Airplane trajectory prediction method based on space-time attention and dynamic optimization bidirectional LSTM

Abstract

The application belongs to the technical field of aircraft trajectory prediction, and specifically discloses an aircraft trajectory prediction method based on space-time attention and dynamic optimization bidirectional LSTM, which comprises the following steps: collecting multi-dimensional feature data of an aircraft, calculating the horizontal speed and vertical speed of the aircraft based on the difference of the feature data, determining fusion features based on the horizontal speed, vertical speed and feature data, performing noise reduction processing and feature normalization on the fusion features to obtain a normalized sequence, inputting the normalized sequence into an improved LSTM model to obtain a predicted trajectory, and performing Kalman filtering on the predicted trajectory to obtain a smoothed trajectory after smoothing correction. The application improves the accuracy and reliability of trajectory prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of aircraft trajectory prediction, and more particularly relates to an aircraft trajectory prediction method based on space-time attention and dynamic optimization bidirectional LSTM. BACKGROUND

[0002] Aircraft trajectory prediction technology is a core link of air traffic management, which realizes accurate prediction of future motion state through time sequence modeling analysis of multi-dimensional parameters such as aircraft position, speed and attitude, and provides key decision support for conflict resolution, air traffic flow scheduling and safety monitoring.

[0003] The prior art generally uses a unidirectional long short-term memory (LSTM) network to extract time sequence features, and completes model training based on a mean square error (MSE) loss function, and the technical implementation covers standardized processes such as data collection, feature engineering, model iteration and trajectory generation.

[0004] However, this scheme has significant technical defects: unidirectional LSTM can only capture forward time sequence dependency, and cannot model the space-time coupling characteristics of historical state and potential future trajectory (such as the dynamic correlation between heading angle and position change during aircraft maneuvering), resulting in cumulative effects of prediction deviation in complex scenarios; the fixed learning rate optimization strategy exacerbates the problems of slow model convergence speed and local optimal trap, significantly reducing training efficiency; in addition, the traditional method does not introduce a dynamic feature attention mechanism, and the response to key events such as heading angle mutation and height sudden change has a lag, and the imbalance in feature importance allocation causes the prediction error to expand; more importantly, the MSE loss function directly calculates the latitude and longitude error in the plane Euclidean distance, ignoring the nonlinear space mapping relationship caused by the curvature of the earth, resulting in nonlinear amplification of the prediction error in high-latitude regions with the distance; at the same time, the existing technology lacks a dynamic post-processing optimization mechanism, and the predicted trajectory has noise interference and local jitter problems, resulting in insufficient trajectory smoothness and continuity in actual application. The above defects seriously limit the trajectory prediction accuracy, algorithm robustness and engineering application value.

[0005] Therefore, how to improve the trajectory prediction accuracy of the aircraft is a technical problem that needs to be solved at present. SUMMARY

[0006] In view of the defects of the prior art, the purpose of the present application is to provide an aircraft trajectory prediction method based on space-time attention and dynamic optimization bidirectional LSTM, aiming to solve the problem of poor prediction accuracy of aircraft trajectory in the prior art.

[0007] To achieve the above purpose, in a first aspect, the present application provides an aircraft trajectory prediction method based on space-time attention and dynamic optimization bidirectional LSTM, comprising:

[0008] Collect multi-dimensional feature data of the aircraft, calculate horizontal speed and vertical speed of the aircraft based on difference of the feature data, determine fusion features based on the horizontal speed, the vertical speed and the feature data;

[0009] Perform noise reduction processing and feature normalization on the fusion features to obtain a normalized sequence;

[0010] Input the normalized sequence into an improved LSTM model to obtain a predicted trajectory;

[0011] Perform Kalman filtering on the predicted trajectory to obtain a smoothed trajectory after smoothing correction;

[0012] The LSTM model is obtained by bidirectional time sequence modeling, dynamic weighting of spatio-temporal attention mechanism, mixed geographic space loss function and dynamic learning rate optimization training.

[0013] Optionally, the method for obtaining the predicted trajectory comprises:

[0014] Extract forward features from the normalized sequence by a forward LSTM, extract backward features from the normalized sequence by a backward LSTM, and obtain splicing features based on the forward features and the backward features;

[0015] Dynamically weight the splicing features by a spatio-temporal attention mechanism to obtain time step weights, and weight all time step weights to obtain a global context representation;

[0016] Combine the global context and the latest time step features of the LSTM by a fully connected layer to obtain a predicted trajectory covering all features.

[0017] Optionally, the process of dynamic learning rate optimization comprises:

[0018] Initialize a maximum learning rate and a minimum learning rate, and determine an annealing period of model iteration;

[0019] Adjust the learning rate according to the maximum learning rate, the minimum learning rate, the current iteration number and the annealing period.

[0020] Optionally, the method for constructing the mixed geographic space loss function comprises:

[0021] Calculate the geographic error between the predicted position and the real position by using Haversine distance;

[0022] Calculate the motion parameter error except the position parameter by using mean absolute error;

[0023] Weight and fuse the geographic error and the motion parameter error to construct a mixed loss function.

[0024] Optionally, the process of dynamically weighting the spatio-temporal attention mechanism comprises:

[0025] The spliced features are combined with the attention weight matrix, the bias term and the learnable compression matrix by using an activation function to obtain a time step weight.

[0026] The spliced features of all time steps are weighted and summed based on the time step weight to obtain a global context representation.

[0027] Optionally, the Kalman filtering of the predicted trajectory to obtain a smoothed trajectory after smoothing comprises:

[0028] A kinematic model of the aircraft is constructed to define a state vector, a uniform motion matrix, a process noise covariance, an observation matrix and an observation noise covariance.

[0029] The Kalman gain is determined according to the observation matrix, the prior state covariance and the observation noise covariance according to the predicted trajectory input into the kinematic model.

[0030] The posterior state of the state vector is determined based on the Kalman gain, the observation input, the observation matrix and the prior state estimation of the state vector to obtain the smoothed trajectory after smoothing.

[0031] Optionally, the noise reduction processing and feature normalization of the fusion features to obtain a normalized sequence comprises:

[0032] The size of the time window is determined, and a linear decay weight corresponding to the target time is determined based on the size of the time window.

[0033] The noise weighted average value of the target time is determined based on the linear decay weight, the fusion features and the size of the time window.

[0034] The feature data of longitude and latitude are robustly normalized by quartile range, and the state parameters are normalized by maximum and minimum value, and the normalized sequence is obtained by comprehensively normalizing all feature data.

[0035] Optionally, the feature data comprises longitude, latitude, height, heading angle, pitch angle, roll angle, airspeed, horizontal speed and vertical speed.

[0036] The application provides an electronic device, comprising: at least one memory for storing a program; at least one processor for executing the program stored in the memory, when the program stored in the memory is executed, the processor is used to execute the method described in the first aspect or any possible implementation manner of the first aspect.

[0037] The application provides a computer readable storage medium, which stores a computer program, and when the computer program runs on a processor, the processor executes the method described in the first aspect or any possible implementation manner of the first aspect.

[0038] The application provides a computer program product, and when the computer program product runs on a processor, the processor executes the method described in the first aspect or any possible implementation manner of the first aspect.

[0039] It can be understood that the beneficial effects of the second aspect to the fifth aspect can be referred to the related description in the first aspect, and will not be repeated here.

[0040] Overall, compared with the prior art, the above technical solutions conceived by the application have the following beneficial effects:

[0041] (1) The application introduces bidirectional LSTM for time series modeling, fully captures the time before and after the correlation information of the flight trajectory, dynamically weights combined with the spatio-temporal attention mechanism, and enhances the understanding and expression ability of the model for the complex dynamic characteristics of the aircraft. At the same time, the Haversine distance and MAE mixed geographic space loss function are used, so that the model can more accurately reflect the actual spatial error on the curved surface of the earth during the training process, thereby significantly improving the accuracy and reliability of the trajectory prediction.

[0042] (2) The application adjusts the dynamic learning rate periodically, effectively avoids falling into local optimum or oscillation in the training process, and improves the efficiency of gradient descent. This strategy enables the model to quickly approach the global optimal region in the early training stage, and then finely adjusts the parameters in the later stage, improves the training convergence speed and stability, significantly shortens the training time and saves computing resources, and balances the model performance and training cost.

[0043] (3) The application uses the Haversine geographic loss function, so that the model can more reasonably handle the nonlinear spatial relationship on the curved surface of the earth, and avoid the error accumulation bias problem caused by the traditional Euclidean distance. At the same time, the features are processed by dimensionless normalization, which effectively reduces the differences between features of different dimensions and scales, and suppresses the nonlinear accumulation effect of errors in each dimension. Through the above two measures, the robustness and stability of the geographic spatial error are significantly improved, so as to ensure the physical reasonableness and accuracy of the trajectory prediction result. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is one of the flowcharts of the aircraft trajectory prediction method based on spatio-temporal attention and dynamic optimization bidirectional LSTM provided by the embodiments of the application;

[0045] Figure 2Figure 2 is a flowchart of a method for aircraft trajectory prediction based on spatio-temporal attention and dynamic optimization bidirectional LSTM according to an embodiment of the present application;

[0046] Figure 3 Figure 3 is a diagram of an LSTM network structure according to an embodiment of the present application;

[0047] Figure 4 Figure 4 is a diagram of a network training result according to an embodiment of the present application;

[0048] Figure 5 Figure 5 is a diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0049] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application.

[0050] The term "and / or" used herein is a description of an association relationship between associated objects, which means that there can be three relationships, for example, A and / or B can mean that there are three cases of A alone, A and B together, and B alone. The symbol " / " used herein represents an or relationship between associated objects, for example, A / B represents A or B.

[0051] The terms "first" and "second" and the like in the description and claims herein are used to distinguish different objects, rather than to describe a specific order of the objects. For example, the first response message and the second response message are used to distinguish different response messages, rather than to describe a specific order of the response messages.

[0052] In the embodiments of the present application, the words "exemplary" or "for example" are used to mean serving as an example, instance, or illustration. Any embodiment or design presented as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or advantageous than other embodiments or design solutions. Rather, the use of "exemplary" or "for example" is intended to present concepts in a concrete manner.

[0053] In the description of the embodiments of the present application, unless otherwise specified, the meaning of "a plurality of" is two or more, for example, a plurality of processing units means two or more processing units, and the like; a plurality of elements means two or more elements, and the like.

[0054] The embodiments of the present application will be described below with reference to the accompanying drawings in the embodiments of the present application.

[0055] Reference Figure 1The application provides an aircraft trajectory prediction method based on space-time attention and dynamic optimization bidirectional LSTM, comprising:

[0056] S101. Collecting multi-dimensional feature data of the aircraft, calculating the horizontal speed and vertical speed of the aircraft based on the difference of the feature data, and determining the fusion features based on the horizontal speed, vertical speed and feature data;

[0057] S102. Noise reduction processing and feature normalization are performed on the fusion features to obtain a normalized sequence;

[0058] S103. The normalized sequence is input into an improved LSTM model to obtain a predicted trajectory;

[0059] S104. Kalman filtering is performed on the predicted trajectory to obtain a smoothed trajectory after smoothing correction;

[0060] The LSTM model is obtained by bidirectional time series modeling, dynamic weighting of space-time attention mechanism, mixed geographic space loss function and dynamic learning rate optimization training.

[0061] Specifically, in the above step S101, multi-dimensional flight feature data (such as longitude, latitude, height, acceleration, etc.) is collected, and the horizontal speed and vertical speed of the aircraft are accurately extracted by using the difference calculation method. The direct calculation method based on the original data can retain the dynamic characteristics of the flight state, and avoid the lag error that may be introduced by the traditional speed estimation method. By fusing the calculated speed information with other original features, a comprehensive feature vector containing kinematic characteristics is constructed, providing more comprehensive input information for subsequent modeling.

[0062] In step S102, noise reduction processing can use wavelet transform or Kalman filtering method, and noise suppression is performed on different feature dimensions, especially for the features obtained by differential calculation such as speed. This targeted noise reduction strategy can retain effective signals while significantly reducing the interference of high-frequency noise on model training. Feature normalization is performed according to the physical meaning and numerical range of each feature using Min-Max or Z-score normalization method, which solves the problem of non-uniform dimension of multi-dimensional features.

[0063] Further, in step S103, the improved LSTM model captures the forward and backward time sequence dependence of the flight trajectory through the bidirectional structure, solving the problem of insufficient utilization of historical information in traditional unidirectional LSTM. The introduction of space-time attention mechanism enables the model to dynamically adjust the attention weight of different time points and different features, especially in key maneuvers such as sharp turns and climbs, automatically enhancing the contribution of related features, and significantly improving the prediction accuracy of complex actions.

[0064] The model training adopts a hybrid geospatial loss function, which combines the Haversine spherical distance loss and the MAE loss to directly optimize the geospatial error of trajectory prediction. The dynamic learning rate strategy automatically adjusts the learning rate size according to the training process, quickly converges in the early training stage, and finely tunes the parameters in the later stage. These improvements enable the model to maintain training efficiency while having stronger modeling capability for the spatiotemporal characteristics of flight trajectories, especially in long-time flight and complex maneuver scenarios.

[0065] Finally, through step S104, Kalman filter post-processing checks and corrects the predicted trajectory output by the LSTM based on the kinematic model of the aircraft. By establishing a state space model containing position and velocity, the Kalman gain is used to adaptively balance the weights of model prediction and observation data, effectively suppressing the non-physical fluctuations that may exist in LSTM prediction. This post-processing method based on kinematic constraints significantly improves the smoothness and continuity of the trajectory.

[0066] The final smoothed trajectory not only retains the accurate maneuver characteristics of LSTM prediction, but also conforms to the physical motion law of the aircraft, significantly improving the overall prediction results in terms of accuracy and reasonableness. By combining data-driven and model-driven methods, the advantages of both methods are fully utilized.

[0067] Specifically, referring to Figure 2 , Figure 2 is a complete flowchart of the embodiments of the present application, including the following steps:

[0068] Start;

[0069] Data input;

[0070] Multi-feature fusion and preprocessing;

[0071] Improved bidirectional LSTM network;

[0072] Hybrid geospatial loss function;

[0073] Dynamic learning rate optimization training;

[0074] Trajectory prediction output;

[0075] Kalman filter post-processing;

[0076] Final smoothed trajectory;

[0077] End.

[0078] The embodiments of the present application introduce bidirectional LSTM for time series modeling, fully capture the time before and after the correlation information of the flight trajectory, dynamically weight combined with the spatio-temporal attention mechanism, and enhance the understanding and expression ability of the model for the complex dynamic characteristics of the aircraft. At the same time, the Haversine distance and the MAE mixed geographical space loss function are used, so that the model can more accurately reflect the actual spatial error on the curved surface of the earth during the training process, thereby significantly improving the accuracy and reliability of the trajectory prediction.

[0079] Optionally, the method for obtaining the predicted trajectory comprises:

[0080] The normalized sequence is extracted by a forward LSTM to obtain forward features, and the normalized sequence is extracted by a backward LSTM to obtain backward features, and the splicing features are obtained based on the forward features and the backward features;

[0081] The splicing features are dynamically weighted by a spatio-temporal attention mechanism to obtain time step weights, and the global context representation is obtained by weighting and fusing all the time step weights;

[0082] The global context and the recent time step features of the LSTM are combined by a fully connected layer to obtain the predicted trajectory covering all the features.

[0083] Referring to Figure 3 , Figure 3 is an LSTM network structure diagram of the embodiments of the present application.

[0084] The preprocessed 9-dimensional features are input into the input layer as input data;

[0085] Forward LSTM layer → forward hidden state;

[0086] Reverse LSTM layer → reverse hidden state;

[0087] Bidirectional feature splicing;

[0088] Spatio-temporal attention weighting;

[0089] Fully connected output layer

[0090] Predicted trajectory.

[0091] Specifically, the normalized multi-dimensional feature sequence (including position, speed, etc.) is first input into the forward and backward LSTM networks. The forward LSTM processes the sequence in time order, capturing the influence of historical information on the current state; the backward LSTM processes in reverse order, learning the constraints of future state on the current time.

[0092] Input normalized sequence ( is the sequence length), the features are extracted by the forward and backward LSTM, and the splicing is :

[0093]

[0094] where, denotes the input normalized sequence of multiple dimensions, denotes the time step, R denotes the real number set, denotes the forward feature, denotes the backward feature, denotes the fusion feature, denotes the feature dimension.

[0095] It should be noted that the forward and backward LSTM networks both adopt 128-dimensional hidden layers to selectively remember long-term dependencies through the gating mechanism, and are particularly sensitive to capturing the triggering features of sudden turns and other maneuvering actions.

[0096] The spatio-temporal attention mechanism calculates the importance weight of each time step feature through a learnable parameter matrix ht . Specifically, the "query-key value" mode is adopted: first, the 256-dimensional feature is mapped to the 128-dimensional latent space by , and then the attention score is generated through the activation of and the linear transformation of tanh .

[0097] Optionally, the process of dynamic weighting of the spatio-temporal attention mechanism includes:

[0098] The activation function is used to combine the spliced features, the attention weight matrix, the bias term, and the learnable compression matrix to obtain the time step weight;

[0099] Based on the time step weight, the spliced features of all time steps are weighted and summed to obtain the global context representation.

[0100] Specifically, the time step weight is calculated to strengthen the key states (such as turns and climbs):

[0101]

[0102] where, is a learnable weight matrix, tanh denotes the activation function, denotes the bias of the weight matrix, is the weight of the time step t.

[0103] Finally, the prediction layer inputs the global context c and the recent time step feature hT in parallel to form a 512-dimensional mixed feature vector.​c provides overall motion trend, hT retains the latest state details, both complementing to avoid feature dilution problem in long-term prediction. By Wo ∈R9×512weight matrix mapping, the final output contains 9-dimensional prediction values such as longitude, latitude, height, etc. The geometric meaning can be interpreted as local state fine-tuning under the constraint of global optimal trajectory.

[0104] Combining global context and recent state, the prediction result covering all features is generated:

[0105]

[0106] wherein, , respectively represent the weight and bias of the output layer, c is the global context vector, is the feature of the nearest time step.

[0107] Optionally, the process of dynamic learning rate optimization comprises:

[0108] Initialize the maximum learning rate and the minimum learning rate, and determine the annealing period of model iteration;

[0109] Adjust the learning rate according to the maximum learning rate, the minimum learning rate, the current iteration number and the annealing period.

[0110] Specifically, the dynamic learning rate optimization training of the embodiment of the application avoids the model from falling into local optimum and accelerates the convergence process through periodic learning rate. The maximum learning rate , the minimum learning rate , the learning rate is adjusted periodically according to the Cosine function:

[0111]

[0112] wherein, is the current iteration number, is the annealing period, and the learning rate is reset to after completing an annealing period.

[0113] Referring to the following table, the following table is a schematic table of network training parameters:

[0114]

[0115] Referring to Figure 4 , Figure 4 is a schematic diagram of network training results.

[0116] Optionally, the construction method of the hybrid geographic space loss function comprises:

[0117] geographical error between the predicted position and the real position is calculated by Haversine distance;

[0118] motion parameter error other than the position parameter is calculated by mean absolute error;

[0119] geographical error and motion parameter error are fused by weighting to construct a hybrid loss function.

[0120] Specifically, Haversine distance optimizes geographical error. Haversine distance solves the problem of non-linear amplification of error in high latitude area by traditional plane Euclidean distance through accurate modeling of earth curvature.

[0121]

[0122] wherein, is the earth radius, , are the real latitude and longitude respectively, , are the predicted latitude and longitude respectively, and B is the batch size.

[0123] Mean absolute error (MAE) constrains the accuracy of other parameters. MAE balances the optimization weight of multiple parameters and improves the sensitivity of the model to key motion states.

[0124]

[0125] wherein, is the batch size, is the predicted value of the i-th dimension of the i-th sample, is the real value of the i-th dimension of the i-th sample.

[0126] Finally, the weighted hybrid loss is obtained by using geographical error and motion parameter error, and the geographical error and parameter accuracy are comprehensively optimized by multi-objective collaborative optimization, so as to solve the problem that a single loss function cannot consider global position accuracy and local motion state.

[0127]

[0128] wherein, is the comprehensive loss, is the geographical error, is the motion parameter error, is the mixing coefficient.

[0129] Optionally, the predicted trajectory is Kalman filtered to obtain a smoothed trajectory after smoothing correction, comprising: ​​​​

[0130] constructing a kinematic model of the aircraft, defining a state vector, a uniform motion matrix, a process noise covariance, an observation matrix and an observation noise covariance;

[0131] inputting the predicted trajectory into the kinematic model, determining a Kalman gain based on the observation matrix, a prior state covariance and the observation noise covariance;

[0132] determining a posterior state of the state vector based on the Kalman gain, an observation input, the observation matrix and a prior state estimation of the state vector, to obtain a smoothed trajectory after smoothing correction.

[0133] Specifically, the embodiment of the application performs trajectory smoothing processing through state space modeling and Kalman filtering.

[0134] modeling aircraft motion as a dynamic system in a state space, and optimizing the smoothing and rationality of the LSTM predicted trajectory through physical constraint.

[0135] define , construct a uniform motion model:

[0136]

[0137] reflecting model uncertainty, set as a diagonal matrix, process noise covariance .

[0138] observation update and smoothing. Take the LSTM prediction value as the observation input , fuse the prediction and the model through the Kalman gain, output the smoothed trajectory, and update the state:

[0139]

[0140] wherein, is the Kalman gain, H is the observation matrix, is the prior state covariance, is the observation noise covariance, R reflects the LSTM prediction error, set as a diagonal matrix, observation noise covariance , is the prior state estimation, is the observation input, is the posterior state of the state vector;

[0141] that is, based on the model to predict the current position and velocity;

[0142] that is, the covariance increases due to the accumulation of model errors.

[0143] Optionally, the denoising and feature normalization of the fusion features to obtain a normalized sequence comprises:

[0144] determining a time window size, determining a linear decay weight corresponding to the target moment based on the time window size;

[0145] determining a noise weighted average value of the target moment based on the linear decay weight, the fusion features and the time window size;

[0146] performing robust normalization of the feature data of latitude and longitude by quartile range, performing normalization of the state parameters by maximum and minimum value, and obtaining a normalized sequence by comprehensively normalizing all feature data;

[0147] Specifically, the embodiment of the application performs feature normalization on the fusion features. Due to the significant difference in physical meaning and data distribution of different features, a feature-independent normalization strategy is adopted. For geographical position (longitude, latitude), robust normalization by interquartile range (IQR) is adopted. IQR is not sensitive to outliers and is suitable for processing outliers (such as temporary jump of GPS signal) that may exist in geographical position data, and can also preserve the spatial relative relationship.

[0148]

[0149] For other parameters such as height, attitude angle, speed, etc., Min-Max normalization is adopted to the interval.

[0150]

[0151] Optionally, the feature data includes longitude, latitude, height, heading angle, pitch angle, roll angle, airspeed, horizontal speed and vertical speed.

[0152] Specifically, the embodiment of the application constructs a multi-dimensional feature vector, calculates the horizontal speed based on the latitude difference and calculates the vertical speed based on the height difference, and constructs a 9-dimensional input feature :

[0153]

[0154] wherein, and are longitude and latitude respectively, and the unit is degree; is height, and the unit is meter; , , are heading angle, pitch angle and roll angle respectively, and the unit is degree; is airspeed, and the unit is ; is horizontal speed, and the unit is , the calculation formula is as follows:

[0155]

[0156] wherein, is the radius of the earth, and is the longitude difference and the latitude difference, is the time interval.

[0157] is the vertical speed, and the unit is , the calculation formula is as follows:

[0158]

[0159] wherein, is the time interval.

[0160] By physical calculation of horizontal / vertical speed, the position change is converted into speed characteristics, and the model's representation ability for motion state is enhanced.

[0161] The technical effects and engineering application value of the present application are illustrated below in combination with specific test data and evaluation standards:

[0162] Effect 1: The trajectory prediction accuracy is significantly improved

[0163] By bidirectional LSTM spatiotemporal modeling and constructing Haversine distance and MAE mixed geographic space loss function, the trajectory data in the international public flight data set are tested and trained, including complex actions such as sharp turns, climbing / dive, etc., and the horizontal position error, heading angle error, and height error are used as evaluation indexes. The test results are shown in Table 1 as follows:

[0164] Table 1:

[0165]

[0166] Effect 2: Training efficiency and model convergence speed are optimized

[0167] The learning rate is adjusted periodically by using the Cosine function to improve the model training efficiency and convergence speed. The test results are shown in Table 2 as follows:

[0168] Table 2:

[0169]

[0170] Effect 3: Nonlinear accumulation of geographic space error is inhibited

[0171] The geographic spatial error is reduced by the Haversine geographic loss function and feature normalization. The high-latitude flight trajectory from 60° to 80° north latitude is predicted, and the cumulative error amplification rate (error growth per 100 kilometers) is taken as the evaluation index. The test results are shown in Table 3:

[0172] Table 3:

[0173]

[0174] The application will be described in detail below in combination with specific embodiments.

[0175] Application scenario: In the cruising phase (height ≥ 10,000 meters) of civil aviation flights, the future flight trajectory is predicted in real time based on ADS-B broadcast data, and conflict warning and airspace optimization support are provided for the air traffic control system.

[0176] Step 1: Data acquisition. Deploy SDR (software defined radio) receivers to capture latitude, longitude, height, ground speed, heading angle and flight identification information broadcast by flights in real time, with a data update frequency of 1 second;

[0177] Step 2: Data preprocessing. Remove sudden changes in latitude and longitude (such as instantaneous jumps exceeding the threshold) and invalid flight identification. For partially missing data such as vertical speed, use time series interpolation method to complete. Calculate derived features and fuse original observation values to construct 9-dimensional input features;

[0178] Step 3: Bidirectional LSTM time series modeling. Input 30 seconds of flight state data (latitude, longitude, height, ground speed, etc.), extract forward and reverse time series dependency features through bidirectional LSTM layer;

[0179] Step 4: Spatio-temporal attention mechanism. Dynamically allocate weights to different time steps to strengthen the influence of features such as heading mutations and acceleration phases;

[0180] Step 5: Training optimization strategy. Adopt a periodic learning rate scheduling strategy, quickly converge in the initial stage, fine-tune parameters in the later stage, and balance multiple objective constraints by multi-objective collaborative optimization, integrating geographic error and parameter precision;

[0181] Step 6: Fully connected output layer. Output the predicted trajectory points for the next 1 minute, including latitude, longitude, height and ground speed;

[0182] Step 7: Kalman filter smoothing. Establish a flight kinematics model, input the predicted trajectory as an observation value, and combine the uniform speed assumption for smoothing correction. Set the noise covariance parameters according to the sensor error characteristics to eliminate high-frequency jitter in the predicted trajectory;

[0183] Step 8: Conflict warning and air traffic control linkage. Based on the predicted trajectory, the horizontal and vertical minimum separation between multiple flights is calculated in real time. When the minimum separation is lower than the safety threshold, a hierarchical warning is sent to the air traffic control system;

[0184] This embodiment realizes high-precision prediction and real-time conflict management of civil aviation flight cruise trajectory through deep fusion of ADS-B data, bidirectional LSTM-attention collaborative modeling and dynamic post-processing optimization.

[0185] With reference to Figure 5 Based on the method in the above embodiment, an electronic device provided by the embodiment of the present application can include a processor (Processor) 510, a communication interface (Communications Interface) 520, a memory (Memory) 530, and a communication bus 540. The processor 510, the communication interface 520, and the memory 530 can communicate with each other through the communication bus 540. The processor 510 can invoke the logical instructions in the memory 530 to execute the method in the above embodiment.

[0186] In addition, the logical instructions in the memory 530 described above can be implemented in the form of a software function unit and sold or used as an independent product. When used, it can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application.

[0187] Based on the method in the above embodiment, a computer readable storage medium is provided in the embodiment of the present application. The computer readable storage medium stores a computer program. When the computer program runs on the processor, the processor executes the method in the above embodiment.

[0188] Based on the method in the above embodiment, a computer program product is provided in the embodiment of the present application. When the computer program product runs on the processor, the processor executes the method in the above embodiment.

[0189] It can be understood that the processor in the embodiments of the present application can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. The general-purpose processor can be a microprocessor or any conventional processor.

[0190] The method steps in the embodiments of the present application can be implemented in the form of hardware or by a processor executing software instructions. The software instructions can be composed of corresponding software modules, and the software modules can be stored in a random access memory (RAM), a flash memory, a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically EPROM (EEPROM), a register, a hard disk, a mobile hard disk, a CD-ROM or any other form of storage medium well known in the art. An exemplary storage medium is coupled to the processor, so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and the storage medium can be located in an ASIC.

[0191] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable apparatus. The computer instructions can be stored in or transmitted by a computer readable storage medium. The computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) manner. The computer readable storage medium can be any available medium accessible by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media sets. The available media can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0192] It can be understood that various numerical numbers involved in the embodiments of the present application are only distinguished for convenience of description, and are not used to limit the scope of the embodiments of the present application.

[0193] Those skilled in the art easily understand that the above only describes the preferred embodiments of the present application and is not used to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. An aircraft trajectory prediction method based on spatio-temporal attention and dynamic optimization bidirectional LSTM, characterized in that, The application relates to an aircraft trajectory prediction method based on a multi-dimensional feature data fusion and a Kalman filter. The method comprises the following steps: collecting multi-dimensional feature data of an aircraft, calculating horizontal speed and vertical speed of the aircraft based on feature data difference, determining fusion features based on the horizontal speed, the vertical speed and the feature data, carrying out noise reduction processing and feature normalization on the fusion features to obtain a normalized sequence, inputting the normalized sequence into an improved LSTM model to obtain a predicted trajectory, and carrying out Kalman filtering on the predicted trajectory to obtain a smoothed trajectory. The LSTM model is obtained through bidirectional time sequence modeling, dynamic weighting of a space-time attention mechanism, a hybrid geographical space loss function and dynamic learning rate optimization training. The method for constructing the hybrid geographical space loss function comprises the following steps: calculating geographical error between a predicted position and an actual position by using a Haversine distance; calculating motion parameter error except for position parameters by using a mean absolute error; and weighting and fusing the geographical error and the motion parameter error to construct a hybrid loss function. The method for obtaining the predicted trajectory comprises the following steps: extracting forward features of the normalized sequence by using a forward LSTM, extracting backward features of the normalized sequence by using a backward LSTM, and obtaining splicing features based on the forward features and the backward features; dynamically weighting the splicing features by using a space-time attention mechanism to obtain time step weights, weighting and fusing all the time step weights to obtain a global context representation; and obtaining a predicted trajectory covering all features by combining the global context and the latest time step features of the LSTM through a full connection layer. The process of dynamic learning rate optimization comprises the following steps: initializing a maximum learning rate and a minimum learning rate, and determining an annealing period of model iteration; and adjusting the learning rate according to the maximum learning rate, the minimum learning rate, a current iteration number and the annealing period. The process of dynamic weighting of the space-time attention mechanism comprises the following steps: obtaining time step weights by combining splicing features, attention weight matrices, bias items and learnable compression matrices through an activation function; and weighting and summing all the time step splicing features based on the time step weights to obtain a global context representation. The method for obtaining the smoothed trajectory by carrying out Kalman filtering on the predicted trajectory comprises the following steps: constructing a kinematic model of the aircraft, defining a state vector, a uniform motion matrix, a process noise covariance, an observation matrix and an observation noise covariance; inputting the predicted trajectory into the kinematic model, and determining a Kalman gain according to the observation matrix, a prior state covariance and an observation noise covariance; and determining a posterior state of the state vector based on the Kalman gain, an observation input, the observation matrix and a prior state estimation of the state vector to obtain a smoothed trajectory. The method for carrying out noise reduction processing and feature normalization on the fusion features to obtain the normalized sequence comprises the following steps: determining a time window size, determining a linear decay weight corresponding to a target moment based on the time window size; and determining a noise weighted average value of the target moment based on the linear decay weight, the fusion features and the time window size. ​ 2. The aircraft trajectory prediction method based on spatio-temporal attention and dynamic optimization bidirectional LSTM according to claim 1, characterized in that, ​ ​ ​ ​ 3. The aircraft trajectory prediction method based on spatio-temporal attention and dynamic optimization bidirectional LSTM according to claim 1, characterized in that, ​ ​ ​ 4. The aircraft trajectory prediction method based on spatio-temporal attention and dynamic optimization bidirectional LSTM of claim 1, characterized in that, ​ ​ ​ 5. The aircraft trajectory prediction method based on spatio-temporal attention and dynamic optimization bidirectional LSTM of claim 1, wherein, ​ ​ ​ ​ 6. The aircraft trajectory prediction method based on spatio-temporal attention and dynamic optimization bidirectional LSTM of claim 1, wherein, ​ ​ ​ The characteristic data of longitude and latitude are robustly normalized by quartile range, and the state parameters are normalized by maximum and minimum value, and the normalization sequence is obtained by comprehensively normalizing all characteristic data.

7. The aircraft trajectory prediction method based on spatio-temporal attention and dynamic optimization bidirectional LSTM of any one of claims 1-6, characterized in that, The characteristic data includes longitude, latitude, height, heading angle, pitch angle, roll angle and airspeed.

8. An electronic device, comprising: The computer program product comprises: at least one memory for storing a computer program; at least one processor for executing the program stored in the memory, and when the program stored in the memory is executed, the processor is configured to execute the method according to any one of claims 1-7.

9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. When the computer program runs on the processor, the processor is configured to execute the method according to any one of claims 1-7.

Images