Intelligent prediction method for aircraft taxiing trajectory by fusing space-time features and motion constraints

Abstract

In order to solve the key problems of the existing aircraft ground sliding trajectory prediction method, such as multi-source information fusion difficulty, space topology modeling deficiency and long-term prediction accuracy decay, the history trajectory encoder and the pavement path encoder parallel processing architecture are designed, the long short-term memory network is used to extract the trajectory time sequence characteristics, the graph attention network is used to model the space topology relationship of the control path, and the heterogeneous information is effectively integrated through the feature fusion layer; The multi-component loss function is proposed to fuse position, speed and acceleration, and the continuity and smoothness of the predicted trajectory are constrained; The lightweight data enhancement strategy and the adaptive residual connection mechanism are designed to improve the model generalization ability and relieve the long-term prediction error accumulation. The method can effectively fuse multi-source trajectory information, and significantly improve the accuracy and stability of the aircraft ground sliding trajectory prediction.

Description

Technical Field

[0001] This invention belongs to the field of aircraft trajectory prediction technology, and in particular relates to an intelligent prediction method for aircraft taxiing trajectories that integrates spatiotemporal characteristics and motion constraints. Background Technology

[0002] Aircraft taxiway prediction, as a core technology for modern intelligent airport management, is of great significance for improving airport operational efficiency and ensuring flight safety. Accurate trajectory prediction can provide airport controllers with critical decision support, enabling real-time conflict detection and resolution, optimizing taxiway planning, and enhancing airport situational awareness, thereby driving airport operations towards digitalization and intelligence.

[0003] Aircraft trajectory prediction falls under the category of time series forecasting problems, and research methods are mainly divided into two categories: model-based methods and neural network-based methods. Traditional model-based methods are mainly based on dynamic models and various Kalman filters. Although they have clear physical meanings, they are difficult to accurately represent uncertainties during aircraft taxiing in the complex environment of airports, such as changes in air traffic control instructions, weather effects, and equipment failures. This results in limited prediction accuracy and makes them unsuitable for the needs of efficient and safe operation of aircraft.

[0004] With the development of aircraft surveillance and artificial intelligence technologies, various neural networks have become the mainstream method for trajectory prediction. However, methods based on a single neural network can only process the temporal information of the trajectory and ignore the spatial topology of the airport pavement network, resulting in insufficient ability to model the differences in the movement patterns of aircraft in different taxiing areas.

[0005] The existing research on real-time prediction of aircraft ground taxiing trajectories still has the following problems: (1) Existing methods often ignore the taxiing path control instructions received by the aircraft during taxiing. The instructions often significantly change the aircraft's taxiing behavior, but the current prediction models have failed to effectively take them into account; (2) Traditional time series models mainly focus on the time dependence of the trajectory, while graph network methods focus on spatial relationship modeling and lack models that effectively integrate the spatial topology of the airport pavement network with the time series characteristics of the trajectory; (3) Current research mainly focuses on short-term prediction of 1 to 5 seconds, but in applications such as conflict avoidance and path optimization that require longer-term predictability, prediction capabilities of 8 seconds or even longer are needed. Existing methods have serious error accumulation problems in long-term prediction; (4) Aircraft trajectory data acquisition is costly and difficult to label, and training samples are relatively scarce. Existing methods often lack effective data augmentation strategies to improve the generalization ability and robustness of the model. Summary of the Invention

[0006] In view of this, the present invention aims to overcome the shortcomings of the above-mentioned problems in the prior art and proposes an intelligent prediction method for aircraft taxiing trajectory that integrates spatiotemporal characteristics and motion constraints, so as to achieve high-precision long-term prediction of aircraft ground taxiing trajectory and provide reliable technical support for airport intelligent management and conflict avoidance system.

[0007] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0008] A method for intelligent prediction of aircraft taxiing trajectories that integrates spatiotemporal features and motion constraints includes the following steps:

[0009] Step 1: Obtain raw aircraft taxiing data, including taxiing trajectory data, air traffic control instructions data, and airport taxiing route maps; construct a dataset and divide it into training, validation, and test sets.

[0010] Step 2: Perform enhanced feature extraction, data standardization, and data augmentation on the acquired data;

[0011] Step 3: Construct a trajectory prediction model, including a trajectory encoder, a path encoder, a feature fusion module, and a trajectory decoder;

[0012] Step 4: Design a multi-component loss function and optimization strategy, and use the training set and validation set to train and optimize the trajectory prediction model;

[0013] Step 5: Output the predicted trajectory and evaluate the model's prediction performance using the test set.

[0014] Furthermore, in step 1, the taxiing trajectory data includes the aircraft ID, position coordinates, and speed; the air traffic control instruction data includes the aircraft ID, taxiing path, and path information; and the airport taxiing route map includes the taxiing network structure, node and edge information, and spatial topology.

[0015] Furthermore, in step 2, the extracted enhancement features include taxiing speed components, acceleration, heading angle, and angular velocity; the data standardization process includes coordinate standardization and zero-mean unit variance; and the data augmentation process includes noise reduction and trajectory smoothing.

[0016] Furthermore, the trajectory encoder uses a two-layer LSTM to encode the aircraft's historical taxiing trajectory, extracts trajectory temporal features, and outputs a trajectory representation vector.

[0017] Furthermore, the path encoder uses a graph attention network to model the spatial topological relationships of the control path and outputs a path feature vector.

[0018] Furthermore, the feature fusion module includes feature alignment and splicing, and nonlinear interaction modeling. It splices the trajectory time-series features obtained from aircraft historical trajectory encoding with the path features obtained from control path encoding, and learns the interaction relationship between the two types of features through a two-layer fully connected network.

[0019] Furthermore, the trajectory decoder generates future trajectories using LSTM and introduces a residual connection mechanism with learnable weights.

[0020] Furthermore, the multi-component loss function includes position loss, velocity loss, and acceleration loss, which jointly constrain and optimize the position, velocity, and acceleration of the aircraft's taxiing trajectory.

[0021] Furthermore, the optimization strategy employs OneCycleLR learning rate scheduling, AdamW optimizer, and gradient pruning training optimization strategy. By dynamically adjusting the learning rate, controlling weight decay, and preventing gradient explosion, the stability and convergence speed of model training are improved.

[0022] Furthermore, the average displacement error and the final displacement error are used as core evaluation indicators. The model's predictive performance is evaluated by measuring the average deviation and the final position deviation between the predicted trajectory and the actual trajectory over the entire time series.

[0023] Compared with existing technologies, the intelligent prediction method for aircraft taxiing trajectories that integrates spatiotemporal features and motion constraints described in this invention has the following advantages:

[0024] This invention's model effectively integrates the temporal motion characteristics of aircraft taxiing with the spatial structure characteristics of airport pavements. Through taxiing enhancement feature extraction and standardization, combined with a distance threshold-based graph construction strategy, it significantly improves the model's ability to represent the dynamic characteristics of aircraft taxiing. It achieves an effective conversion from raw GPS data to high-quality feature representation, providing rich spatiotemporal information input for trajectory prediction. The designed multi-component loss function, optimized through joint constraints of position, velocity, and acceleration, ensures the kinematic consistency of the predicted trajectory. Combined with residual connectivity and optimized training strategies, the proposed method exhibits rapid convergence, providing reliable technical support for airport ground traffic control systems. Attached Figure Description

[0025] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0026] Figure 1 This is a flowchart illustrating the intelligent prediction method for aircraft taxiing trajectories that integrates spatiotemporal features and motion constraints according to the present invention.

[0027] Figure 2 This diagram illustrates the training loss and verification error results of the three methods of this invention. Detailed Implementation

[0028] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0029] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0030] like Figure 1 As shown, this invention provides an aircraft ground taxiing trajectory prediction method (GAT-LSTM) that fuses a Graph Attention Network (GAT) and a Long Short-Term Memory (LSTM) network. This method designs a parallel processing framework of a historical trajectory encoder and a pavement path encoder. It captures the temporal dynamic features of the trajectory using LSTM, models the spatial topological relationships of the control path using GAT, and achieves effective integration of heterogeneous information through a feature fusion layer. A multi-component loss function fusing position, velocity, and acceleration is proposed, and Huber loss is used to mitigate the impact of outliers while constraining the position accuracy, velocity continuity, and acceleration smoothness of the predicted trajectory. To address the problem of scarce training data, a lightweight trajectory noise injection and smoothing data augmentation strategy is designed. A residual connection mechanism with learnable weights is introduced into the decoder to adaptively fuse the linear extrapolation baseline with the neural network prediction.

[0031] Based on the characteristics of gliding data and considering the trajectory prediction task, this invention establishes a trajectory prediction method based on GAT-LSTM. The problem definition and mathematical modeling are as follows:

[0032] (1) The aircraft taxiing trajectory prediction problem is defined as:

[0033] Given aircraft taxiing instructions and airport pavement structure information, predict future taxiing trajectories.

[0034] Let the aircraft's historical taxiing trajectory be...

[0035]

[0036] in, Let T be the set of all real numbers. hist d represents the historical gliding time step; d represents the feature dimension; x represents the historical gliding time step. t =[pos x , pos y , v, a, θ, ω] T Let pos represent the state vector at time t, where pos xLet x be the horizontal coordinate of the aircraft, y be the vertical coordinate of the aircraft, v be the velocity, a be the acceleration, θ be the heading angle, and ω be the angular velocity.

[0037] The predicted target gliding trajectory is:

[0038]

[0039] Among them, T future To predict the gliding time step; y t =[pos x , pos y ] T This represents the aircraft's position coordinates at time t.

[0040] During the taxiing process of an aircraft at an airport, the structure information of the airport taxiway is represented using a graphical structure:

[0041] (1)

[0042] In the formula, This represents the airport pavement structure diagram; υ={v1, v 2, …, v N,} represents the set of nodes for the taxiway path, v i =[x i ,y i ] T ε represents the coordinates of the i-th gliding path point; ε represents the gliding path connection relationship between nodes.

[0043] (2) Quantitative representation of control orders

[0044] During ground taxiing, air traffic controllers issue taxiing clearances to aircraft via radio, specifying the taxiing path from the current position to the target position. This invention concretizes control instructions as a sequence of critical path points in a taxiway network. This sequence clearly identifies key locations such as taxiway intersections, turning points, and runway thresholds that the aircraft must traverse. Let the control instruction path be an ordered sequence of M critical nodes:

[0045] (2)

[0046] In the formula, M is the number of path points (usually 5-30 nodes); This represents the planar coordinates of the i-th path point. These path points correspond to key locations in the airport taxiway network, such as intersections, turning points, and runway entrances, constituting the spatial constraints for the future taxiing of aircraft.

[0047] Processing taxiing data includes: calculating enhanced features such as taxiing speed components, acceleration, heading angle, and angular velocity from the aircraft's original position and speed data, and fully expressing the kinematic information of the aircraft's taxiing process as a multi-dimensional feature vector. This provides rich input for deep learning models to describe the dynamic characteristics of aircraft taxiing, enabling the models to learn the motion laws and spatiotemporal dependencies of aircraft taxiing, thereby improving the accuracy and physical rationality of trajectory prediction.

[0048] (1) Extraction of aircraft taxiing enhancement features

[0049] Based on the acquired raw aircraft taxiing data, the following enhanced features are calculated to facilitate subsequent model training:

[0050] Glide speed component:

[0051] (3)

[0052] In the formula, This represents the gliding velocity component in the x-direction; This represents the gliding velocity component in the y-direction.

[0053] Glide acceleration:

[0054] (4)

[0055] In the formula, Indicates the acceleration at time t; This represents the gliding speed at time t.

[0056] Aircraft heading angle calculation

[0057] (5)

[0058] In the formula, θ t This represents the heading angle at time t.

[0059] Aircraft angular velocity calculation:

[0060] (6)

[0061] In the formula, ω t This represents the angular velocity at time t.

[0062] (2) Standardization of gliding data

[0063] For subsequent model training and processing, the aircraft position coordinates are normalized using Z-score:

[0064] (7)

[0065] In the formula, (μ x,σ x ) and (μ y ,σ y ) are the mean and standard deviation of the x and y coordinates of the aircraft in the training set, respectively.

[0066] To effectively integrate the temporal motion characteristics of aircraft taxiing with the spatial structural characteristics of airport pavement, this invention designs a deep neural network architecture that integrates trajectory encoders, path encoders, and multi-component constraint loss. Through deep fusion of spatiotemporal features and explicit modeling of motion constraints, it achieves high-precision intelligent prediction of aircraft taxiing trajectories.

[0067] The complete process of transforming control path instructions from their raw form into features usable by the model and ultimately integrating them into trajectory prediction can be summarized into the following seven steps:

[0068] The first step is to obtain the taxiing permission instructions issued by the controller from the airport surface monitoring system and record the taxiing and path that the aircraft should follow in the form of text or structured data.

[0069] The second step is to parse the control instruction text into a specific taxiway node sequence.

[0070] The third step is to map the abstract sequence of taxiway nodes into a sequence of coordinate points in a Cartesian coordinate system, with each node corresponding to a two-dimensional coordinate point, thus forming a numerical representation of the path.

[0071] The fourth step is to construct a graph structure based on the sequence of coordinate points. Each path point is treated as a node in the graph, and connecting edges are established between nodes according to rules such as sequence order and spatial distance, forming a graph network containing spatial topological relationships.

[0072] The fifth step involves inputting the constructed graph structure into a graph attention network for encoding. Through multi-layer graph convolution and attention mechanisms, the spatial structural features of the path are extracted, including information such as the path's geometry, turning radius, and key point locations, ultimately yielding a feature vector representation of the path.

[0073] The sixth step involves deep fusing the path feature vector with the temporal feature vector of the aircraft's historical trajectory. First, the two feature vectors are concatenated. Then, a multi-layer neural network is used to learn the interaction between them, achieving an adaptive balance between historical motion trends and path space constraints, and outputting the fused comprehensive features.

[0074] The seventh step involves inputting the fused features into the decoder, which then uses a recurrent neural network to progressively generate trajectory coordinate predictions for the next eight time steps. During the decoding process, the path constraint information contained in the fused features guides the predicted trajectory toward the controlled path.

[0075] Through the complete processing chain of the above seven steps, the control path instructions, originally in text or symbolic form, undergo a series of transformations including parsing, quantization, graph modeling, feature encoding, and deep fusion, and are ultimately effectively embedded into the trajectory prediction model as spatial constraint information. This mechanism achieves an organic combination of control intentions and the actual motion patterns of aircraft, ensuring that the prediction results conform to both physical motion characteristics and the path specification requirements of air traffic control.

[0076] (1) Aircraft taxiing trajectory encoder

[0077] Using a two-layer LSTM to encode the aircraft's historical taxiing trajectory:

[0078] (8)

[0079] In the formula, Represents the 6-dimensional state characteristics of an aircraft; This indicates the previously hidden state; The previous state is represented by W; the corresponding weight matrix is ​​represented by b; the corresponding bias vector is represented by f. t Represents historical information that controls forgetting; i t Indicates control over the receipt of new information; o t Information indicating control output; Indicates the current candidate information; C t Indicates long-term memory state; h t σ represents the short-term output state, and σ refers to the activation function, which is usually the sigmoid function.

[0080] Aircraft taxiing characteristics representation:

[0081] (9)

[0082] In the formula, This represents the normalized characteristics of an aircraft's taxiing trajectory. This represents the hidden state output of the LSTM encoder at the last historical moment; T hist The historical gliding time step is represented by a value of 20, indicating the length of the observed historical trajectory; LayerNorm(⋅) is the layer normalization function, calculated as follows:

[0083] (10)

[0084] In the formula, μ represents the mean of the input; σ 2 Indicates the variance of the input; Represents the numerical stability constant; This represents the learnable scaling and offset parameters. This refers to convolution operations.

[0085] (2) Airport pavement path encoder

[0086] The control command paths are represented by a graph structure, where their spatial topology is determined. Specifically, the sequence of path points R is transformed into a graph G=(V,E), where the node set V contains all path points, and the edge set E is constructed according to the following rules:

[0087] 1) Adjacent connection: Connecting adjacent points in a path sequence ;

[0088] 2) Skip connection: Connects points that are separated by one node. Capture mid-range dependencies;

[0089] 3) Distance-based connection: Connecting points with an Euclidean distance less than a threshold. The multi-scale graph construction strategy of modeling spatial proximity relationships between point pairs can simultaneously capture the sequential constraints of paths, local geometric features, and global spatial layout, enabling graph attention networks to learn the spatial guidance information contained in control instructions.

[0090] Considering the establishment of GTA, the formula for the graph attention mechanism is as follows:

[0091] (11)

[0092] In the formula, W∈ Represents the weight matrix; a∈ N represents the attention vector; i Let || denote the set of neighbors of taxiway node i, and || denote the vector concatenation operation. ij This represents the edge formed by nodes i and j in the graph structure. The unnormalized raw attention score quantifies the importance of node j to node i. After normalization, we get a. ij Represents the edge formed by nodes i and j The final attention coefficient, e ik The meaning is the same as e ij Same, but different edge choices This represents the feature vector of node i.

[0093] Based on the construction of multi-head attention, the final feature representation of node i after multi-head attention processing is as follows:

[0094] (12)

[0095] In the formula, K=4 is the number of attention heads, indicating that 4 parallel attention mechanisms are used; i represents the i-th path point; ∈ The feature representation of node i generated by the k-th attention head is denoted by ; k∈{1,2,3,4} represents the attention head index. ∈ This represents the final feature representation of node i after multi-head attention processing.

[0096] Global pooling operations:

[0097] (13)

[0098] In the formula, H∈ The feature matrix representing all taxiway nodes,

[0099] N is the number of nodes; This means averaging the features of all nodes to obtain the global mean feature. This means finding the maximum value of the features across all nodes to obtain the global maximum feature. A global feature representation of the airport pavement.

[0100] (2) Feature fusion module

[0101] The core task of the feature fusion module is to effectively integrate the temporal features of historical trajectories with the spatial features of control paths. The fusion process consists of two stages:

[0102] 1) First stage: Feature alignment and stitching

[0103] Time-series features obtained by encoding aircraft historical trajectories Spatial features obtained from control path coding To splice:

[0104] (14)

[0105] In the formula, Represents the feature vector of an aircraft's taxiing trajectory; Represents the airport pavement path feature vector; This represents the vector concatenation operator, which joins along the feature dimension; This represents the combined feature vector after concatenation.

[0106] 2) Second stage: Nonlinear interaction modeling

[0107] The interaction relationship between two types of features is learned through a two-layer fully connected network:

[0108] (15)

[0109] In the formula, This represents the first-level weight matrix, which maps 256 dimensions to 128 dimensions; This represents the first-level bias vector; This represents the weight matrix of the second layer; This represents the second-layer bias vector; This indicates the characteristics after fusion.

[0110] The first layer maps 256-dimensional combined features to a 128-dimensional latent space, learning higher-order interaction patterns between trajectories and paths; the second layer further refines and integrates these features. This design enables the model to adaptively learn the trade-off between historical motion trends and regulatory path constraints, ensuring that trajectory predictions conform to the laws of physical motion while also ensuring that the prediction results follow the spatial guidance of regulatory instructions.

[0111] (3) Aircraft taxiing trajectory decoder

[0112] The decoder input is a repetition of the fused features:

[0113] (16)

[0114] In the formula, repeat(⋅,n) represents the operation of repeating the vector n times; This represents the decoder input sequence matrix.

[0115] The input to the decoder is the fused features. This feature integrates historical movement trends and regulatory constraints. The decoder uses LSTM to unfold over eight time steps, progressively generating the future trajectory:

[0116] (17)

[0117] In the formula, This represents the hidden state of the encoder at time 20, which serves as the initial hidden state of the decoder. This represents the cell state of the encoder at time 20, which serves as the initial cell state of the decoder. This represents the feature sequence output by the LSTM decoder; This indicates the final hidden state of the decoder; This indicates the final cell state of the decoder.

[0118] The control command path can affect the decoding process due to fusion characteristics. The spatial constraints governing the path are already embedded within the code, and the decoder implicitly considers the guiding role of the path when generating the position for each future time step. Specifically, this manifests as follows:

[0119] 1) Spatial attraction: Predicted trajectory points will gradually move closer to the controlled path;

[0120] 2) Direction correction: When the historical direction of motion is inconsistent with the path direction, the decoder will adjust the heading;

[0121] 3) Speed ​​adjustment: When approaching a turning point, the decoder tends to predict a lower speed.

[0122] The above constraints avoid the rigidity of hard rules, enabling the model to maintain flexible modeling of actual motion laws while complying with regulatory instructions.

[0123] Output projection operation:

[0124] (18)

[0125] In the formula, This represents the first-level projection weight matrix; This represents the first-layer projection bias vector; Indicates intermediate projection output; This represents the second-layer projection weight matrix, and outputs 2D coordinates. This represents the second-layer projection bias vector; Predict the final location sequence.

[0126] (4) Residual connection

[0127] Residual connections combine the complex pattern recognition capabilities of deep learning with the fundamental laws of physical motion. By automatically balancing the contributions of both through a learnable parameter λ, they improve the accuracy of predictions and the physical plausibility of the data.

[0128] Linear extrapolation baseline calculation formula:

[0129] (19)

[0130] In the formula, λ is a learnable parameter, initialized to 0.15. This indicates the aircraft's position coordinates at time 20; This indicates the aircraft's position coordinates at time 19; This represents the aircraft's taxiing velocity vector at the last moment; Indicates a future time step index; This represents the linear extrapolation position at time t; Represents a complete linear extrapolation trajectory sequence; This represents the predicted trajectory learned by the neural network. This indicates the final predicted aircraft taxiing trajectory.

[0131] To ensure that the predicted aircraft taxiing trajectory is not only accurate in position, but also conforms to the kinematic laws of aircraft taxiing in terms of velocity and acceleration changes, this invention designs a multi-component loss function that includes position loss, velocity loss, and acceleration loss. By jointly constraining and optimizing the position, velocity, and acceleration of the aircraft taxiing trajectory, high-precision trajectory prediction under kinematic consistency constraints is achieved.

[0132] (1) The formula for calculating position loss is as follows.

[0133] (20)

[0134] In the formula, N is the number of aircraft samples in the batch; T future =8 indicates the number of future time steps in the prediction; This represents the predicted position of the i-th aircraft at time t; This represents the actual position of the i-th aircraft at time t; Let L2 be the square of the L2 norm, and calculate the Euclidean distance.

[0135] (2) The formula for calculating speed loss is as follows.

[0136] (twenty one)

[0137] In the formula, This represents the predicted velocity vector calculated through position difference; This represents the velocity vector calculated using the difference between the actual positions. This indicates Huber's losses.

[0138] (3) The formula for calculating acceleration loss is as follows.

[0139] (twenty two)

[0140] In the formula, This represents the predicted acceleration vector calculated through velocity difference; This represents the acceleration vector calculated using the actual velocity difference.

[0141] (4) The formula for calculating the total loss is as follows.

[0142] (twenty three)

[0143] The weight is set as: ω pos =1.0, position loss weight (highest priority); ω vel =0.3, velocity loss weight (medium importance); ω acc =0.1, acceleration loss weight (low weight, to avoid over-constraint).

[0144] To ensure that the aircraft taxiing trajectory prediction model can converge to the optimal solution efficiently and stably, advanced training optimization strategies such as OneCycleLR learning rate scheduling, AdamW optimizer, and gradient pruning are adopted. By dynamically adjusting the learning rate, controlling weight decay, and preventing gradient explosion, the stability and convergence speed of model training are improved, thereby achieving better prediction performance.

[0145] (1) OneCycleLR learning rate scheduling

[0146] The learning rate changes as follows:

[0147] (twenty four)

[0148] In the formula, the minimum learning rate lr min =0.001, maximum learning rate lr max =0.003; t is the current training step number; T total Total training steps;

[0149] (2) The gradient clipping calculation formula is as follows.

[0150] (25)

[0151] In the formula, τ=1.0 is the clipping threshold; A ij This represents the elements of the adjacency matrix, where 1 indicates a connection and 0 indicates no connection. This represents the original gradient vector, where P is the total number of all parameters in the model. This represents the scaling factor, which is used to scale the gradient norm when it exceeds a threshold.

[0152] To comprehensively quantify the performance of aircraft taxiing trajectory prediction methods, the average displacement error (ADE) and final displacement error (FDE) are used as core evaluation indicators. By measuring the average deviation and final position deviation between the predicted trajectory and the actual trajectory over the entire time series, the accuracy level and practical value of the model in aircraft taxiing trajectory prediction tasks are objectively evaluated.

[0153] (1) Formula for calculating average displacement error (ADE):

[0154] (26)

[0155] (2) Formula for calculating final displacement error (FDE):

[0156] (27)

[0157] To effectively represent the complex spatial structure of airport pavements and the connections between taxiways, this invention employs a graph construction strategy based on adjacency connections, skip connections, and distance connections. By constructing a graph structure containing taxiway path points and their adjacency relationships, it provides a suitable input representation for the graph attention network, enabling the model to learn the spatial constraint information of the airport pavement and integrate it into the aircraft taxiway trajectory prediction process.

[0158] The formula for constructing the adjacency matrix is ​​as follows.

[0159] (28)

[0160] In the formula, θ1 is the distance threshold, which is taken as the 30th percentile of all distances; A ij This represents the elements of the adjacency matrix, where 1 indicates a connection and 0 indicates no connection. Represents the coordinates of the i-th and j-th taxiway nodes.

[0161] The final adjacency matrix is ​​calculated using the following formula.

[0162] (29).

[0163] The following specific examples illustrate the solution of the present invention.

[0164] To evaluate the practical application effectiveness of the proposed method, real-time monitoring data from a hub airport in North China was used. The data collection spanned two periods: the first period was from late November to early December 2023 (12 days), and the second period began in mid-December 2023 and lasted for 10 days. The raw data acquired through the multi-point positioning monitoring system underwent rigorous screening, retaining only the trajectory data from the aircraft's ground operation phase.

[0165] The raw monitoring data contains the following core elements: time sequence identifier, aircraft unique code (TRACK_ID), GPS coordinate information (latitude and longitude), and instantaneous speed of motion (SPEED). In the data preprocessing stage, the WGS84 geographic coordinates acquired by GPS are first converted to the CGCS2000 Cartesian coordinate system suitable for calculation. Then, a continuous motion sequence is constructed based on the aircraft code and time sequence information. To eliminate the influence of absolute position on the model, the geometric center of the airport is selected as the reference point, and data normalization is achieved through coordinate translation. Finally, the vector representation transformation is completed according to the trajectory serialization method.

[0166] (2) Control order data

[0167] The air traffic control instruction data originates from actual taxiing permission recorded by the airport surface monitoring system. Each taxiing instruction includes the aircraft ID and the corresponding sequence of planned path points. The path points are indexed by taxiway node numbers in the airport's electronic map and then converted into planar coordinates in the CGCS2000 coordinate system. The path lengths in the dataset range from 5 to 30 nodes, averaging approximately 15 nodes, covering typical taxiing scenarios such as from parking stand to runway and from runway to parking stand.

[0168] Matching with historical trajectory data is done via aircraft ID and timestamp. For each training sample, the aircraft's actual trajectory within the historical observation window (first 20 seconds) and the corresponding valid air traffic control instructions for that period are extracted to form paired input data.

[0169] (3) Airport pavement network structure

[0170] Taxiway network modeling requires consideration of the geometric layout and topological relationships of airport infrastructure. The test airport used in this invention has a typical multi-runway configuration and a complex taxiway network structure. Through digital processing, the physical taxiways are converted into a graph data structure containing nodes and connections. By combining the path allocation instructions from the air traffic control system and the historical movement patterns of aircraft, the sequence of taxiway nodes that aircraft may pass through in the future can be accurately identified.

[0171] (4) Sliding window segmentation strategy

[0172] Training samples were constructed using a fixed window segmentation method. For each complete aircraft taxiing trajectory, the historical observation window length was set to 20 time steps, and the future prediction window length was set to 8 time steps. Since the monitoring system's sampling frequency is 1Hz, the historical window corresponds to 20 seconds of observation data, and the prediction window corresponds to the trajectory for the next 8 seconds.

[0173] The specific splitting process is as follows:

[0174] 1) Read the complete trajectory sequence from the file and filter out the valid trajectories with a length of no less than 28 time steps (20+8);

[0175] 2) Extract the first 20 time steps from each trajectory as a historical observation sequence;

[0176] 3) Extract time steps 21 to 28 as the future truth value sequence;

[0177] 4) Simultaneously associate the control path instruction data corresponding to the aircraft.

[0178] The fixed window segmentation method described above ensures that each training sample contains complete observation-prediction sequence pairs, providing a consistent time span for model learning. For trajectories longer than 28 steps, this invention employs a starting point segmentation strategy, using only the first 28 time steps of the trajectory to maintain temporal consistency of the dataset.

[0179] (5) Dataset partitioning scheme

[0180] The collected raw trajectory data, after filtering and segmentation, was divided into training, validation, and test sets in a 70%:15%:15% ratio. The specific segmentation strategy is as follows:

[0181] 1) Training set: Contains 1050 samples (70% of the total), used for model parameter learning and weight optimization. To enhance the model's generalization ability, the training set samples underwent data augmentation, including trajectory noise injection (Gaussian noise level 0.005) and trajectory smoothing (sliding window size 3), with the augmentation probability set to 30%.

[0182] 2) Validation set: Contains 225 samples (15% of the total), used for hyperparameter tuning and early stopping detection during training. The validation set does not undergo data augmentation and maintains the original data distribution to truly reflect the model's generalization performance on unseen data.

[0183] 3) Test set: Contains 225 samples (15% of the total), used for final performance evaluation and method comparison. The test set is also not augmented to ensure the objectivity of the evaluation results.

[0184] The dataset was partitioned by randomly shuffling the data and then proportionally dividing it to ensure a balanced distribution of samples in each subset. There was no overlap between the training, validation, and test sets, guaranteeing the independence of the evaluation.

[0185] To comprehensively evaluate the performance of the method proposed in this invention in trajectory prediction tasks, STGAT and WIMP were selected as comparative baseline methods, which are highly representative and relevant. STGAT, as a typical representative of spatiotemporal graph attention networks, can effectively model spatiotemporal dependencies in trajectory prediction. Its graph attention mechanism is similar in technical approach to the GAT encoder used in this invention, facilitating the verification of the differences in performance between different graph attention network architectures when handling taxiway topology. WIMP, as an advanced method in the field of multi-agent motion prediction, is specifically designed for multi-objective trajectory prediction problems in traffic scenarios. Its multimodal prediction capabilities and modeling methods for environmental constraints are highly compatible with aircraft taxiing prediction tasks, enabling the verification of the advantages of the method proposed in this invention in complex airport environments from different technical perspectives. These two methods represent the mainstream technical routes of graph attention networks and multi-agent prediction, respectively. Comparative analysis fully demonstrates the effectiveness and innovation of the trajectory encoding-path fusion architecture proposed in this invention in aircraft taxiing trajectory prediction tasks.

[0186] To ensure the fairness and reliability of the comparative experiments, all three methods used the same training configuration parameters. The training epochs were set to 25, and the batch size was 32 samples. The AdamW optimizer was used for all methods, with an initial learning rate of 0.001 and a weight decay coefficient of 1×10⁻⁻⁻⁶. 4The momentum parameters β1=0.9 and β2=0.999. The learning rate scheduling strategy uses OneCycleLR with a maximum learning rate of 0.003, and the warm-up phase accounts for 10% of the total training time. To prevent gradient explosion, a gradient clipping threshold of 1.0 is set. An early stopping mechanism is employed, terminating training when the validation set ADE metric fails to improve for six consecutive epochs to prevent overfitting. The hidden layer dimension is uniformly set to 128 for all methods to maintain comparability in model complexity. The dropout rate is set to 0.1 for regularization to prevent overfitting.

[0187] (2) Setting the prediction time window

[0188] Based on the actual needs of airport taxiing scenarios, this invention sets the historical observation window to 20 time steps, predicting the trajectory coordinates for the next 8 time steps. With a data sampling rate of 1Hz, each time step corresponds to 1 second of real time. Therefore, the model predicts the aircraft's position change within the next 8 seconds based on the historical trajectory information of the previous 20 seconds. This time window setting provides sufficient historical information for motion pattern learning while also meeting the practical application needs of short-term trajectory prediction in airport ground traffic control.

[0189] Using the collected gliding data, trajectory prediction analysis is conducted using the proposed method and comparative methods. For example... Figure 2 As shown, firstly, the loss values ​​of the proposed method and the comparison model during the training process are compared ( Figure 2 a), and ADE in the verification process ( Figure 2 b) and FDE ( Figure 2 c).

[0190] Depend on Figure 2 As can be seen, the method of this invention exhibits excellent performance in trajectory prediction tasks. Regarding training loss, the method of this invention demonstrates the fastest convergence speed and the most stable training process, rapidly decreasing from an initial 0.346 to below 0.006. In contrast, while WIMP and STGAT methods can also converge effectively, their convergence speeds are relatively slower. In the evaluation metrics during the validation phase, the method of this invention achieved optimal results on both the ADE and FDE, two key metrics. Specifically, the ADE of the method of this invention decreased from an initial 0.270 to below 0.029, and the FDE decreased from 0.250 to around 0.032, significantly outperforming the final performance of the WIMP (ADE: 0.043, FDE: 0.042) and STGAT (ADE: 0.040, FDE: 0.044) methods. This result demonstrates that by combining graph attention networks with long short-term memory networks, the proposed method can better capture the spatiotemporal dependencies in trajectory data, thereby significantly improving the accuracy of gliding trajectory prediction. This verifies the effectiveness and superiority of the proposed method in trajectory prediction tasks in complex scenarios.

[0191] Furthermore, the performance results of the above methods are shown in Table 1.

[0192] Table 1

[0193]

[0194] As shown in Table 1, the method of this invention exhibits significant performance advantages in aircraft taxiing trajectory prediction tasks. Its average displacement error on the test set is only 0.0254 meters, representing improvements of 30.2% and 40.8% compared to STGAT and WIMP, respectively. Simultaneously, it maintains the best prediction accuracy of 0.0293 meters in terms of final displacement error. Although the model complexity of the three methods is similar (the number of parameters is between 360,000 and 590,000), the method of this invention achieves more accurate trajectory prediction results through a multi-component loss function, data augmentation strategies, and an effective fusion design of GAT and LSTM. Its median indices reach an ADE of 0.0163 meters and an FDE of 0.0218 meters, indicating that the method can achieve centimeter-level high-precision prediction on most prediction samples. It is worth noting that the current test set size of the method of this invention is relatively small. To further verify the generalization ability and statistical reliability of the method, subsequent validation on a larger-scale test dataset is needed to ensure the consistency and stability of the performance advantages. From the perspective of model complexity, the parameter sizes of the three methods are comparable. This method contains 593,347 parameters, which is roughly the same as STGAT (593,346 parameters) and slightly higher than WIMP (363,123 parameters). Despite the similar number of parameters, this method achieves higher parameter utilization efficiency through optimized design of multi-component loss functions, GAT-LSTM fusion architecture, and data augmentation strategies, resulting in a significant performance improvement at similar computational costs.

[0195] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent prediction method for aircraft taxiing trajectories that integrates spatiotemporal features and motion constraints, characterized by: Includes the following steps: Step 1: Obtain raw aircraft taxiing data, including Taxi trajectory data: including at least aircraft ID, timestamp, position coordinates, speed, acceleration, heading angle, and angular velocity; Control instructions data: including aircraft ID, taxiway instructions, and route information; Airport taxiway map: including taxiway network structure, node and edge information, and spatial topology; Construct the dataset and divide it into training, validation, and test sets; Step 2: Perform enhanced feature extraction, data standardization, and data augmentation on the training set data, and perform enhanced feature extraction and data standardization on the validation and test set data; The extracted enhancement features include taxiing speed components, acceleration, heading angle and angular velocity; the data standardization process includes coordinate standardization and zero mean unit variance; and the data augmentation process includes noise reduction and trajectory smoothing. Step 3: Construct a trajectory prediction model, including a trajectory encoder, a path encoder, a feature fusion module, and a trajectory decoder; The trajectory encoder uses a two-layer long short-term memory network to encode the aircraft's historical taxiing trajectory, extracts the trajectory temporal features, and outputs a trajectory representation vector. Path Encoder: A graph attention network is used to model the spatial topological relationships of the controlled path. A graph structure is constructed based on the coordinate point sequence, and each path point is treated as a node in the graph. Connection edges are established between nodes according to the sequence order and spatial distance rules, including adjacency connections, skip connections, and distance connections, forming a graph network containing spatial topological relationships. Adjacency connections are used to connect adjacent points in the path sequence, skip connections are used to connect points separated by one node to capture mid-range dependencies, and distance connections are used to connect point pairs with a Euclidean distance of less than a threshold to model spatial proximity relationships. The constructed graph structure is input into the graph attention network for encoding. Through multi-layer graph convolution and attention mechanisms, the spatial structural features of the path are extracted, and the path feature vector is output. Feature fusion module: The trajectory representation vector and the path feature vector are deeply fused by learning the interaction between them through a multi-layer neural network, and a comprehensive feature vector is output. Trajectory Decoder: Reshapes the comprehensive feature vector into a sequence form, inputs it into the LSTM decoder, and generates trajectory coordinate prediction values; Step 4: Design a multi-component loss function and optimization strategy. The loss function includes: Position loss: Calculates the positional deviation between the predicted trajectory and the actual trajectory; Velocity loss: Calculates the velocity deviation between the predicted trajectory and the actual trajectory; Acceleration loss: Calculates the acceleration deviation between the predicted trajectory and the actual trajectory; The optimization strategy employs the AdamW optimizer, combined with OneCycleLR learning rate scheduling and gradient pruning. The trajectory prediction model is trained and optimized using the training and validation sets; Step 5: Input the historical taxiing data of the aircraft to be predicted into the trained model, output the future taxiing trajectory, and evaluate the prediction performance using the average displacement error (ADE) and final displacement error (FDE).

2. The intelligent prediction method for aircraft taxiing trajectories integrating spatiotemporal features and motion constraints according to claim 1, characterized in that: The feature fusion module includes feature alignment and splicing, and nonlinear interaction modeling. It splices the trajectory representation vector obtained by encoding the aircraft's historical trajectory with the path feature vector obtained by encoding the control path, and learns the interaction relationship between the two types of features through a two-layer fully connected network.

3. The intelligent prediction method for aircraft taxiing trajectories integrating spatiotemporal features and motion constraints according to claim 1, characterized in that: The trajectory decoder generates future trajectories through a long short-term memory network and introduces a residual connection mechanism with learnable weights.

4. The intelligent prediction method for aircraft taxiing trajectory integrating spatiotemporal features and motion constraints according to claim 1, characterized in that: The optimization strategy employs OneCycleLR learning rate scheduling, AdamW optimizer, and gradient pruning training optimization strategy. By dynamically adjusting the learning rate, controlling weight decay, and preventing gradient explosion, the stability and convergence speed of model training are improved.

Images