missing ratio requirement, so as to obtain point-like missing flight trajectory sample data and segment-like missing flight trajectory sample data through the span mask.

[0013] As a method for predicting sparse flight trajectory data of high-speed targets of the present invention, further, when the span length is 1, sampling is performed from the complete flight trajectory sample data using uniform distribution based on the span and point-like missing flight trajectory sample data is obtained through the span mask; when the span length is greater than 1, sampling is performed from the complete flight trajectory sample data using uniform distribution based on the span and segment-like missing flight trajectory sample data is obtained through the span mask.

[0014] As a high-speed target sparse flight trajectory data prediction method of the present invention, further, the trajectory prediction model is trained using the flight trajectory sample data, including:

[0015] A model training loss function is constructed using flight trajectory prediction loss and covariance loss, wherein the covariance loss is the covariance loss of the moving target attribute in the fused flight trajectory and the target flight trajectory, and the fused flight trajectory is a flight trajectory formed by combining the observed flight trajectory and the predicted flight trajectory;

[0016] The trajectory prediction model is trained based on the model training loss function and using the flight trajectory sample data.

[0017] As a method for predicting sparse flight trajectory data of high-speed targets of the present invention, further, the trajectory prediction model is trained using flight trajectory sample data, and it also includes: selecting a span cover of the flight trajectory, and predicting and training the span cover based on global information by covering two left and right boundary points on the span and combining position embedding, wherein the two left and right boundary points are used to obtain the motion trend of the missing flight trajectory according to the motion law of the boundary points.

[0018] As the high-speed target sparse flight trajectory data prediction method of the present invention, further, the model training loss function table

[0019] L Traj_Bert =α k l imp +β k l cov

[0020] Shown as: α k =2(1-0.5k / epoch), where l imp is the prediction loss, α k is the adaptive weight of the prediction loss, l cov For the cooperation

[0021] β k = k / epoch

[0022] Difference loss, β k is the adaptive weight of the covariance loss, k is the current training round number, and epoch is the total number of training rounds.

[0023] As a method for predicting sparse flight trajectory data of high-speed targets of the present invention, further, the moving target attributes include: dynamic characteristics of the moving target, and the dynamic characteristics include but are not limited to: position of the moving target and speed of the moving target.

[0024] On the other hand, the present invention also provides a high-speed target sparse flight trajectory data prediction system, comprising: a sample acquisition module, a model training module and a trajectory prediction module, wherein:

[0025] A sample acquisition module is used to acquire flight trajectory sample data, wherein the flight trajectory sample data includes complete flight trajectory sample data, point-like missing flight trajectory sample data and segment-like missing flight trajectory sample data, wherein the point-like missing flight trajectory sample data and segment-like missing flight trajectory sample data are obtained by performing random span masking on the complete flight trajectory sample data using a span masking method, wherein the complete flight trajectory is a trajectory with continuous latitude and longitude coordinates and timestamps, wherein the point-like missing refers to the presence of discontinuous timestamps in the flight trajectory that are missing or invalid, and the segment-like missing refers to the presence of continuous timestamps in the flight trajectory that are missing or invalid;

[0026] A model training module is used to build a trajectory prediction model based on a Transformer encoder, and train the trajectory prediction model using flight trajectory sample data to obtain a trajectory prediction target model, wherein the trajectory prediction model uses position embedding and dimension embedding to extract temporal features and spatial features in the input flight trajectory data, and uses an attention mechanism to learn the correlation between various attributes of flight trajectory points in the flight trajectory sample data and the reasoning relationship between flight trajectory points;

[0027] The trajectory prediction module is used to input the flight trajectory data of the target to be predicted into the trajectory prediction target model, and use the trajectory prediction target model to obtain the complete flight trajectory data of the target to be predicted.

[0028] Beneficial effects of the present invention:

[0029] The invention adopts a span masking strategy to expand the granularity of random masking on training data, adopts an optimization target of missing flight trajectory boundary prediction, and enhances the robustness of the interpolation model to segment missing; the Encoder part in the Transformer model is used to construct the model, the position embedding is used to retain the time series information in the flight trajectory data, the flight trajectory points are mapped to a high-dimensional space through dimension embedding to extract the spatial features in the flight trajectory data, and the attention mechanism enables the model to learn the correlation between each flight trajectory point in the flight trajectory and each attribute of each flight trajectory point; the root mean square error between the predicted value and the true value is used as the prediction loss, and the observed value and the predicted missing value are used to form a fused flight trajectory. During the training process, the error of the covariance matrix between the fused flight trajectory and the true flight trajectory is calculated to constrain the predicted missing segment to have the same movement trend as the true flight trajectory, and an adaptive learning strategy is adopted to dynamically adjust the weights of the prediction loss and the covariance loss during the training process, so that the model focuses on different aspects in different training stages, and the flight trajectory data missing at different times and continuous times have good prediction effects, can be applied to the scenarios of exploration, search and rescue, fire fighting, etc. with unrestricted routes, and has good application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a schematic diagram of the high-speed target sparse flight trajectory data prediction process in the embodiment;

[0031] Figure 2 This is a schematic diagram of the trajectory prediction model structure in the embodiment;

[0032] Figure 3 This is an illustration of the span mask principle in the embodiment;

[0033] Figure 4 This is an illustration of boundary prediction in the embodiment;

[0034] Figure 5It is a schematic diagram of the motion characteristics and covariance matrix of different flight trajectories in the embodiment;

[0035] Figure 6 It is a comparison diagram of the loss curves in the embodiment;

[0036] Figure 7 This is a heat map of the prediction span covariance matrix in the embodiment;

[0037] Figure 8 Schematic diagram of the performance comparison trajectory pre-training model structure in the embodiment. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of the present invention clearer and more understandable, the present invention is further described in detail below in conjunction with the accompanying drawings and technical solutions.

[0039] For the prediction of flight trajectories of aircraft with restricted flight paths at different times and / or consecutive times, the present invention provides an embodiment of the present invention, see Figure 1 As shown, a method for predicting sparse flight trajectory data of a high-speed target is provided, comprising:

[0040] S101. Acquire flight trajectory sample data, wherein the flight trajectory sample data includes complete flight trajectory sample data, point-missing flight trajectory sample data and segment-missing flight trajectory sample data. The point-missing flight trajectory sample data and segment-missing flight trajectory sample data are obtained by performing random span masking on the complete flight trajectory sample data using a span masking method. The complete flight trajectory is a trajectory with continuous longitude and latitude coordinates and timestamps. The point-missing means that discontinuous timestamps are missing or invalid in the flight trajectory, and the segment-missing means that continuous timestamps are missing or invalid in the flight trajectory.

[0041] Specifically, the span masking method is used to perform random span masking on the complete flight trajectory sample data, which can be designed to include the following content:

[0042] Get complete flight trajectory sample data;

[0043] A span of random length is obtained by randomly sampling a specified geometric distribution, wherein the span length includes but is not limited to 1 under a specified measurement unit; and a span starting point is obtained by sampling from the complete flight trajectory sample data using a uniform distribution based on the span, and a span mask is performed on the corresponding position of the flight trajectory using the span starting point;

[0044] Return to obtain a span of random length and iterate until the flight trajectory meets the missing ratio requirement, so as to obtain point-like missing flight trajectory sample data and segment-like missing flight trajectory sample data through the span mask.

[0045] Among them, when the span length is 1, the point-like missing flight trajectory sample data is sampled from the complete flight trajectory sample data based on the span using uniform distribution and the span mask is used to obtain the point-like missing flight trajectory sample data. When the span length is greater than 1, the segment-like missing flight trajectory sample data is sampled from the complete flight trajectory sample data based on the span using uniform distribution and the span mask is used to obtain the segment-like missing flight trajectory sample data.

[0046] S102, constructing a trajectory prediction model based on the Transformer encoder, and training the trajectory prediction model using the flight trajectory sample data to obtain a trajectory prediction target model, where Figure 2 As shown, the trajectory prediction model uses position embedding and dimension embedding to extract the temporal features and spatial features in the input flight trajectory data, and uses the attention mechanism to learn the correlation between the attributes of the flight trajectory points in the flight trajectory sample data and the reasoning relationship between the flight trajectory points.

[0047] Specifically, the trajectory prediction model is trained using the flight trajectory sample data, which can be designed to include:

[0048] A model training loss function is constructed using flight trajectory prediction loss and covariance loss, wherein the covariance loss is the covariance loss of the moving target attribute in the fused flight trajectory and the target flight trajectory, and the fused flight trajectory is a flight trajectory formed by combining the observed flight trajectory and the predicted flight trajectory;

[0049] The trajectory prediction model is trained based on the model training loss function and using the flight trajectory sample data.

[0050] In order to improve the robustness of the trajectory prediction model to segmental missing, in this embodiment, the span is masked using the model training method of the flight trajectory random span masking (Span Mask Trajectory Model, SMTM). Random sampling from a geometric distribution obtains a span of random length l~Geo(p), and then uniformly distributes it from Traj i The starting point of the span is obtained by sampling, and this process is iterated until the missing ratio requirement is met, such as Figure 3 As shown in the figure, because the length of the span is randomly selected, the length may be 1 or greater than 1, so the flight trajectories in the training set have both point-like and segment-like missing.

[0051] The missing span boundary is predicted and the span cover of the flight trajectory is selected. Based on the global information, the span cover is trained by covering two boundary points on the left and right and combining position embedding. The two boundary points on the left and right are used to obtain the movement trend of the missing flight trajectory according to the movement law of the boundary points.

[0052] In this embodiment, the optimization goal of missing span prediction (Mask Span Prediction) is proposed based on the attention mechanism, and the missing span is predicted using global information. At the same time, the span double boundary prediction (Span Boundary Prediction, SBP) is introduced to predict the flight trajectory. Select a span to cover Mtraj j ={mtr s ,mtr s+1 ,…mtr e}, using two boundary points on the left and right s-1 ,tr s-2 ,tr e+1 ,tr e+2 Combined with position embedding P i-s+1 Make predictions about span coverage:

[0053] y i =f(tr s-1 ,tr s-2 ,tr e+1 ,tr e+2 ,P i-s+1 )

[0054] There are many ways to choose the number of boundary points. The disadvantage of a single boundary point is that the boundary point will not contain the target's motion trend information, and the relative position of the missing segment cannot be known through the boundary point. Figure 4 As shown in a in . The double boundary points can be used to know the approximate movement trend of the missing flight trajectory according to the movement law of the boundary points, making the prediction result more credible, such as Figure 4 As shown in b.

[0055] In existing research, a normal distribution assumption is made for flight trajectory data. The mean vector and covariance matrix can uniquely determine a high-dimensional normal distribution. If the mean vector and covariance matrix of the fused flight trajectory are the same as those of the target flight trajectory, it can be said that they obey the same distribution. This provides a theoretical basis for using covariance loss as the optimization target.

[0056] In this embodiment, the role of covariance loss is described from another perspective. For moving targets such as airplanes, the covariance matrix can describe the correlation between the various dimensions of the moving target (such as position, speed, etc.), reveal the intrinsic connection between these attributes, and reflect the dynamic characteristics of the moving target. Figure 5As shown in the figure, the covariance matrices of flight trajectories of different shapes are different. By comparing the covariance matrices of different moving targets and extracting their feature information, the classification and recognition of moving targets can be realized. The fused flight trajectory composed of the observed flight trajectory and the predicted flight trajectory can reflect the similarity of the motion characteristics of the two moving targets by using the similarity measurement of the covariance matrix of the fused flight trajectory and the target flight trajectory. Based on this, the error of the corresponding elements in the covariance matrix of the fused flight trajectory and the target flight trajectory is calculated. The covariance loss is used as the optimization target to constrain the prediction of the model to have a similar motion trend as the target flight trajectory.

[0057] In the early training stage of the model, the prediction error is large and the value of covariance loss is unstable. It is expected that the model will focus on reducing the prediction error of each covered flight trajectory point. As the iteration proceeds, when the prediction error of each covered flight trajectory point is small, it is expected that the model's attention will be inclined to constrain the predicted flight trajectory span to have the same motion trend as the covered flight trajectory. Based on this, an adaptive loss function is designed, which can be expressed as:

[0058] L Traj_Bert =α k l imp +β k l cov

[0059] α k =2(1-0.5k / epoch)

[0060] β k = k / epoch

[0061] In the formula, l imp is the prediction loss, α k is the adaptive weight of the prediction loss, l cov is the covariance loss, β k is the adaptive weight of the covariance loss. Initially, the weight of the prediction loss is 1, and the weight of the covariance loss is 0. As the number of training rounds increases, the weight of the prediction loss gradually decreases, and the weight of the covariance loss gradually increases. Finally, the weight of the prediction loss and the weight of the covariance loss reach a balance.

[0062] S103: inputting the flight trajectory data of the target to be predicted into a trajectory prediction target model, and using the trajectory prediction target model to obtain the complete flight trajectory data of the target to be predicted.

[0063] Furthermore, based on the above method, an embodiment of the present invention also provides a high-speed target sparse flight trajectory data prediction system, comprising: a sample acquisition module, a model training module and a trajectory prediction module, wherein:

[0064] A sample acquisition module is used to acquire flight trajectory sample data, wherein the flight trajectory sample data includes complete flight trajectory sample data, point-like missing flight trajectory sample data and segment-like missing flight trajectory sample data, wherein the point-like missing flight trajectory sample data and segment-like missing flight trajectory sample data are obtained by performing random span masking on the complete flight trajectory sample data using a span masking method, wherein the complete flight trajectory is a trajectory with continuous latitude and longitude coordinates and timestamps, wherein the point-like missing refers to the presence of discontinuous timestamps in the flight trajectory that are missing or invalid, and the segment-like missing refers to the presence of continuous timestamps in the flight trajectory that are missing or invalid;

[0065] A model training module is used to build a trajectory prediction model based on a Transformer encoder, and train the trajectory prediction model using flight trajectory sample data to obtain a trajectory prediction target model, wherein the trajectory prediction model uses position embedding and dimension embedding to extract temporal features and spatial features in the input flight trajectory data, and uses an attention mechanism to learn the correlation between various attributes of flight trajectory points in the flight trajectory sample data and the reasoning relationship between flight trajectory points;

[0066] The trajectory prediction module is used to input the flight trajectory data of the target to be predicted into the trajectory prediction target model, and use the trajectory prediction target model to obtain the complete flight trajectory data of the target to be predicted.

[0067] In order to verify the effectiveness of this solution, the following is a further explanation based on experimental data:

[0068] Experimental content: The flight trajectory data used contains five-dimensional attributes such as longitude, latitude, and altitude. The optimization goal during the training process is to reduce the overall error between the predicted trajectory points and the actual trajectory points. The state root mean square error is selected during the test to measure the overall prediction accuracy. The calculation of the state root mean square error does not take into account the influence of the correlation between the attributes in the data. The flight trajectory data follows the dynamic law of the aircraft target. Data of different attributes may be correlated. Using only the state root mean square error to measure the overall prediction accuracy is not accurate enough. For this reason, the Mahalanobis distance that can eliminate the correlation between different attributes is introduced as an evaluation indicator of the overall prediction accuracy. Data of different attributes have different probability distributions. The prediction accuracy of each attribute cannot be known through the overall prediction error. In order to intuitively display the prediction of different attributes, the average relative error is introduced to measure the prediction accuracy of different attributes. In the experiment, the method of controlling variables is used. First, on the same data set, the same model structure is used to verify the effectiveness of the adaptive loss function, boundary prediction, and verification covariance loss proposed in this study. Then, on a mixed data set containing both point-like missing and segment-like missing, after the same number of training rounds, the prediction accuracy is compared with Trajectory-BERT to verify the prediction performance of this case.

[0069] Experimental data: ADS-B data covering 64 days from 2021 to 2022 is used in the range of longitude [-98,-68] and latitude [33,48]. The update interval of the trajectory in this dataset is 20s, including attributes such as timestamp, position and speed. During the data cleaning process, missing values ​​are deleted, and the continuous missing flight trajectories are truncated. After data enhancement by the sliding window method, the entire dataset contains approximately 5 million tracks. After data cleaning, standard deviation normalization is used to standardize the trajectory attributes to unify the data scale and range. In order to evaluate the performance of the model, the trajectories of the first 60 days are selected as the training data set, and the trajectories of the last four days are selected as the test set.

[0070] Evaluation indicators: The state root mean square error, Mahalanobis distance error and mean relative error are used to evaluate the performance of the trajectory prediction algorithm.

[0071] 1. State root mean square error, calculate each state of the predicted flight trajectory The state m corresponding to the actual flight trajectory ij The root mean square error between the predicted state and the actual state reflects the degree of deviation between the predicted state and the actual state. When evaluating the performance of the flight trajectory prediction algorithm, the smaller the state root mean square error, the closer the predicted state is to the actual state. The state root mean square error is calculated as:

[0072]

[0073] In the formula represents the predicted value of the jth state in the missing i-th segment, M ij is the true value of the corresponding state.

[0074] 2. Mahalanobis distance error, based on the evaluation index of Euclidean distance, Euclidean distance does not consider the impact of the correlation between data dimensions on the distance measurement of multivariate data. In the process of Mahalanobis distance calculation, the data is rotated to the principal component to make the dimensions independent of each other. In this way, the Mahalanobis distance can reflect the difference in the variation of each variable, as well as the correlation that sometimes exists, and it is independent of the units used for each variable. Assume that the true state of the span to be predicted And span prediction results The Mahalanobis distance calculation formula is:

[0075]

[0076] where ∑ -1 Get the inverse matrix of the target flight trajectory covariance matrix.

[0077] 3. Average relative error. During the performance test, the flight trajectory data was standardized to eliminate the impact of different dimensional dimensions, but this does not guarantee that the data of different dimensions have the same distribution after standardization. The average relative error can further eliminate the impact of the distribution of data of different dimensions on the performance test. The average relative error of different dimensions is calculated separately to measure the prediction accuracy of the algorithm for data of different dimensions. The average relative error of the kth dimension is calculated as:

[0078]

[0079] In the formula represents the predicted value of the k-th dimension attribute, M k is the corresponding true value.

[0080] Ablation experiment: The missing rate of the test set is set to 30%, 40%, 50% and 60% respectively. Experiments are conducted on test sets with different missing rates to make the experimental results more confident. First, the effectiveness of the adaptive loss function is verified, and then the effectiveness of boundary prediction and covariance loss are verified in turn when the adaptive loss function is used.

[0081] 1. Verification of the effectiveness of adaptive loss function

[0082] On the same dataset, with the same model structure and parameter scale, the loss of model2 is calculated normally: L Traj_Bert = l imp +l cov ; The first 25 rounds of model1 use the adaptive loss function L Traj_Bert =α k l imp +β k l cov , the adaptive weight α after 25 rounds k and β k The values ​​of are all iterated to 1, which is the same as the weight of each loss in model1. Figure 6 It can be seen that during the training process, model1 converges faster than model2, the convergence process is smoother, and the final test error is lower and the training level is better. Tables 1 and 2 compare the prediction accuracy of model1 and model2 on test sets with different missing rates. The results show that the prediction longitude of model1 in the three test indicators of state root mean square error, Mahalanobis distance error and mean absolute error is higher than that of model2, which fully demonstrates that the adaptive loss function proposed in this study is beneficial to model training.

[0083] Table 1 Overall prediction error

[0084]

[0085] Table 2 Prediction error of each attribute

[0086]

[0087] 2. Validation of Boundary Prediction

[0088] Tables 3 and 4 are the ablation experiment results of boundary prediction. Model 1 adds a boundary prediction method based on model 2. Table 3 shows the state root mean square error and Mahalanobis distance error to measure the overall prediction accuracy. The model with the boundary prediction layer has higher prediction accuracy at various missing rates than the model without boundary prediction. Table 4 shows the mean absolute error of the prediction accuracy of different attributes at different missing rates. Similar to the ablation experiment results of the adaptive loss function, the prediction error of lon is the smallest and vel is the largest for each model in the five-dimensional attribute. This indirectly reflects the different distribution of data of different attributes, but the prediction accuracy of the model with the boundary prediction layer is still higher than that without the boundary prediction layer. Judging from the overall prediction error and the prediction error results of each attribute, adding a boundary prediction layer does improve the prediction accuracy of the model.

[0089] Table 3 State root mean square error and Mahalanobis distance error

[0090]

[0091] Table 4 Mean absolute error

[0092]

[0093] 3. Verification of the effectiveness of covariance loss

[0094] The covariance matrices of flight trajectories of different shapes are different. The higher the prediction accuracy, the more similar the shape of the predicted flight trajectory is to the true flight trajectory, and thus the covariance matrix is ​​closer. Figure 6 The following figure shows the prediction results of model1 and model2 for three missing spans in the same flight trajectory. To make the comparison more intuitive, the covariance matrix of each prediction result is displayed as a heat map. Compared with model1, model2 adds covariance loss during training, such as Figure 7 As shown in the figure, the covariance matrix of the predicted span of model2 with covariance loss is closer to the covariance matrix of the true span. Tables 5 and 6 show the experimental results of the two models on the test set. It can be seen from the table that the prediction accuracy of model2 in various evaluation indicators is higher than that of model1. The above experimental results show that adding covariance loss can effectively constrain the predicted results to have similar shape characteristics to the true flight trajectory.

[0095] Table 5 State RMS error and Mahalanobis distance error

[0096]

[0097] Table 6 Mean absolute error

[0098]

[0099] 4. Performance comparison with Trajectory-BERT

[0100] In the BERT-based trajectory pre-training model, the masked trajectory is input, the training model predicts the missing position based on the context information, and outputs the predicted value of the corresponding mask position. The overall framework of the model is as follows Figure 8 As shown in the figure, position embedding retains the temporal information of the flight trajectory data, and dimension embedding maps low-dimensional features to higher dimensions to assist the model in better extracting data features. The MTM layer predicts the position of the Mask, and the parallel FC-Decoder layer restores the entire input trajectory. The overall loss includes three parts: covariance matrix (Loss1), Decoder (Loss2), and MTM (Loss3). The covariance loss is calculated on the flight trajectory restored by the Decoder layer, while the covariance loss of this solution is calculated on the fused trajectory.

[0101] The prediction accuracy of Trajectory-BERT and the proposed solution was tested on a mixed test set containing point-like and segment-like missing information. The results are shown in Tables 7 and 8. Model 1 is the prediction result of Trajectory-BERT, and model 2 is the prediction result of the proposed method.

[0102] Table 7 State RMS error and Mahalanobis distance error

[0103]

[0104] Table 8 Mean absolute error

[0105]

[0106] The above data show that the solution in this case can be applied to the prediction of missing flight trajectories at different times and / or continuous times, can meet the application needs in scenarios with unrestricted routes, and has good application prospects.

[0107] Unless otherwise specifically stated, the relative steps, numerical expressions and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present invention.

[0108] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0109] The units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation is not considered to be beyond the scope of the present invention.

[0110] Those skilled in the art will appreciate that all or part of the steps in the above method can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk or an optical disk. Optionally, all or part of the steps in the above embodiment can also be implemented using one or more integrated circuits, and accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or in the form of software function modules. The present invention is not limited to any specific form of combination of hardware and software.

[0111] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for predicting sparse flight trajectory data of high-speed targets, characterized in that: Include: Acquire flight trajectory sample data, the flight trajectory sample data including complete flight trajectory sample data, point-like missing flight trajectory sample data and segment-like missing flight trajectory sample data, the point-like missing flight trajectory sample data and segment-like missing flight trajectory sample data are obtained by performing random span masking on the complete flight trajectory sample data using a span masking method, the complete flight trajectory is a trajectory with continuous latitude and longitude coordinates and timestamps, the point-like missing means that there is a discontinuous timestamp in the flight trajectory that is missing or invalid, and the segment-like missing means that there is a continuous timestamp in the flight trajectory that is missing or invalid; A trajectory prediction model is constructed based on a Transformer encoder, and the trajectory prediction model is trained using the flight trajectory sample data to obtain a trajectory prediction target model, wherein the trajectory prediction model uses position embedding and dimension embedding to extract temporal features and spatial features in the input flight trajectory data, and uses an attention mechanism to learn the correlation between the attributes of the flight trajectory points in the flight trajectory sample data and the reasoning relationship between the flight trajectory points; The flight trajectory data of the target to be predicted is input into the trajectory prediction target model, and the trajectory prediction target model is used to obtain the complete flight trajectory data of the target to be predicted.

2. The high-speed target sparse flight trajectory data prediction method according to claim 1, characterized in that: The span masking method is used to perform random span masking on the complete flight trajectory sample data, including: Get complete flight trajectory sample data; A span of random length is obtained by randomly sampling a specified geometric distribution, wherein the span length includes but is not limited to 1 under a specified measurement unit; and a span starting point is obtained by sampling from the complete flight trajectory sample data using a uniform distribution based on the span, and a span mask is performed on the corresponding position of the flight trajectory using the span starting point; Return to obtain a span of random length and iterate until the flight trajectory meets the missing ratio requirement, so as to obtain point-like missing flight trajectory sample data and segment-like missing flight trajectory sample data through the span mask.

3. The high-speed target sparse flight trajectory data prediction method according to claim 2, characterized in that: When the span length is 1, the complete flight trajectory sample data is sampled using uniform distribution based on the span and the point-like missing flight trajectory sample data is obtained through the span mask. When the span length is greater than 1, the complete flight trajectory sample data is sampled using uniform distribution based on the span and the segment-like missing flight trajectory sample data is obtained through the span mask.

4. The high-speed target sparse flight trajectory data prediction method according to claim 1 or 2, characterized in that: Use flight trajectory sample data to train the trajectory prediction model, including: A model training loss function is constructed using flight trajectory prediction loss and covariance loss, wherein the covariance loss is the covariance loss of the moving target attribute in the fused flight trajectory and the target flight trajectory, and the fused flight trajectory is a flight trajectory formed by combining the observed flight trajectory and the predicted flight trajectory; The trajectory prediction model is trained based on the model training loss function and using the flight trajectory sample data.

5. The high-speed target sparse flight trajectory data prediction method according to claim 4, characterized in that: The trajectory prediction model is trained using flight trajectory sample data, and the method also includes: selecting the span cover of the flight trajectory, and performing prediction training on the span cover based on global information by covering two boundary points on the left and right sides of the span and combining position embedding. The two boundary points on the left and right sides are used to obtain the movement trend of the missing flight trajectory according to the movement law of the boundary points.

6. The high-speed target sparse flight trajectory data prediction method according to claim 4, characterized in that: The model training L Traj_Bert ＝α k l imp +β k l cov The loss function is expressed as: α k =2(1-0.5k / epoch), where l imp is the prediction loss, α k is the adaptive weight of the prediction loss, β k = k / epoch l cov is the covariance loss, β k is the adaptive weight of the covariance loss, k is the current training round number, and epoch is the total number of training rounds.

7. The high-speed target sparse flight trajectory data prediction method according to claim 4, characterized in that: The moving target attributes include: moving target dynamic features, and the dynamic features include but are not limited to: moving target position and moving target speed.

8. A high-speed target sparse flight trajectory data prediction system, characterized in that: It includes: sample acquisition module, model training module and trajectory prediction module, among which, A sample acquisition module is used to acquire flight trajectory sample data, wherein the flight trajectory sample data includes complete flight trajectory sample data, point-like missing flight trajectory sample data and segment-like missing flight trajectory sample data, wherein the point-like missing flight trajectory sample data and segment-like missing flight trajectory sample data are obtained by performing random span masking on the complete flight trajectory sample data using a span masking method, wherein the complete flight trajectory is a trajectory with continuous latitude and longitude coordinates and timestamps, wherein the point-like missing refers to the presence of discontinuous timestamps in the flight trajectory that are missing or invalid, and the segment-like missing refers to the presence of continuous timestamps in the flight trajectory that are missing or invalid; A model training module is used to build a trajectory prediction model based on a Transformer encoder, and train the trajectory prediction model using flight trajectory sample data to obtain a trajectory prediction target model, wherein the trajectory prediction model uses position embedding and dimension embedding to extract temporal features and spatial features in the input flight trajectory data, and uses an attention mechanism to learn the correlation between various attributes of flight trajectory points in the flight trajectory sample data and the reasoning relationship between flight trajectory points; The trajectory prediction module is used to input the flight trajectory data of the target to be predicted into the trajectory prediction target model, and use the trajectory prediction target model to obtain the complete flight trajectory data of the target to be predicted.

9. An electronic device, characterized in that: include: at least one processor, and a memory coupled to the at least one processor; The memory stores a computer program, and the computer program can be executed by the at least one processor to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed, the method according to any one of claims 1 to 7 can be implemented.

Images