Data-driven airport scene sliding path intelligent generation method

Abstract

The invention relates to the technical field of airport operation management, in particular to a data-driven airport scene sliding path intelligent generation method, which comprises the following steps: acquiring a plurality of taxiway sequences to form a taxiway sequence set; establishing a neural network model and a target function, and training the neural network model to obtain a taxiway embedded vector generator; starting and ending point pair coding, environment feature coding and related flight coding are carried out based on the taxiway embedded vector generator, and heterogeneous fusion features are obtained; carrying out autoregression generation based on Transform to obtain a plurality of candidate taxiways and corresponding score vectors; the multiple candidate taxiways are expanded and screened according to the score vectors of all the candidate taxiways, multiple candidate paths are generated, and the high-quality airport scene sliding path can be generated.

Description

Technical Field

[0001] This invention relates to the field of airport operation management technology, specifically to a data-driven intelligent generation method for airport surface taxiway paths. Background Technology

[0002] With the rapid development of the global air transport industry, airport surface traffic flow continues to grow, and airport taxiing systems have become a bottleneck restricting the overall capacity and operational efficiency of airports. Statistics show that airport surface taxiing delays account for more than 30% of total delays, and the economic losses caused by low aircraft taxiing efficiency remain high annually. Therefore, developing efficient and intelligent taxiing path planning technology is of great significance for improving airport operational efficiency and alleviating airport surface traffic congestion.

[0003] At large hub airports, taxiway planning faces complex challenges, including real-time traffic condition changes, weather conditions, runway configuration adjustments, and various safety separation requirements. Existing taxiway planning technologies suffer from the following problems: Traditional rule-based methods rely on manually defined fixed rule sets, which lack adaptability to complex and ever-changing operating environments, often resulting in overly conservative or inefficient paths. Mathematical optimization algorithms seek theoretically optimal solutions by constructing optimization models, but often neglect actual airport operational constraints and controller path selection preferences, leading to poor feasibility in practical applications. More importantly, experienced controllers' path selection preferences, developed through long-term practice, contain a wealth of valuable tacit knowledge, such as path selection tendencies under different operating conditions, accurate predictions of traffic flow trends, and real-time path adjustment decisions based on safety considerations. These controller decision preferences exhibit nonlinear characteristics, multi-objective trade-offs, and strong context dependence, making them difficult to accurately model and effectively utilize using traditional rule-based or mathematical optimization methods.

[0004] In recent years, deep learning technology has made groundbreaking progress in fields such as sequence modeling, pattern recognition, and knowledge mining. In particular, sequence generation models based on the Transformer architecture have demonstrated powerful pattern learning and generation capabilities in fields such as natural language processing, providing a novel technical approach to solving complex path planning problems. However, existing research on the application of deep learning technology in airport taxiway path planning remains very limited. It lacks specialized designs for the specific characteristics of taxiway path planning tasks, and even more so, it lacks systematic technical solutions that can effectively integrate multi-source heterogeneous information and learn controllers' path selection decision preferences. Summary of the Invention

[0005] In view of the above problems, the present invention provides a data-driven intelligent generation method for airport surface taxiways, which solves the technical problem that the prior art cannot generate high-quality taxiways.

[0006] This invention provides a data-driven intelligent generation method for airport surface taxiways, comprising the following steps: Step S1: Based on the collected aircraft surface taxiway tracks and airport taxiway network topology data, obtain multiple taxiway sequences to form a taxiway sequence set; the taxiway sequence includes multiple taxiway identifiers to mark different taxiways; Step S2: Establish a neural network model and an objective function. Train the neural network model based on the set of taxiway sequences and the objective function. Use the trained neural network model as a taxiway embedding vector generator. Step S3: Based on the origin-endpoint pair of the path generation task, the airport operating environment, and the taxiing time of the current flight, the origin-endpoint pair encoding, environmental feature encoding, and related flight encoding are performed respectively based on the taxiway embedding vector generator to obtain heterogeneous fusion features; Step S4: Based on the starting point of the path generation task and the heterogeneous fusion features, perform autoregressive generation based on Transformer. During the autoregressive generation process, the heterogeneous fusion features are processed by a dual-source parallel cross-attention mechanism to obtain multiple candidate gliding paths and their corresponding score vectors. Step S5: Update the score vector of each candidate taxiway based on the airport taxiway network topology data; expand and filter multiple candidate taxiways based on the updated score vector to generate multiple candidate paths, wherein the candidate path is a sequence of multiple candidate taxiways.

[0007] Preferably, step S1 specifically includes: Step S1-1: Match the aircraft surface taxiway track with the road segment edges in the airport taxiway network topology data to obtain the road segment edge identifiers for each track point. Step S1-2: Merge consecutive track points with the same road segment connection identifier to obtain a taxiway sequence; finally, obtain multiple taxiway sequences to form a taxiway sequence set.

[0008] Preferably, step S2 specifically includes: Step S2-1: Preprocess the set of taxiway sequences to obtain training samples; The preprocessing steps specifically include: constructing a taxiway vocabulary from all taxiway identifiers appearing in the taxiway sequence, and replacing taxiway identifiers in the taxiway vocabulary that appear less than a preset threshold with special markers. ; For each taxiway identifier in the taxiway vocabulary, training samples are built. The training samples include features and labels. The one-hot encoded vector of the taxiway identifier is used as the feature, and the one-hot encoded vector of all taxiway identifiers in the taxiway identifier context window is used as the label. Step S2-2: Establish a Skip-gram neural network model based on a three-layer fully connected architecture. The Skip-gram neural network is used to input the features to predict the label. Step S2-3: Determine the objective function, and train the neural network model based on the training samples and the objective function; use the trained neural network model as a taxiway embedding vector generator.

[0009] Preferably, in steps S2-3, the expression of the objective function is:

[0010] in, Describe the objective function. This represents the total number of training samples. This is the index of the taxiway sequence. This represents the total number of taxiway sequences. For taxiway indexing, This represents the total number of taxiways. For the context window size, This refers to any slide in the context window. This represents the i-th taxiway in the j-th taxiway sequence. for The context window, This indicates the calculation of the natural logarithm. Represents weight parameters Under the condition of Prediction when the center taxiway is The conditional probability.

[0011] Preferably, step S3 specifically includes: Step S3-1: Obtain the start and end point pairs of the path generation task. Obtain the vector representations of the start and end points according to the taxiway embedding vector generator. After concatenating the vector representations of the start and end points, pass them through the fusion network to obtain the start and end point features. Step S3-2: Obtain the airport operating environment, which includes numerical features and categorical features. The Wide component processes the numerical features, and the Deep component processes the categorical features. The output vectors of the Wide component and the Deep component are concatenated to obtain the environmental features. Step S3-3: Obtain the relevant flight set based on the taxiing time and time overlap of the current flight. Encode the relevant flight set based on BiLSTM, orientation embedding and temporal embedding methods to obtain the relevant flight features.

[0012] Preferably, step S3-3 specifically includes: A preliminary set of related flights is obtained based on the taxiing start and end times and the effective taxiing time of the current flight; the flights in the preliminary set of related flights are sorted and filtered according to the time overlap to obtain the set of related flights. The vector representation of each taxiway in the taxiway sequence of the flights in the relevant flight set is obtained by the taxiway embedding vector generator. All vector representations are passed through a bidirectional LSTM network to obtain the path semantic features of each relevant flight. The path semantic features are modified based on the relationship between the movement directions of related flights and the current flight to obtain the direction embedding enhancement features of each related flight; The orientation embedding enhancement features are modified based on the time interval between related flights and the current flight to obtain the time interaction features of each related flight. The relevant flight features are obtained by concatenating and filling in the time interaction features of all related flights.

[0013] Preferably, step S4 specifically includes: A Transformer decoder is constructed, which includes multiple decoding layers. Each layer contains a masked multi-head self-attention module, a multi-head cross-attention module, and a feedforward neural network module. The Transformer decoder takes static context and relevant flight features as input, and after processing through multiple decoding layers, obtains the decoder's hidden state; the static context includes origin and destination features and environmental features. Map the hidden state at the last position in the decoder's hidden state to the taxiway vocabulary space to obtain the score vector for each candidate taxiway.

[0014] Preferably, in step S4, the multi-head cross-attention module performs multi-head self-attention processing on the decoder hidden state and static context to obtain static features; Multi-head self-attention processing is applied to the decoder's hidden state and related flight features to obtain dynamic features; Static and dynamic features are concatenated and then processed sequentially through a fusion weight matrix, a ReLU activation function, and layer normalization to obtain the fused cross-attention features. These cross-attention features are then input into the cross-attention feature feedforward neural network module for further processing.

[0015] Preferably, step S5 specifically includes: Step S5-1: Initialize multiple candidate paths; Step S5-2: Based on the airport taxiway network topology data, obtain the topology constraint mask vector according to the set of taxiways directly connected to the candidate taxiways, and update the score vector of each candidate taxiway using the topology constraint mask vector; Step S5-3: Based on the probability distribution of the score vector, select a preset number of candidate slides with the highest scores from the probability distribution; Multiple new candidate paths are generated based on the expansion of the candidate taxiway, and the candidate paths are updated based on the multiple new candidate paths; Step S5-4: Return to step S5-2 and end the search when any of the following conditions are met: (1) A preset number of complete paths to the destination have been generated; (2) The length of all candidate paths has reached the preset upper limit; (3) The highest probability of all candidate paths after expansion is lower than the preset probability threshold; Finally, the current multiple candidate paths are taken as the candidate path set.

[0016] Compared with the prior art, the present invention has at least the following beneficial effects: (1) This invention takes into account the complex controller path selection decision preferences contained in airport taxiway planning. It automatically learns the controller decision patterns in historical operation data through deep learning technology, effectively expressing the implicit knowledge and decision patterns that are difficult to capture by traditional methods.

[0017] (2) The multi-source heterogeneous feature fusion encoder is used to encode the multi-source heterogeneous information in a unified manner, which can integrate dynamic information such as origin and destination constraints, environmental situation features, and traffic situation in real time, so that the generated path can adapt to different operating scenarios and real-time changing constraints.

[0018] (3) An autoregressive generative model based on Transformer is adopted. Through a dual-source parallel cross-attention mechanism and hard topology constraints, the generated taxiway path is ensured to meet the physical connectivity requirements and reflect the actual decision preferences of the controller, providing practical intelligent decision support for airport surface operation management. Attached Figure Description

[0019] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0020] Figure 1 A flowchart of the data-driven intelligent generation method for airport surface taxiways provided by the present invention.

[0021] Figure 2 This is a schematic diagram of the multi-source heterogeneous feature fusion encoder architecture provided by the present invention.

[0022] Figure 3 This is a schematic diagram of the dual-source parallel cross-attention architecture provided by the present invention. Detailed Implementation

[0023] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0024] This invention proposes a data-driven intelligent generation method for airport surface taxiways. Employing deep learning-based sequence generation technology, it transforms taxiway planning into a conditional sequence generation task, fully leveraging controller route selection preferences embedded in historical operational data. First, map matching technology is used to convert raw flight track data into standardized taxiway sequences, and a Skip-gram model is used to learn dense vector representations of taxiways, capturing spatial topological relationships and usage patterns between taxiways. Then, a multi-source heterogeneous feature fusion encoder is constructed to effectively integrate multi-source heterogeneous information such as origin / destination information, environmental conditions, and surface traffic conditions, providing rich contextual representations for path generation. Finally, an autoregressive path generation model is designed based on the Transformer architecture, using a dual-source parallel cross-attention mechanism to process static environmental information and dynamic traffic information respectively, and introducing hard topological constraints to ensure the physical feasibility of the generated paths, achieving dynamic generation of diverse taxiways.

[0025] To illustrate the effectiveness of the method proposed in this invention, the following detailed description of the above technical solution is provided through a specific embodiment, such as... Figure 1 As shown, a data-driven intelligent generation method for airport surface taxiways is disclosed, and the specific implementation steps are as follows: Step S1: Based on the collected aircraft surface taxiway tracks and airport taxiway network topology data, obtain multiple taxiway sequences to form a taxiway sequence set; the taxiway sequence includes multiple taxiway identifiers to mark different taxiways; In this step, the present invention converts the original aircraft taxiway track data into standardized taxiway sequence data suitable for training deep learning models.

[0026] (1) Extract the edge sequence of the sliding segment This invention acquires aircraft taxiing tracks based on existing airport surface surveillance radar systems. Aircraft taxiing track data typically exists in the form of GPS coordinate sequence.

[0027] In this step, the present invention pre-acquires airport taxiway network topology data, which is used to represent the topology of the airport taxiway network. The topology may include nodes and road segment edges connecting nodes, which are used to represent the connection relationship between road segments of the airport taxiway.

[0028] The goal of map matching technology is to map a continuous sequence of track points onto a discrete topological network of taxiway segments on an airport surface. This invention employs a map matching algorithm based on the shortest distance. For each observed track point in the track data sequence, the shortest distance from that track point to all candidate taxiway segments is calculated, and the taxiway segment with the smallest distance that is less than a preset threshold is selected as the matching result.

[0029] To improve the continuity of map matching results, this invention provides a track continuity verification mechanism. Specifically, for two consecutive track points in a track point sequence... and Check the edges of the matched road segments. and Whether the topological continuity constraint is satisfied means that the two road segments are topologically connected or belong to the same road segment. When a discontinuous match is detected, an edge is added to the preceding road segment. Select the distance from the current track point from the set of adjacent road segments. The nearest road segment is connected to the edge as the correction result.

[0030] The above track continuity verification mechanism can ensure that the matching results meet the physical constraints of continuous aircraft taxiing and avoid unreasonable jumps in road segment connections caused by GPS data noise.

[0031] The final result is a standardized sequence of edges connecting the gliding sections.

[0032] (2) Convert the taxiway segment connection sequence into a taxiway sequence. In this step, to reduce sequence complexity, the present invention further converts the fine-grained road segment connection sequence into a taxiway-level sequence.

[0033] The topology data of the airport surface taxiway network includes information on the association of road segment edges, with each road segment edge labeled with its associated taxiway identifier. The conversion process maps road segment edges to taxiway identifiers through a table lookup operation.

[0034] Next, consecutive identical taxiway identifiers are merged to verify the topological connectivity of adjacent taxiways, ensuring that the sequence is physically feasible and maintains temporal consistency.

[0035] The standardized taxiway sequence obtained after conversion is typically much shorter than the original segment connection sequence. The converted taxiway sequence retains the core information of the aircraft's taxiway path while possessing structural features suitable for processing by deep learning models.

[0036] In this step, the present invention obtains multiple taxiway sequences, forming a taxiway sequence set.

[0037] Step S2: Establish a neural network model and an objective function. Train the neural network model based on the set of taxiway sequences and the objective function. Use the trained neural network model as a taxiway embedding vector generator. In this step, the present invention trains a taxiway embedding vector generator, which is able to construct dense embedding vectors for each taxiway in the airport surface taxiway topology network. These embedding vectors can capture the spatial topological relationships, usage frequency patterns and semantic similarities between taxiways.

[0038] (1) Preprocess the set of taxiway sequences to obtain training samples. In this step, the set of taxiway sequences obtained in step S1 is used as the training corpus, each taxiway sequence is regarded as a sentence, and each taxiway identifier in the sequence is regarded as a word.

[0039] A taxiway vocabulary was built by statistically analyzing the taxiway identifiers that appeared in all training sequences. .

[0040] Glossary of taxiways This should include all taxiway identifiers actively used in the training data at the airport. Therefore, in this step, low-frequency taxiways that appear less frequently than a preset threshold are replaced with special markers. , It indicates unknown or special words.

[0041] In this step, the glossary of taxiway terms will be compiled. All taxiway identifiers are converted into one-hot encoded vectors, and training samples are constructed based on these one-hot encoded vectors. When given... When the taxiway is centered, its one-hot encoding is: Establish feature-label pairs, where the features are the one-hot encodings of the center taxiway, and the labels are formed by concatenating the taxiway encodings within the context window. The expression is:

[0042] in, Indicates features, Indicates a label, These represent the j-th taxiway sequence and their respective... , , , A ski track For the context window size, This indicates a vector concatenation operation; the context window contains information about the vectors before and after the center taxiway. There are 10 adjacent ski lanes, totaling 1 A contextual glide path.

[0043] (2) Establish a Skip-gram neural network model This invention employs a Skip-gram neural network to learn a dense vector representation of the taxiway. The network uses a three-layer fully connected architecture.

[0044] (3) Determine the objective function, and train the neural network model based on the training samples and the objective function. In this step, the present invention determines the objective function of the Skip-gram neural network, and the training objective is to maximize the co-occurrence probability of the center skip lane and its real context skip lane.

[0045] To facilitate numerical optimization, the objective function is transformed into a negative log-likelihood loss function, expressed as:

[0046] in, Describe the objective function. This represents the total number of training samples. This is the index of the taxiway sequence. This represents the total number of taxiway sequences. For taxiway indexing, This represents the total number of taxiways. For the context window size, This refers to any slide in the context window. This represents the i-th taxiway in the j-th taxiway sequence. for The context window, This indicates the calculation of the natural logarithm. Represents weight parameters Under the condition of Prediction when the center taxiway is The conditional probability, .

[0047] Based on the training samples, and with the objective function minimization as the goal, the weight parameters are optimized using backpropagation and stochastic gradient descent. Input and output weights and The neural network model is trained to obtain a trained neural network model, which is then used as a taxiway embedding vector generator.

[0048] After training, the vocabulary list for the gliding track will be available. Each taxiway identifier in the vocabulary has its final embedding vector representation obtained by inputting the corresponding row of the weight matrix. Each taxiway identifier has an embedding vector as follows:

[0049] in, Indicates the first Embedded vectors of taxiway identifiers Indicates input weights Glossary of corresponding taxiways A row vector of taxiway identifiers; Indicates the dimension size is The real number field, This represents the total dimension of the embedding vector.

[0050] In this step, the present invention trains a taxiway embedding vector generator. By inputting taxiway identifiers into the taxiway embedding vector generator, taxiway embedding vectors can be obtained. The taxiway embedding vectors encode the topological relationships and usage patterns of taxiways throughout the network, providing a foundation for subsequent heterogeneous feature fusion and path generation.

[0051] Step S3: Based on the origin-endpoint pair of the path generation task, the airport operating environment, and the taxiing time of the current flight, the origin-endpoint pair encoding, environmental feature encoding, and related flight encoding are performed respectively based on the taxiway embedding vector generator to obtain heterogeneous fusion features; In this step, the present invention constructs a heterogeneous feature fusion encoder architecture, which fuses multi-source heterogeneous information affecting taxiing path selection. This encoder fully utilizes multi-dimensional information from the airport operating environment, including spatial constraint information, environmental state information, and dynamic traffic information. Through encoding components, it transforms these heterogeneous data into a unified feature representation, providing a decision-making basis for subsequent path generation. The multi-source heterogeneous feature fusion encoder architecture of the present invention is as follows: Figure 2 As shown.

[0052] The starting and ending points of this step are specifically described below for coding, environmental feature coding, and related flight codes.

[0053] (1) Start and end point pair coding First, obtain the start and end point pairs for the path generation task. These start and end point pairs can be pre-input according to requirements. Both the start and end points are taxiway identifiers used to mark the taxiways where the start and end points are located.

[0054] The start and end point information constitutes the spatial constraints of the taxiway path generation task. The start and end point encoder converts discrete position identifiers into dense vector representations. In this step, the taxiway identifiers of the start and end points are input into the taxiway embedding vector generator to obtain the vector representations of the start and end points.

[0055] The embedding vector is normalized by the start-endpoint pair encoder to obtain the normalized vector of the start-endpoint pair, where the normalized vector of the start point is... The standardized vector of the endpoint is The vector representations of the start and end points are concatenated and then deeply fused through a fusion network. The expression is:

[0056] in, These are the start and end point characteristics after fusion. Indicates a converged network. This represents the normalized vector of the start-endpoint pair. The fusion network employs a combination of linear transformation, ReLU activation, layer normalization, and Dropout regularization to learn the spatial relationship between the start and end points.

[0057] (2) Environmental feature coding In this step, the present invention first acquires the airport operating environment based on sensor data or external data sources. The airport operating environment includes numerical features, such as meteorological conditions like wind speed and visibility, and categorical features, such as airlines, aircraft types, and runway configurations. The environmental feature encoder of the present invention employs a Wide&Deep architecture to process these heterogeneous features.

[0058] The Wide component processes continuous numerical features through linear transformations.

[0059] The Deep component maps discrete categorical features into dense vectors through an embedding layer. These vectors are then concatenated and input into a multilayer perceptron to learn feature interactions, yielding the Deep component's output vector. The output vectors of the Wide and Deep components are then concatenated to form an environmental feature representation containing complete environmental information, expressed as:

[0060] in, As an environmental feature, This represents the output vector of the Wide component. This represents the output vector of the Deep component.

[0061] (3) Relevant flight codes The taxiing status information of related flights constitutes a part of the dynamic traffic environment and influences the route selection decision of the target flight. The goal of the related flight encoder is to extract traffic situation information from flights with spatiotemporal proximity, providing dynamic environmental awareness for the route generation of the target flight.

[0062] The selection of relevant flights employs a time window-based method. First, the effective taxiing time of the target flight is determined as the criterion. For departing flights, engine start-up preparation time and effective taxiing time are considered; for arriving flights, their taxiing start time is used directly. Then, flights whose taxiing time intervals overlap with this effective time point are selected as the initial set of relevant flights.

[0063] When the number of relevant flights initially screened exceeds the preset limit At that time, sorting and filtering are performed based on time overlap. Time overlap is defined as the proportion of the overlap time between the taxiing time windows of the relevant flight and the target flight to the total taxiing time of the target flight, expressed as:

[0064] in, Indicates flight Time overlap These represent the start and end times of the target flight's taxiing, respectively. These represent taking the minimum and maximum values, respectively. and They represent flights The start and end times of the glide.

[0065] Sort by time overlap from highest to lowest, and select the top... These flights constitute the final set of related flights. Flights with higher time overlap are more likely to interact with the target flight in time, and their impact on route generation is more significant.

[0066] For each spatiotemporally relevant flight, its historical or planned taxiway sequence is sequentially encoded. A bidirectional long short-term memory network is used to encode the flight's taxiway sequence. This network simultaneously captures the forward and backward dependencies of the path, generating a vector representation containing the complete path semantics. For the... The taxiway sequence of each related flight is obtained by pre-training an embedding lookup layer to obtain the vector representation of each taxiway in the sequence, and then a bidirectional long short-term memory network is used to encode the taxiway sequence of the flight.

[0067] Considering the impact of the relative motion direction between flights on traffic conflicts and route selection, this invention employs a relative direction embedding mechanism. By analyzing the motion direction relationship between related flights and the target flight, they are classified into two types: "movement in the same direction" and "movement in opposite directions," and a specific embedding vector is learned for each type to obtain direction embedding enhancement features.

[0068] Temporal relationships are a factor influencing the intensity of interactions between flights. This invention provides a time difference embedding module to encode the time interval between related and target flights. This module calculates the time interval in minutes using timestamp differences, then converts it into a feature representation through an embedding layer, obtaining the features resulting from the time interactions between related flights.

[0069] Finally, the feature representations of all relevant flights are concatenated and standardized to obtain the relevant flight features. .

[0070] Through the above steps, the present invention obtains the fused start and end point features. Environmental characteristics and related flight characteristics They are collectively considered as the heterogeneous fusion feature.

[0071] Step S4: Based on the starting point of the path generation task and the heterogeneous fusion features, perform autoregressive generation based on Transformer. During the autoregressive generation process, the heterogeneous fusion features are processed by a dual-source parallel cross-attention mechanism to obtain multiple candidate gliding paths and their corresponding score vectors. In this step, the present invention constructs a sequence generation model based on the Transformer architecture to achieve autoregressive generation of taxiing paths. The model adopts an encoder-decoder architecture, progressively predicting each taxiway in the taxiing path through an autoregressive approach, fusing multi-source contextual information, and ensuring the physical feasibility of the generated path.

[0072] (1) Establish the Transformer decoder structure This invention employs a Transformer decoder architecture, optimized for the glide path generation task. The decoder includes... The layer decoding layer contains three sub-modules: masked multi-head self-attention, multi-head cross-attention, and feedforward neural network.

[0073] The autoregressive generation process begins with the starting taxiway. At the initial time step, the starting taxiway of the path generation task is used as the initial input, and its vector representation is obtained through a pre-trained embedding layer. In subsequent time steps, the decoder autoregressively predicts the next taxiway based on the generated path prefixes.

[0074] The decoder uses the glide path embedding vector generator trained in the previous step to map the embedding dimension to the hidden dimension through the embedding projection layer.

[0075] At each time step The decoder predicts the next taxiway based on the generated partial path and context information. The context information used by the decoder includes static context. And related flight characteristics (i.e., dynamic context). Among them, static context From the start and end point characteristics after fusion and environmental characteristics It is pieced together:

[0076] in, Indicates a static context.

[0077] The decoder adopts a stacked structure and includes Layer decoding layer. Projected sequence. Passing in sequence Layer-by-layer decoding processing, each decoding layer contains three sub-modules: First, the masked multi-head self-attention submodule ensures the causality of the generation process, ensuring that each position can only focus on positions preceding it. This is achieved by... The attention heads are processed in parallel to enhance the model's expressive power, and sinusoidal positional encoding is used to add positional information to the sequence.

[0078] Second, the multi-head cross-attention submodule handles the interaction between decoder state and context information. This invention improves the standard Transformer's single cross-attention to a dual-source parallel cross-attention mechanism, which processes static information (origin and destination and environment) and dynamic information (related flights) separately, and then fuses them.

[0079] Third, the feedforward neural network submodule performs a nonlinear transformation on the representation of each position.

[0080] go through After layer-by-layer processing of the decoding layer, the decoder's hidden state is obtained. The hidden state at the last position is taken. (Right now The The output projection layer maps the vectors to the taxiway vocabulary space to obtain the score vectors for each candidate taxiway. .

[0081] The dual-source parallel cross-attention architecture of the present invention is as follows: Figure 3As shown below, the multi-head cross-attention submodule in the Transformer decoder architecture is described in detail.

[0082] To address the specific characteristics of the taxiing path generation task, the single cross-attention mechanism in the standard Transformer is improved into a dual-source parallel processing architecture. This architecture processes static environmental information and dynamic traffic information separately, and then fuses them. This mechanism is implemented in each decoding layer, replacing the standard cross-attention submodule.

[0083] This mechanism uses dual-source context information. As input, two independent cross-attention branches are constructed respectively.

[0084] Static feature cross-attention handles relatively stable factors such as start-end point constraints and environmental conditions. Through a multi-head attention mechanism, the decoder's hidden state... With static context Interact to obtain static features .

[0085] Dynamic feature cross-attention processing is used to process real-time taxiing status information of relevant flights. This information changes over time and has a significant impact on path selection. The decoder's hidden state... Characteristics of related flights Dynamic features are obtained through interaction. .

[0086] The dual-source information is integrated through a neural network fusion layer. The two attention outputs are concatenated and then subjected to linear transformation and normalization to obtain the fused cross-attention feature. The fused features are added to the output of the mask self-attention through residual connections, and then enter the feedforward neural network submodule to complete the processing of this layer's decoding layer.

[0087] Step S5: Update the score vector of each candidate taxiway based on the airport taxiway network topology data; expand and filter multiple candidate taxiways based on the updated score vector to generate multiple candidate paths, wherein the candidate path is a sequence of multiple candidate taxiways.

[0088] (1) Update the score vector of each candidate taxiway based on the hard topology constraint mechanism. In this step, to ensure that the generated path conforms to the physical connectivity of the airport taxiway network, the present invention modifies the original score vector obtained in step S4. A hard topology constraint is applied. This constraint filters physically reachable candidate taxiways through masking operations to ensure the topological validity of the generated paths.

[0089] An adjacency table is pre-constructed based on airport taxiway network topology data, and a topology constraint mask vector is constructed based on the set of taxiways directly connected to the taxiways. The mask vector is added to the original score vector to obtain the score vector after constraint update. .

[0090] (2) Diversified generation of cluster search Multiple candidate paths are generated using a cluster search strategy, and then maintained. Multiple candidate taxiways are expanded in parallel, ultimately outputting multiple taxiing paths for decision-making. Based on the updated score vector, multiple candidate taxiways are expanded and filtered to generate multiple candidate paths. The implementation steps are as follows: The steps preceding this step include: Candidate path initialization. Creation. There are 10 candidate paths, each with its initial state being the starting taxiway. Each candidate path records its current path sequence and cumulative log probability score.

[0091] Candidate path expansion. For each candidate path that has not reached the destination, the current path is used as the input sequence, and the original score vector is obtained through a transformer decoder. Then, a topological constraint mechanism is applied to construct a mask based on the path endpoints, obtaining the constrained score, and the probability distribution is calculated.

[0092] This step includes: Select the top-K highest-scoring candidate slides from the probability distribution. ), extended generation A new candidate path is generated, and the cumulative log probability of each new path is the original path score plus the log probability of the newly selected glide path.

[0093] Candidate path selection: Sort all expanded candidate paths from highest to lowest according to their cumulative log probability score, and retain the highest-scoring paths. Each path is selected as a candidate set for the next iteration.

[0094] Termination condition judgment. The search ends when any of the following conditions are met: (1) A generation has been completed. The line reaches the finish line (1) The complete path; (2) The length of all candidate paths reaches the preset upper limit. (3) The highest probability of all candidate paths after expansion is lower than a preset threshold. For paths that do not reach the destination, a graph search algorithm is used to verify the probability from the end of the path to the destination. The connectivity is evaluated, and if the path is connected, the shortest path segment is completed; otherwise, the path is discarded.

[0095] Output a set of candidate paths. The algorithm output includes... A complete set of candidate paths, sorted from highest to lowest cumulative log probability score.

[0096] While the specific embodiments of the present invention depict actions or steps in a particular order, this should be understood as requiring such actions or steps to be performed in the specific order shown or in sequential order, or requiring all illustrated actions or steps to be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0097] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A data-driven intelligent generation method for airport surface taxiway, characterized in that, The method comprises the following steps: Step S1, obtaining a plurality of taxiway sequences based on the collected aircraft taxiway and the airport taxiway network topology data, to form a taxiway sequence set; the taxiway sequence comprises a plurality of taxiway identifiers for marking different taxiways; Step S2, establishing a neural network model and an objective function, training the neural network model based on the taxiway sequence set and the objective function, and taking the trained neural network model as a taxiway embedding vector generator; Step S3, based on the start and end point pair of the path generation task, the airport operating environment and the taxi time of the current flight, the start and end point pair coding, the environment feature coding and the related flight coding are performed based on the taxiway embedding vector generator to obtain a heterogeneous fusion feature; Step S4, based on the start point of the path generation task and the heterogeneous fusion feature, a Transformer-based autoregressive generation is performed, and in the autoregressive generation process, the heterogeneous fusion feature is processed by a double-source parallel cross-attention mechanism to obtain a plurality of candidate taxiways and a corresponding score vector; Step S5, updating the score vector of each candidate taxiway based on the airport taxiway network topology data; based on the updated score vector, the plurality of candidate taxiways are expanded and screened to generate a plurality of candidate paths, and the candidate paths are sequences formed by the plurality of candidate taxiways.

2. The data-driven airport taxi route intelligent generation method of claim 1, wherein, Step S1 specifically comprises: Step S1-1, matching the aircraft taxiway and the road segment edge in the airport taxiway network topology data to obtain the road segment edge identifier of each track point; Step S1-2, merging the continuous track points with the same road segment edge identifier to obtain a taxiway sequence; finally, a plurality of taxiway sequences are obtained to form a taxiway sequence set.

3. The data-driven airport taxi route intelligent generation method of claim 2, wherein, Step S2 specifically comprises: Step S2-1, preprocessing the taxiway sequence set to obtain a training sample; The preprocessing step specifically includes: constituting a runway vocabulary table from runway identifiers appearing in all runway sequences, and replacing runway identifiers appearing in the runway vocabulary table below a preset threshold with a special marker ; For each taxiway identifier in the taxiway vocabulary, a training sample is established, and the training sample comprises a feature and a label, wherein the one-hot encoding vector of the taxiway identifier is taken as the feature, and the one-hot encoding vectors of all taxiway identifiers in the context window of the taxiway identifier are taken as the label; Step S2-2, establishing a Skip-gram neural network model based on a three-layer fully connected architecture, and the Skip-gram neural network is used to input the feature to predict the label; Step S2-3, determining an objective function, training the neural network model based on the training sample and the objective function; taking the trained neural network model as a taxiway embedding vector generator.

4. The data-driven airport taxi route intelligent generation method of claim 3, wherein, In step S2-3, the expression of the objective function is: where, denotes the objective function, denotes the total number of training samples, is the index of the sliding window sequence, is the total number of sliding window sequences, is the index of the sliding window, is the total number of sliding windows, is the size of the context window, denotes any sliding window in the context window, denotes the ith sliding window in the jth sliding window sequence, is the context window of denotes the natural logarithm, denotes the weight parameter the conditional probability of given the center sliding window under the condition that is the center sliding window.​ 5. The data-driven airport taxi route intelligent generation method of claim 4, wherein, Step S3 specifically comprises: Step S3-1, obtaining the start and end point pair of the path generation task, obtaining the vector representation of the start and end point according to the taxiway embedding vector generator, splicing the vector representation of the start and end point through a fusion network to obtain the start and end point feature; Step S3-2, an airport operating environment is obtained, the airport operating environment includes numerical features and category features, numerical features are processed by a Wide component, category features are processed by a Deep component, and output vectors of the Wide component and the Deep component are spliced to obtain environment features; Step S3-3, a set of related flights is obtained according to a taxiing time and a time overlap degree of the current flight, and related flight features are obtained by encoding the set of related flights based on a BiLSTM, direction embedding and time embedding method.

6. The data-driven airport taxi route intelligent generation method of claim 5, wherein, Step S3-3 specifically includes: A preliminary set of related flights is obtained according to taxiing start and end times and effective taxiing time of the current flight, and the set of related flights is sorted and filtered according to a time overlap degree to obtain a set of related flights; A vector representation of each taxiway in a taxiway sequence of a flight in the set of related flights is obtained by a taxiway embedding vector generator, and path semantic features of each related flight are obtained by passing all vector representations through a bidirectional LSTM network; The path semantic features are modified according to a motion direction relationship between the related flights and the current flight to obtain direction embedding enhanced features of each related flight; The direction embedding enhanced features are modified based on a time interval between the related flights and the current flight to obtain time-interacted features of each related flight; The time-interacted features of all related flights are spliced and padded to obtain related flight features.

7. The data-driven airport taxi route intelligent generation method of claim 6, wherein, Step S4 specifically includes: A Transformer decoder is established, the Transformer decoder includes multiple decoding layers, each layer includes a masked multi-head self-attention module, a multi-head cross-attention module and a feedforward neural network module; The Transformer decoder takes static context and related flight features as input, and obtains decoder hidden states after layer-by-layer processing of multiple decoding layers; the static context includes terminal point features and environment features; The hidden state of the last position in the decoder hidden state is mapped to a taxiway vocabulary space to obtain a score vector of each candidate taxiway.

8. The data-driven airport taxi route intelligent generation method of claim 7, wherein, In step S4, the multi-head cross-attention module performs multi-head self-attention processing on the decoder hidden state and the static context to obtain static features; The decoder hidden state and the related flight features are processed by multi-head self-attention to obtain dynamic features; The static features and the dynamic features are spliced, and then sequentially subjected to weighting by a fusion weight matrix, ReLU activation function and layer normalization processing to obtain fused cross-attention features, and then the cross-attention features are input into a cross-attention feature feedforward neural network module for subsequent processing.

9. The data-driven airport taxi route intelligent generation method of claim 8, wherein, Step S5 specifically includes: Step S5-1, a plurality of candidate paths are initialized; Step S5-2, based on the airport taxiway network topology data, a topology constraint mask vector is obtained according to a set of taxiways directly connected to a candidate taxiway, and the score vector of each candidate taxiway is updated by the topology constraint mask vector; Step S5-3, according to the probability distribution of the score vector, a preset number of candidate taxiways with the highest scores are selected from the probability distribution; generating a plurality of new candidate paths according to the candidate slipway expansion, updating the candidate paths based on the plurality of new candidate paths; Step S5-4, return to step S5-2 until the search is ended when any of the following conditions is met: (1) a preset number of complete paths reaching the terminal point have been generated; (2) the lengths of all candidate paths reach a preset upper limit; (3) the highest probability of all candidate paths after expansion is lower than a preset probability threshold; finally, the current plurality of candidate paths are taken as the candidate path set.

Images