Airport path planning method and system based on dynamic directed graph and DQN algorithm

Abstract

The invention discloses an airport path planning method based on a dynamic directed graph and a DQN algorithm, and the method comprises the steps: S1, building a scene directed graph model based on actual airport data, collecting the latitude and longitude coordinates of an aircraft in real time through a multi-point positioning system, and dynamically updating the occupation state of node dimensions; s2, based on the real-time node data, planning an optimal path of the aircraft from a current node to a target node by adopting a DQN algorithm; s3, updating the planned path to the directed graph model, locking an occupied node state, and dynamically releasing node resources through data updated in real time; and S4, generating a sliding path according to the node state which changes in real time. A dynamic directed graph model is achieved by updating data in the directed graph in real time, the whole sliding path is divided into the shortest sliding time range of single intermediate nodes through the DQN algorithm, so that the intermediate nodes have the condition that the intermediate nodes can be repeatedly used, and the total sliding time of multiple aircrafts sliding on the scene is greatly saved.

Description

Technical Field

[0001] This invention relates to the field of aviation, and specifically to an airport path planning method and system based on dynamic directed graphs and the DQN algorithm. Background Technology

[0002] Research on airport taxiway path planning has evolved from static optimization to dynamic intelligence, mainly focusing on three major directions: mathematical models, intelligent algorithms, and graph theory methods. The following sections describe their development trajectory and limitations. Traditional mathematical models and optimization methods, in their early stages, centered on mathematical modeling, constructing accurate path planning models through linear programming, mixed-integer programming, and other methods. Marín Á.G. (2006) proposed an airport taxiway model based on a linear multi-commodity flow network, achieving high-precision path planning through edge constraints and binary variables; however, its complex model resulted in high computational resource requirements. The mixed-integer linear programming method developed by PCRoling and HG Visser (2008) can optimize traffic scheduling, but it relies on pre-set taxiway plans, and frequent adjustments during dynamic updates affect real-time performance. Wang Yu-ting (2009) used colored real-time Petri nets (CTPN) to model airport dynamic traffic, transforming multi-aircraft path optimization into a dynamic obstacle avoidance problem with time parameters; however, the computational complexity increased significantly in multi-aircraft collaborative scenarios. Adriaan W. ter Mors' (2010) environmentally perceptive path planning method faces a bottleneck of escalating computation time as the number of agents increases, and the flexibility of path selection is limited by preset rules. To improve dynamic adaptability, researchers have gradually introduced and improved intelligent optimization algorithms, such as genetic algorithms and multi-objective optimization. Yu Jiang (2014) improved the genetic algorithm for path selection and taxiing time control, but the algorithm's convergence speed is affected by the airport scale. Robert Morris's team (2016) combined multi-agent planning and probabilistic prediction modeling, which enhanced dynamic response capabilities, but the system implementation complexity and data dependence still need to be reduced. Huang Bangju et al. (2022) proposed a multi-objective optimization model based on NSGA-II, which significantly improved the overall path efficiency, but its reliance on historical data leads to insufficient adaptability to sudden scenarios. Zhou Xiaotong (2024) used a two-step genetic algorithm to comprehensively optimize taxiway and runway scheduling, which improved computational efficiency, but the complex constraints in large-scale airport scenarios may still cause solution delays. Recent research on the integration of reinforcement learning and deep learning has focused on the deep application of reinforcement learning (RL) and spatiotemporal graph neural networks (STGNN). Zhou Yunfan's team (2019) first introduced deep Q-networks (DQN) into coasting speed management, reducing conflict resolution time by 13.16%, but the scalability of the model in multi-machine collaborative scenarios has not yet been verified. Lv Liangheng et al. (2019) improved path planning convergence efficiency by evaluating the network with empirical values ​​and parallel exploration structures, but did not solve the real-time replanning delay problem under high-density traffic.Fujun Wang (2024) predicted intersection states based on STGNN (a fusion of GCN and RNN) and reduced the problem size by combining it with MILP. However, the model is sensitive to the quality of historical data, and prediction bias requires additional replanning overhead. The deepening and innovation of graph theory methods have provided a structured representation basis for path planning, leading to various efficient algorithms. Ravizza's team (2014) proposed the shortest path algorithm with time windows (QPPTW), which significantly reduces the probability of path conflicts through spatiotemporal conflict constraints. Brownlee et al. (2018) abstracted airports as directed graphs G=(V,E) and used fuzzy membership functions to quantify the uncertainty of taxiing time, laying a theoretical foundation for dynamic programming. Adacher (2018) designed an alternative graph model, using fixed arcs and alternative arcs to achieve resource priority allocation, effectively avoiding deadlock problems. Dabachine et al. (2018) improved the hybrid A* algorithm, increasing search efficiency by over 40% compared to the traditional Dijkstra algorithm, but real-time path adjustment still needs further optimization. Summary of the Invention

[0003] The purpose of this invention is to provide an airport route planning method and system based on dynamic directed graphs and the DQN algorithm. The specific technical solution is as follows: An airport path planning method based on dynamic directed graphs and the DQN algorithm is characterized by the following steps: S1, collecting airport surface data and aircraft data within the airport surface and preprocessing it to obtain real-time updated directed graph node distribution data and node occupancy data; S2, planning the aircraft's path from the current node to the target node using the DQN algorithm model based on the real-time updated directed graph node distribution data and node occupancy data in S1; S3, calculating the time range for the aircraft to occupy each intermediate node based on the intermediate nodes required for the aircraft's path from the current node to the target node in S2; S4, optimizing the subsequently planned path or increasing the waiting time of the subsequently planned path based on the time range for occupying each intermediate node in S3.

[0004] When collecting and preprocessing airport surface data in S1, after obtaining the two-dimensional planar structure diagram of the airport's taxiways, the taxiways are segmented and numbered through their intersections. Each taxiway segment is abstracted as a node in a directed graph, resulting in the distribution data of the directed graph nodes of the surface.

[0005] When collecting and preprocessing data of aircraft within the airport surface in S1, the MLAT latitude and longitude data of the aircraft are collected in real time, the nodes occupied by the aircraft are identified, and the data is updated in real time in the directed graph node distribution data of the surface, thus obtaining the real-time updated directed graph node distribution data and node occupancy data.

[0006] When planning the path of an aircraft from the current node to the target node in S2, the node occupancy status of the current node occupied by the aircraft and the node usage status that the aircraft is preparing to use during taxiing are updated in real time.

[0007] The usage status of nodes to be used during aircraft taxiing can be planned by two paths simultaneously.

[0008] The network structure formula for the DQN algorithm model is: in, It is the action-value function, representing the state. Take action below At that time, the "value" (usually the expected return) that can be obtained; It is a state-value function, representing the state. Its inherent "value". It is the dominance function, used to measure the dominance in a state. Next, specific actions In DQN, the network outputs the degree of good or bad relative to the average level of this state. and Then, by combining the above formula (4.13), we can form... ; Indicates all possible actions Summing the dominant function; It represents the number of possible actions in the action space.

[0009] In S3, when calculating the time range of an aircraft occupying each intermediate node, the time when the aircraft's nose enters node a and the time when the aircraft's tail leaves node a are calculated based on the taxiway length of node a and the aircraft's taxiing speed at node a, thus obtaining the time range of the aircraft occupying node a.

[0010] Node a includes taxiway nodes and intersection nodes. When an aircraft is at an intersection node, its taxiing speed is 15-30 km / h, and when an aircraft is at a taxiway node, its taxiing speed is 40-60 km / h.

[0011] When optimizing the later-planned path in S4, it is necessary to ensure that the time interval between the earlier and later paths occupying the same node is greater than 9 seconds. That is, the time interval of the first path occupying node a is (16:30:10-16:30:40), and the time interval of the second path occupying node a should be outside the time interval (16:30:01-16:30:49).

[0012] An airport path planning system based on dynamic directed graphs and the DQN algorithm includes: a data acquisition and preprocessing module, which receives real-time MLAT data from aircraft on the airport surface, including information such as the aircraft's position, speed, and heading; performs noise reduction, format standardization, and status synchronization processing on the raw data to ensure data integrity and timeliness, providing high-precision airport surface geographic coordinate input for path planning; a DQN path planning client module, which calls the DQN algorithm model to perform path planning based on the real-time updated directed graph node distribution and node occupancy data; and a web server module, which is used for... This module updates the directed graph node distribution and occupancy data in real time. It receives preprocessed data from the data acquisition and preprocessing module and path planning data from the DQN path planning client module. It uses the received data to update the directed graph node distribution and occupancy data in real time. If a path planning request is received, it sends the current state of the directed graph to the client's DQN algorithm for path planning. After the client's DQN completes the path planning and the path planning criteria are met, it sends the planned taxiway path nodes to the Web server module. The Web server then updates the nodes in the airport's directed graph based on the received taxiway path nodes.

[0013] The above-mentioned technical solution of the present invention has the following beneficial technical effects: by dividing the entire taxiing path of an aircraft into the time range of a single intermediate node, the intermediate node can be reused, which greatly saves the overall taxiing time of multiple aircraft when taxiing on the ground, and is particularly beneficial to the operational efficiency of the airport during busy periods. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the distribution data of the directed graph nodes in the field in this invention; Figure 2 This is a schematic diagram of the current position and state of the aircraft in the directed graph of this invention; Figure 3 This is a schematic diagram of the aircraft taxiing path established in a directed graph in this invention, considering only the state dimension of a single node. Figure 4 This is a schematic diagram of the actual two-dimensional planar position of the aircraft when it enters the G node from the HG node in this invention. Figure 5 This is a visual diagram illustrating the dynamic changes of the directed graph when an aircraft moves from node HG to node G, considering only the state dimension of a single node in this invention. Figure 6This is a schematic diagram of the aircraft completely entering the G node from the HG node in this invention, considering only the usage state dimension of one node. Figure 7 This is a schematic diagram of a directed graph for visualizing the taxiing path of an aircraft under two usage state dimensions at one node in this invention. Figure 8 In this invention Figure 7 Based on this, a visualization diagram of the node usage status after the subsequent aircraft follows the taxiing path is established. Figure 9 This is a schematic diagram of a visualized directed graph that traverses the taxiing path on the previous aircraft's taxiing path in this invention. Figure 10 This is a schematic diagram of the AGV state construction in this invention; Figure 11 This is a schematic diagram of the DQN algorithm framework in this invention; Figure 12 This is a schematic diagram illustrating the creation of a visualized following sliding path in a dynamic directed graph in this invention; Figure 13 This is a schematic diagram of the path planning method in this invention that visualizes and avoids conflicts between taxiway intersections and taxiway usage nodes; Figure 14 This is a schematic diagram of the path planning method combining dynamic directed graphs and DQN algorithm in this invention.

[0015] Because multiple path planning was involved, different colors were used to represent them. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0017] An airport path planning method based on dynamic directed graphs and the DQN algorithm includes: S1, collecting airport surface data and aircraft data within the airport surface and preprocessing it to obtain real-time updated directed graph node distribution data and node occupancy data; S2, planning the aircraft path from the current node to the target node using the DQN algorithm model based on the real-time updated directed graph node distribution data and node occupancy data in S1; S3, calculating the time range for the aircraft to occupy each intermediate node based on the intermediate nodes required for the aircraft's path from the current node to the target node in S2; S4, optimizing the subsequently planned path or increasing the waiting time of the subsequently planned path based on the time range for occupying each intermediate node in S3. By breaking down the entire aircraft taxiing path into the time range of occupancy of individual intermediate nodes, the intermediate nodes can be reused, significantly reducing the overall taxiing time of multiple aircraft on the surface, which is particularly beneficial to the operational efficiency of the airport during peak hours.

[0018] When collecting and preprocessing airport surface data in S1, after obtaining the two-dimensional planar structure diagram of the airport's taxiways, the taxiways are segmented and numbered through their intersections. Each taxiway segment is abstracted as a node in a directed graph, resulting in the distribution data of the directed graph nodes of the surface.

[0019] When collecting and preprocessing data of aircraft within the airport surface in S1, the MLAT latitude and longitude data of the aircraft are collected in real time, the nodes occupied by the aircraft are identified, and the data is updated in real time in the directed graph node distribution data of the surface, thus obtaining the real-time updated directed graph node distribution data and node occupancy data.

[0020] When planning the path of an aircraft from the current node to the target node in S2, the node occupancy status of the current node occupied by the aircraft and the node usage status that the aircraft is preparing to use during taxiing are updated in real time.

[0021] The usage status of nodes to be used during aircraft taxiing can be planned by two paths simultaneously.

[0022] The network structure formula for the DQN algorithm model is: in, It is the action-value function, representing the state. Take action below At that time, the "value" (usually the expected return) that can be obtained; It is a state-value function, representing the state. Its inherent "value". It is the dominance function, used to measure the dominance in a state. Next, specific actions In DQN, the network outputs the degree of good or bad relative to the average level of this state. and Then combine them using the formula (4.13) above to form ; Indicates all possible actions Summing the dominant function; It represents the number of possible actions in the action space.

[0023] In S3, when calculating the time range of an aircraft occupying each intermediate node, the time when the aircraft's nose enters node a and the time when the aircraft's tail leaves node a are calculated based on the taxiway length of node a and the aircraft's taxiing speed at node a, thus obtaining the time range of the aircraft occupying node a.

[0024] Node a includes taxiway nodes and intersection nodes. When an aircraft is at an intersection node, its taxiing speed is 15-30 km / h, and when an aircraft is at a taxiway node, its taxiing speed is 40-60 km / h.

[0025] When optimizing the later-planned path in S4, it is necessary to ensure that the time interval between the earlier and later paths occupying the same node is greater than 9 seconds. That is, the time interval of the first path occupying node a is (16:30:10-16:30:40), and the time interval of the second path occupying node a should be outside the time interval (16:30:01-16:30:49).

[0026] An airport path planning system based on dynamic directed graphs and the DQN algorithm includes: a data acquisition and preprocessing module, which receives real-time MLAT data from aircraft on the airport surface, including information such as the aircraft's position, speed, and heading; performs noise reduction, format standardization, and status synchronization processing on the raw data to ensure data integrity and timeliness, providing high-precision airport surface geographic coordinate input for path planning; a DQN path planning client module, which calls the DQN algorithm model to perform path planning based on the real-time updated directed graph node distribution and node occupancy data; and a web server module, which is used for... This module updates the directed graph node distribution and occupancy data in real time. It receives preprocessed data from the data acquisition and preprocessing module and path planning data from the DQN path planning client module. It uses the received data to update the directed graph node distribution and occupancy data in real time. If a path planning request is received, it sends the current state of the directed graph to the client's DQN algorithm for path planning. After the client's DQN completes the path planning and the path planning criteria are met, it sends the planned taxiway path nodes to the Web server module. The Web server then updates the nodes in the airport's directed graph based on the received taxiway path nodes.

[0027] An airport route planning method based on dynamic directed graphs and the DQN algorithm includes: S1. Collect and preprocess airport surface data and aircraft data within the airport surface to obtain real-time updated directed graph node distribution data and node occupancy data. In S1, during airport surface data collection and preprocessing, after obtaining the 2D plan view of the airport's taxiways, the taxiways are segmented and numbered based on their intersections. Each taxiway segment is abstracted as a node in the directed graph, resulting in the directed graph node distribution data. During aircraft data collection and preprocessing in S1, the MLAT latitude and longitude data of the aircraft is collected in real-time. The nodes occupied by the aircraft are identified and updated in real-time in the directed graph node distribution data, resulting in the real-time updated directed graph node distribution data and node occupancy data.

[0028] Specifically, such as Figure 1 As shown, a directed graph is constructed using interval partitioning. The established directed graph model includes nodes and edges. Nodes represent taxiway intersections and straight taxiways connecting these intersections, while edges represent the connections and directions between these nodes. This model construction method enables the establishment of a directed graph model that can simulate the taxiing paths and positions of aircraft on the airport surface, providing a foundation for path planning using the DQN algorithm. Figure 1 The upper half of the diagram shows the two-dimensional planar structure of the airport's taxiways. The lower half abstracts this two-dimensional planar structure of the taxiways into a directed graph structure of the airport as described in the method. Since the airport in the diagram does not have unidirectional taxiways, the constructed directed graph is a bidirectional directed graph. The connections between any nodes in the graph are bidirectional. Figure 1 The graph constructed is a bidirectional directed graph; for the sake of simplicity, the directional arrows have been omitted.

[0029] Based on the airport layout, this paper divides runways 14-32 into lines A, taxiways closer to the runways into lines B, and taxiways closer to the apron into lines C. Furthermore, the straight taxiways connecting taxiway intersections are also abstracted as nodes in a directed graph and jointly named after the names of the taxiway intersections at both ends. The taxiway division begins with runway 32, and the turns at both ends of the runway are abstracted as independent nodes in the directed graph, named nodes A and H respectively. Taxiway intersections are named sequentially as B, D, E, F, and G, where B, E, and F are nodes connecting apron taxiways, D is a node connecting runways and apron taxiways, and G is a node connecting runways. On line C, the connection between the parking stands and taxiway line C is divided into nodes, with each parking stand corresponding to a node named C29, C28, ..., C21, C1, C2, ..., C11, and the parking stands themselves named 29, 28, ..., 21, 1, 2, ..., 11. Taxiway intersection node D connects to runways C6, C7, and 14-32. Since runways 14-32 are not considered in the path planning algorithm during this study, node D needs to be connected to nodes C6 and C7, but not to nodes 14-32, when constructing the directed graph. By segmenting and numbering taxiways according to taxiway intersections, the operational structure of the airport surface can be more clearly described. These divisions provide the foundation for the subsequent construction of the directed graph of the airport surface, abstracting each taxiway segment as a node in the directed graph, and the connections between taxiways as directed edges. This abstraction method effectively simplifies the complexity of the airport surface.

[0030] Based on the constructed directed graph, a method for constructing a dynamic directed graph is proposed, aiming to optimize the utilization efficiency and computational complexity of airport taxiway resources and improve the real-time performance of path planning. In the design of the dynamic directed graph, this method employs a real-time resource release mechanism: when an aircraft passes a path node, by detecting the aircraft's MLAT latitude and longitude data, once it is confirmed that the aircraft has fully entered the next node, the resources of that node are immediately released, while the resources of nodes not yet visited remain active. This method of directing and scheduling aircraft is implicit in the daily scheduling of airport controllers. By constructing a dynamic directed graph of the airport surface, taxiway resources can be utilized more efficiently, unnecessary occupation can be avoided, and thus the overall operational efficiency of the airport surface can be improved.

[0031] S2. Based on the real-time updated directed graph node distribution and occupancy data from S1, the DQN algorithm model is used to plan the aircraft's path from the current node to the target node. In S2, when planning the path, the node occupancy status of the current node and the node usage status to be used during taxiing are updated in real time. The node usage status to be used during taxiing can be planned by two paths simultaneously. Specifically, to achieve the construction of a dynamic airport surface directed graph, multiple state feature dimensions are added to each node in the directed graph, including one node occupancy dimension and two node usage status dimensions. These feature dimensions can dynamically reflect the resource occupancy, usage status, and spatial location information of aircraft at each node on the taxiway, thus providing more accurate data support for dynamic airport surface scheduling and path planning. Figure 2 As shown, the red sections represent the occupied dimension used to represent the current position and state of the aircraft in the directed graph. Let... Indicates the first The dimension occupied by each node, whose value is of boolean type: in, This is the index of the node. (Based on the occupied dimension) This allows for real-time tracking of aircraft positions on the ground, providing crucial support for path planning. In a dynamic directed graph, two nodes are used, with the state dimension representing the node's usage. Let... and They represent the i-th The node has two usage state dimensions, whose values ​​are boolean: ; in, This serves as the index for the node. These two usage status dimensions allow for a more detailed description of node usage, providing more accurate support for dynamic scheduling and path planning.

[0032] like Figure 3 As shown, the aircraft taxiing path is constructed in a directed graph considering only one node using the state dimension. This represents the nodes in a directed graph, where red nodes represent the real-time position of the aircraft, and red and blue nodes jointly represent the planned taxiing path. Figure 3 The established gliding path is as follows: in, Represents the nodes in a directed graph. This refers to the name of the node. In the study, path nodes already planned for aircraft taxiing were... The state dimension is uniformly changed to 1 to indicate that these nodes have been occupied or activated.

[0033] like Figure 4 and Figure 5 As shown, considering only one node using the state dimension, the visualization of the dynamic directed graph changes when the aircraft moves from node HG to node G. Figure 4 This represents the actual two-dimensional planar position of the aircraft when it moves from node HG to node G. When an aircraft moves from node HG to node G, it needs to occupy the resources of two nodes: node HG and node G. For example... Figure 5 The red nodes indicate that these two nodes are occupied by an aircraft, meaning the aircraft is located within these two nodes. Other nodes planned as aircraft taxiing paths but not yet reached by the aircraft are represented by blue nodes. Therefore, the occupancy dimension of this taxiing path is represented as follows: in This represents the set of occupied dimensions for nodes HG, G, GF, F, C26, C25, C24, and 24 of this taxiing path. The node usage state dimension of this taxiing path is represented as follows: in This represents the set of usage state dimensions for nodes HG, G, GF, F, C26, C25, C24, and 24 of this gliding path.

[0034] like Figure 6 As shown, this is a visualization of the dynamic directed graph changes when an aircraft completely moves from node HG to node G, considering only the state dimension of a single node. Figure 5 The difference is that the red marker for node HG in the diagram becomes colorless, indicating that the aircraft has completely left node HG and entered node G. Other taxiway nodes are still represented in blue, indicating that the aircraft has not yet reached these planned nodes. The colorless HG node signifies that the node's resources have been released and can be replanned by the algorithm as a taxiway path for other aircraft. The occupancy dimension of this taxiway path is represented as follows: in This represents the set of dimensions occupied by nodes HG, G, GF, F, C26, C25, C24, and 24 of this gliding path.

[0035] The nodes of this gliding path are represented using the state dimension as follows: in This represents the set of usage state dimensions for nodes HG, G, GF, F, C26, C25, C24, and 24 along this gliding path. Figure 5 The difference is This means that node HG can now be used by other aircraft's taxiing paths.

[0036] Considering only one node usage state dimension, the above method cannot handle the common aircraft following taxiing scenarios in apron control. When aircraft need to follow each other in the same direction, since a node can only be activated once, the taxiing path of the following aircraft cannot be established in the dynamic directed graph. Furthermore, when other aircraft occupy the target node, the study found that if a node has already been planned as a taxiing path for another aircraft, but the current aircraft has not yet arrived at that node, a new taxiing path can be established within a redundancy time constraint, thereby improving taxiing efficiency at the airport. Therefore, this study sets two usage state dimensions for the nodes in the directed graph, enabling this path planning method to be learned by the algorithm and applied to actual planning.

[0037] like Figure 7 As shown, a directed graph visualizes the taxiing path established by an aircraft with one node and two usage state dimensions. The taxiing path established by the aircraft in the graph is as follows: in The nodes in a directed graph, Represents the node name.

[0038] Below each node are two numerical identifiers, indicating the node's usage status dimension. Figure 7 In the dynamic directed graph, red nodes represent the current position of the aircraft, blue nodes represent paths that have been planned for use by the aircraft but have not yet been reached, and colorless nodes represent taxiing paths that have not been planned for any aircraft. The usage state dimension of blue and red nodes is [1, 0], indicating that these nodes have been planned as taxiing paths for an aircraft. In addition, yellow nodes represent the starting point of the taxiing path to be planned, and green nodes represent the ending point of the taxiing path. Figure 7 The occupied dimensions of nodes HG, G, GF, F, C26, C25, C24, and 24 are represented as follows: in This represents the occupied dimension set of nodes HG, G, GF, F, C26, C25, C24, and 24.

[0039] The usage status dimensions of nodes HG, G, GF, F, C26, C25, C24, and 24 are as follows: in, Two dimensions representing the use of nodes, Represents the node name. In the above formula... This is because It's an airport tarmac, only big enough for one aircraft at a time. To prevent it from being detected and misused by algorithms, it's designated as a parking area. The nodes are all set to (1,1) when the state dimension is used for path planning.

[0040] Figure 8Shown in Figure 7 Based on this, a visualization of the node usage status after the subsequent aircraft follows the taxiing path is established. In the figure, red nodes represent the position of the previous aircraft in the dynamic directed graph, orange nodes represent the taxiing path nodes planned by the current aircraft, yellow nodes represent the starting point of the current aircraft's taxiing path, blue nodes represent path nodes that have been planned for use by the previous aircraft but have not yet been reached, and green nodes represent the end point of its taxiing path. Figure 8 The nodes (GF, F, C26, C25) within the red box represent taxiway nodes shared by two aircraft. The two numerical identifiers below each node indicate its usage status. Nodes GF, F, C26, and C25, with their identifiers changed to (1,1), indicate that these nodes have been planned as taxiways by both aircraft, their resources are fully utilized, and they cannot be used for taxiway path planning by other aircraft. Figure 7 In the diagram, the state dimension of a node is (1,0), meaning the node usage state of the previous aircraft is established at... The node usage status of the next aircraft is then established on... superior. Figure 8 The occupancy dimensions of nodes HG, G, GF, F, C26, C25, C24, 24, and 25 are represented as follows: in This represents the occupied dimension set of nodes HG, G, GF, F, C26, C25, C24, 24, and 25.

[0041] The usage status dimensions of nodes HG, G, GF, F, C26, C25, C24, 24, and 25 are as follows: in The two dimensions representing the usage state of a node are: Represents the node name. Figure 9 There are two sliding paths, named sliding paths A and B. Sliding path A is as follows: B's sliding path is: in Represents node symbols, Indicates the node name.

[0042] exist Figure 9In the diagram, the small orange arcs on the three blue nodes indicate that nodes C26, C25, and C24 are shared nodes for taxiing paths A and B. The red nodes represent the current position of aircraft on taxiing path A in the dynamic directed graph, and the blue nodes represent nodes that aircraft on taxiing path A have not yet reached. The yellow nodes represent the starting node of aircraft on taxiing path B, the green nodes represent the target node of aircraft on taxiing path B, and the orange nodes represent nodes that aircraft on taxiing path B have not yet reached. Colorless nodes are marked (0,0) as not used by any aircraft; blue nodes are marked (1,0) as activated by aircraft on taxiing path A; orange nodes are marked (0,1) as activated by aircraft on taxiing path B; and blue nodes with small orange arcs are marked (1,1) as being jointly activated and used by aircraft on taxiing paths A and B. Figure 9 The two taxi paths in the data indicate that when the aircraft on taxi path A has not yet reached nodes C26, C25, and C24, taxi path B can be established under the condition of satisfying the redundancy time interval, thereby improving the utilization efficiency of airport surface resource nodes.

[0043] exist Figure 9 The dimension occupied by the nodes of path A in the diagram is represented as follows: in This represents the occupied dimension set of nodes HG, G, GF, F, C26, C25, C24, and 24.

[0044] The node usage state dimension of path A is: in express The usage status of the two nodes of the point. Indicates the node name.

[0045] The dimension occupied by the B path node is represented as follows: in This represents the set of dimensions occupied by nodes C27, C26, C25, C24, C23, ..., C1, E, ..., A.

[0046] The B path node's usage state dimension is: in express The usage status of the two nodes of the node. This represents the node name. The DQN algorithm design is explained below: A graph-based navigation environment, GraphEnv, was designed based on a reinforcement learning framework. Its core objective is to train an agent to find the optimal path from a random starting point to a target node within a graph network. The environment is implemented using the OpenAI Gym interface standard, and its main features include: Topology: A bidirectional directed graph structure with 53 nodes (implemented using NetworkX) is used. Runway nodes 14-32 were excluded during modeling because runway usage is not considered in path planning. In actual airport operations, runway usage is managed by designated personnel and is primarily used for aircraft queuing and takeoff.

[0047] Dynamic node characteristics: The two usage state dimensions of the node are randomly updated in three common airport scheduling scenarios to simulate the dynamic changes in actual operation.

[0048] Orientation awareness: Introducing a coordinate system and orientation awareness. This allows the agent to receive information from surrounding nodes in the environment and make better decisions.

[0049] Action constraints: Validity detection based on adjacency relationships. In a directed graph, not all nodes have the same adjacent nodes. Therefore, action constraint detection is needed to prevent incorrect actions from causing the agent to become stuck in an inaccessible state.

[0050] State Space Setup for DQN Algorithm: The path planning problem in DQN typically uses the agent's current position or internal network information as input. In our algorithm, we adopt an input method based on the agent's current state information. In the deep Q-learning environment, each node's information includes five dimensions: node number, node usage state (1), etc. ), Node usage status 2 ( The system includes taxiing time weights (time) and the node's position coordinates (x, y) in the directed graph. Specifically, node represents the current node position of the agent, while state1 and state2 indicate whether the node has been planned and used by the aircraft.

[0051] if A value of 0 indicates that the node's If the path is not occupied, it can be planned as an aircraft path if certain conditions are met.

[0052] if A value of 0 indicates that the node's The path is not occupied and can also be used for planning.

[0053] when and When both are 1, it means that the node is occupied by two aircraft and the current node is unavailable.

[0054] if and If both are 0, it means that the node has not been planned for use by any aircraft and is in an idle state.

[0055] We incorporate the starting and target nodes in the path planning into the state space vector, and also include the available adjacent nodes around the agent in the state space.

[0056] like Figure 10 As shown in the diagram, the AGV (automatic vehicle) perceives its surrounding environment in a state construction diagram. In the diagram, the AGV is located at node F, which has three adjacent nodes: FE, C26, and GF. Node FE is located in front of node F (set to 0 in the program), node C26 is located to the left of node F (set to 2 in the program), and node GF is located behind node F (set to 3 in the program). There are no nodes to the right of F. Therefore, the state of the surrounding environment perceived by the AGV is as follows: Figure 2 As shown in the small diagram on the right, the front state is 0, the right state is 1, the left state is 0, and the back state is 0. Additionally, if a node has two usage states... A value of (1,1) indicates that both usage states of this node are activated, so the AGV's perception value for this node is 1. If it is... Figure 2 Usage status of node C26 in Then the AGV perceives the state of the left-2 node (C26) as 1, which is unusable.

[0057] The state space settings for this article are as follows: in This represents the current position of the AGV; and Indicates the current node usage status; Indicates the direction currently selected by the AGV; and These represent the x-coordinates of the current AGV and the target node, respectively. and These represent the ordinate values ​​of the current AGV and the target node, respectively. and These represent the relative positions of the current AGV and the target node AGV, respectively. These represent the current environmental conditions surrounding the AGV. Represents information from the front lines. Information on the right side of the table. Represents information on the left. Represents information from the rear. This indicates the starting point location information of the current planned path. This indicates the location information of the target node in the current planned path.

[0058] Action space design in the DQN algorithm: In the path planning problem using the DQN algorithm, the design of the action space directly affects the agent's decision-making ability and the flexibility of path planning. This section will detail the methods for setting up the action space. The action space will be set based on the structural characteristics of the directed graph of the airport surface. For example... Figure 2 The action space is defined by four directions: forward (0), right (1), left (2), and backward (3). Each direction corresponds to a possible movement of the agent in the graph. Through the definition of the action space, the agent can select an effective movement direction at each time step, thereby gradually approaching the target node. Furthermore, an action validity detection mechanism is needed to ensure that the agent does not select invalid actions, such as moving to a direction without adjacent nodes. The agent's action selection logic is based on... Omitted. Specifically, the agent at each time step uses a probability... Randomly select a valid action, or with probability Select the action with the highest current Q value. The action selection process can be divided into the following steps.

[0059] The strategy, the agent's action selection follows The strategy adopted is the one used in the paper "Multi-robot Path Planning Based on Deep Reinforcement Learning DQN Algorithm" by Yang, Y team (2020). Strategy

[50] , expressed as: in, Indicates at time step The action chosen. It is the exploration rate, which controls the probability of randomly selecting an action. Indicates the state Select action The Q value.

[0060] Random action selection, when the agent chooses a probability When selecting a random action, it randomly chooses one from all valid actions in the current state. The definition of a valid action is based on the current state and the constraints of the environment. Specifically, the logic for random action selection is as follows: To obtain the neighboring nodes of the current node: The agent first obtains all the neighboring nodes of the current node.

[0061] Mapping actions to adjacent nodes: Each action is mapped to its corresponding adjacent node through an action mapping table.

[0062] Filter valid actions: Based on environmental constraints (such as node usage status, path validity, etc.), filter out all valid actions.

[0063] Random action selection: Randomly select an action from the list of valid actions.

[0064] The formula for random action selection is expressed as: in, Indicates the state The set of all valid actions under this condition. This means randomly selecting an element from the set.

[0065] Q-value maximization action selection, when the agent chooses action with probability... When selecting the action with the highest Q-value, it calculates the Q-value of each action in the current state using the policy network and selects the action with the highest Q-value. The specific steps are as follows: State input: Enter the current state It is converted into tensor form and input into the policy network.

[0066] Calculate the Q-value: Calculate the Q-value for each action using the policy network.

[0067] Select the action with the highest Q value: Select the action with the highest Q value.

[0068] Verify the validity of the action: If the selected action is valid, return the action; otherwise, return to the random action selection logic.

[0069] The formula for maximizing Q-value action selection is expressed as: in, Indicates the state Select action The Q value.

[0070] Action validity verification involves the AI ​​verifying the validity of a selected action. The validity of an action depends on the current state and environmental constraints. The logic for action validity verification is as follows: Get the current node and the next node. Determine the current node and the target node based on the current state and the selected action.

[0071] Check the node status. Based on the status of the next node (e.g., ... and ), to determine whether the selected action is valid.

[0072] Path validity check. Ensure the next node is not on the previously selected path and that other environmental constraints are met.

[0073] The formula for validating the action is expressed as: Through the above steps, the agent can determine the time step based on... The strategy selects actions and ensures the effectiveness of the selected actions. This process plays a crucial role in the DQN algorithm, helping the agent find a balance between exploration and exploitation, thereby progressively optimizing the strategy.

[0074] The design of the reward function in the DQN algorithm is crucial to the learning performance of the agent. The reward function evaluates the agent's performance after each action and guides the agent towards the target learning direction through positive and negative rewards. This section describes the logic and mathematical expression of the reward function in detail.

[0075] reward function It is to compute the state of the intelligent agent. Next action And transition to state The reward obtained afterward. The design of the reward function takes into account the following factors: Movement penalty: The agent is penalized for each step it takes to encourage efficient path planning.

[0076] Objective function reward: A high reward is given when the agent reaches the target node.

[0077] Distance change reward: The reward value is dynamically adjusted based on the distance change between the current state and the target node (including glide time weight and BFS distance).

[0078] Invalid movement penalty: An additional penalty is imposed if the agent does not change its position.

[0079] The specific logic of the reward function, and the calculation logic of the reward function, can be divided into the following parts: Movement penalty: The agent incurs a fixed penalty for each step it takes, expressed by the formula: No movement penalty: If the agent's position has not changed (i.e.) Then, additional penalties are applied, expressed by the formula: Target Reward: If the agent reaches the target node (i.e. A high reward will be given if the reward is obtained, as expressed in the formula: Glide time weighting change reward: The agent adjusts the reward value based on the change in time distance between the current state and the target node: If the weight of gliding time is reduced (i.e.) If the reward is positive, then a positive reward will be given.

[0080] If the weight of gliding time is increased (i.e.) If the reward is negative, then a negative reward will be given.

[0081] The formula is expressed as: BFS (Breadth-First Search) distance change reward: The agent adjusts its reward value based on the change in the BFS distance between the current state and the target node: If the BFS distance decreases (i.e.) If the reward is positive, then a positive reward will be given.

[0082] If the BFS distance increases (i.e.) If the reward is negative, then a negative reward will be given.

[0083] The formula is expressed as: Combining the above parts, the final formula for calculating the reward function is: in: It's a movement penalty. It is an invalid movement penalty. It is a target reward. Rewards based on changes in gliding time weighting. It is a reward based on the distance change in BFS.

[0084] The DQN algorithm architecture is based on Dueling DQN and a network design that prioritizes experience replay, such as... Figure 11 As shown, specifically: Dueling DQN network architecture enhances the algorithm's ability to evaluate state value by decomposing the Q-value into a value function and an advantage function. The network structure is defined as follows: Network structure formula Let the state be The action is Then the Q-value of Dueling DQN is calculated as follows: in, It is the action-value function, representing the state. Take action below At that time, the "value" (usually the expected return) that can be obtained; It is a state-value function, representing the state. Its inherent "value". It is the dominance function, used to measure the dominance in a state. Next, specific actions In Dueling DQN, the network outputs the degree of good or bad relative to the average level of this state. and Then combine the above formulas to form ; Indicates all possible actions Summing the dominant function; It represents the number of possible actions in the action space.

[0085] Network implementation: Shared feature layer (3 fully connected layers): in , , , For the state dimension.

[0086] Value Branches:

[0087] Advantageous branches: The priority experience replay mechanism improves the utilization rate of key experiences through a non-uniform sampling strategy. Its core logic is as follows: Priority definition: Every experience Priority is based on the absolute value of the timing difference error (TD-error): in, It is the first Prioritize the experience points; It is the first The time difference error (TD-error) corresponding to each empirical point; It is a very small positive constant used to avoid the priority value being zero.

[0088] Sampling probability: The sampling probability and priority have a power-law relationship: in, It is the sampling number The probability of a piece of experience. It is a prioritization hyperparameter used to adjust the degree of influence of priority on the sampling distribution. Then it degenerates into uniform sampling (all samples have the same probability). It will then follow Sampling at an angle of size, The larger the value, the higher the probability of it being sampled. It is a normalization factor used to ensure that the sum of the sampling probabilities of all experiences is 1.

[0089] Importance sampling weights: To compensate for sampling bias, a weight correction is introduced: in, It is the first Importance Sampling Weight. It is the total number of experiences currently stored in the replay buffer. It is the sampling number The probability of a piece of experience. It is a hyperparameter of importance sampling correction, when No corrections are made at that time. Make full corrections in a timely manner, during training, as training progresses. It will gradually increase from a smaller value to 1.

[0090] Priority update rules: Update the priority of selected experiences after each sampling: in, It is the first The experience points will be given a new priority after the update. It is the target Q-value during training. It is the Q-value estimated by the current online network for the state-action pair.

[0091] Explore strategy design, adopt Strategy, dynamically decaying exploration rate: Setting the initial exploration rate attenuation coefficient Minimum exploration rate .

[0092] S3. Calculate the time range for the aircraft to occupy each intermediate node based on the intermediate nodes required for the aircraft's path from the current node to the target node, as calculated in S2. When calculating the time range for the aircraft to occupy each intermediate node in S3, the time it takes for the aircraft's nose to enter node a and the time it takes for the aircraft's tail to leave node a are calculated based on the taxiway length of node a and the aircraft's taxiing speed at node a, thus obtaining the time range for the aircraft to occupy node a. Node a includes taxiway nodes and intersection nodes. When the aircraft is at an intersection node, the taxiing speed is 15-30 km / h; when the aircraft is at a taxiway node, the taxiing speed is 40-60 km / h. S4. Optimize the subsequently planned path or increase the waiting time of the subsequently planned path based on the time range for occupying each intermediate node calculated in S3.

[0093] The routing task for two aircraft taking off and following each other under the design (1) is shown in Table (1) below. The table shows the flight priority, flight number, operation type, start time, starting point of taxiing and target node to be reached.

[0094] Flight schedule for two aircraft following each other as they taxi away from the airport (1) Priority Flight number Run type Start time starting point end 1 KLM807 leaving 14:20:20 2 A 2 CSN6238 leaving 14:21:00 3 A Method 1 uses a dynamic directed graph combined with the DQN algorithm for path planning. Figure 12 As can be seen, when the first aircraft was at node C4, the second aircraft established its taxiway path at 14:21:00. At this time, the first aircraft was located at node C4, which had been designated as a shared taxiway path for both aircraft. This status can be confirmed by the node's index [1,1,1]. The taxiways for the two aircraft are as follows: The taxiing path of the first aircraft when the second aircraft begins taxiing is as follows: The taxiing path of the second aircraft is as follows: Figure 12 In the diagram, red nodes indicate that the second aircraft is at parking position 3. Cyan nodes represent the taxiing path used by the second aircraft. Blue nodes represent the taxiing path planned for the first aircraft. Blue nodes with magenta indices are fully activated and planned as a shared taxiing path for both aircraft; node indices are [0, 1, 1]. The two aircraft arrive at target node A at 14:24:03, sharing a 223s taxiing time weight. A total of 12 nodes are used (including parking position 2 and 3 nodes).

[0095] Method 2, following the traditional aircraft scheduling method, only releases the taxiway node used by the first aircraft to other aircraft when the first aircraft arrives at the target point A. Furthermore, each path node can only be used by one aircraft. Based on calculations, the taxiing time weight for aircraft at position 2 to reach node A is 206 seconds, the waiting taxiing time weight for the second aircraft is 166 seconds, and the taxiing time weight for aircraft at position 3 to reach node A is 183 seconds. Therefore, the arrival time of the two aircraft at the target node A is 14:26:49, totaling 389 seconds of taxiing time weight; the number of nodes used is 21.

[0096] Method 3 considers a path planning method to avoid conflicts between taxiway intersections and taxiway usage nodes. This method ensures that only one aircraft can occupy a taxiway node at a time. The specific planning method is as follows: Figure 13 As shown, 14 seconds after the first aircraft enters node D, the taxiing path for the second aircraft is established. The cyan nodes in the diagram represent the taxiing path established by the second aircraft, and the red nodes represent the node positions of the two aircraft. Figure 13 The blue nodes represent the taxiing paths of the first aircraft before it arrives. The taxiing paths of the two aircraft are shown below: The taxiing path of the first aircraft when the second aircraft begins taxiing is as follows: The taxiing path of the second aircraft is as follows: The calculations require that the taxiing path of the second aircraft to reach point B at C11 should not conflict with the taxiing path of the first aircraft. Therefore, it is necessary for the first aircraft to reach point A and for node A to be released when the second aircraft reaches point B. The departure time of the second aircraft is 14:21:00, requiring a wait of 39 seconds. Of this, 65 seconds is for the first aircraft to fully reach point D, and 14 seconds is for the first aircraft to reach point A and for node A to be fully released. The two aircraft arrive at the target node A at 14:25:01, sharing a total taxiing time weight of 281 seconds; 24 nodes are used.

[0097] The study presented three path planning methods: a dynamic directed graph combined with the DQN algorithm (Method 1), a traditional aircraft scheduling method (Method 2), and a method that considers avoiding taxiway intersections and node conflicts (Method 3). Experimental results show that Method 1 performs best in both taxiing time weight and node usage, with the lowest taxiing time weight (223 seconds) and the fewest nodes (12), achieving the optimal balance between efficiency and resource consumption in airport surface path planning. Specifically, Method 1 saves 42.67% of taxiing time weight compared to Method 2 and 20.64% compared to Method 3, significantly improving operational efficiency.

[0098] In contrast, Method 2, a traditional aircraft scheduling method, has a taxiing time weight as high as 389 seconds (approximately 1.7 times that of Method 1) and uses 21 nodes, resulting in low overall efficiency and failing to fully utilize airport resources. While Method 3 performs moderately in terms of taxiing time weight (281 seconds), it uses the most nodes (24), indicating significant room for optimization in resource utilization and requiring further adjustments to improve overall performance. In conclusion, the path planning method combining dynamic directed graphs and the DQN algorithm outperforms the other two methods in terms of both taxiing efficiency and resource consumption, providing a better solution for airport surface path planning.

[0099] like Figure 14 As shown, the dynamic directed graph combined with the DQN algorithm path planning method is designed and mainly consists of three parts: a data acquisition and preprocessing module, a web server module, and a path planning client module. The design goals and functions of each module are as follows: 1. Data Acquisition and Preprocessing Module: Function Description: This module is responsible for receiving real-time MLAT data from aircraft on the airport ground, including information such as the aircraft's position, speed, and heading on the surface. After receiving the data, it performs noise reduction, format standardization, and status synchronization processing on the raw data to ensure data integrity and timeliness; it also provides high-precision airport surface geographic coordinate input for path planning.

[0100] Implementation: MLAT signals are collected in real time using data receiving equipment deployed at the airport. The received data is preprocessed. During simulation, the processed MLAT data is transmitted to the web server via a Python program on the client side.

[0101] 2. Web server module: Function Description: The web server serves as the core data processing and transmission hub of the system, dynamically updating the state of nodes in the directed graph of the airport scene. It is responsible for receiving pre-processed data uploaded by clients and using the received data to release the nodes in the directed graph G of the airport scene stored on the web server. , and The data is updated by changing nodes in graph G that were initially set to 1 to 0 (available), effectively releasing nodes from the directed graph G. If a path planning request is received, the current state of the directed graph G is sent to the client's DQN algorithm for path planning. After the client's DQN completes path planning and the path planning criteria are met, the planned taxiway path nodes are sent to the web server. The web server then activates the nodes in the airport surface directed graph G based on the received taxiway path nodes, effectively activating the nodes received from the database. or The data changes from a 0 to an active state (1). In short, it receives path planning requests, sends the current directed graph G to the client, and synchronizes the planning results.

[0102] Implementation: A web server is built using Python in PyCharm. The server receives and sends data based on a standard RESTful API interface. The server's internal processing logic includes data reception, dynamic directed graph state updates, bidirectional communication with clients, and the release and activation of node resources.

[0103] 3. DQN route planning client module: Function Description: The DQN path planning client receives a real-time directed graph G of the airport scene from the server by calling a pre-trained DQN algorithm model, performs path optimization planning, and transmits the taxiing paths that meet the criteria to the web server. The planning results must meet the following criteria: (1) If the planned path meets the redundancy time interval requirement, then output the suggested taxiing route; (2) If resource conflicts exist, calculate the expected waiting time and provide an optimization solution; Once the planning is complete, the client will transmit the qualified glide path to the web server to update the node status and allocate resources.

[0104] Implementation: A deep Q-network is implemented based on PyTorch, integrating Priority Experience Replay (PER) and Double DQN techniques to make decisions using input flight status data and airport environment conditions. The route planning client keeps synchronized with the server each time a route is planned to ensure that the planning results are consistent with the actual situation. The specific planning process is as follows: (1) Receive the current directed graph G sent by the server; (2) Calculate the action value using the Q-network and select the optimal path; (3) Execution path conflict detection and redundancy time verification; (4) The time required to return the planning results or wait for the node to be released before proceeding with path planning.

Claims

1. An airport route planning method based on dynamic directed graphs and the DQN algorithm, characterized in that, include: S1. Collect airport surface data and aircraft data within the airport surface and preprocess them to obtain real-time updated surface directed graph node distribution data and node occupancy data. S2. Based on the real-time updated field directed graph node distribution data and node occupancy data in S1, the DQN algorithm model is used to plan the path of the aircraft from the current node to the target node. S3. Calculate the time range of each intermediate node occupied by the aircraft based on the intermediate nodes required for the aircraft to travel from the current node to the target node in S2. S4. Optimize the later planned path or increase the waiting time of the later planned path based on the time range occupied by each intermediate node in S3.

2. The airport route planning method based on dynamic directed graph and DQN algorithm as described in claim 1, characterized in that, In S1, when collecting and preprocessing airport surface data, after obtaining the two-dimensional planar structure diagram of the airport's taxiways, the taxiways are segmented and numbered through the intersections. Each taxiway segment is abstracted as a node in a directed graph, thus obtaining the distribution data of the directed graph nodes of the surface.

3. The airport route planning method based on dynamic directed graph and DQN algorithm as described in claim 2, characterized in that, In step S1, when collecting and preprocessing data of aircraft within the airport surface, the MLAT latitude and longitude data of the aircraft are collected in real time, the nodes occupied by the aircraft are confirmed, and the data is updated in real time in the directed graph node distribution data of the surface, thus obtaining the real-time updated directed graph node distribution data and node occupancy data.

4. The airport route planning method based on dynamic directed graph and DQN algorithm as described in claim 3, characterized in that, In S2, when planning the path of an aircraft from the current node to the target node, the node occupancy status of the current node occupied by the aircraft and the node usage status that the aircraft is preparing to use during taxiing are updated in real time.

5. The airport route planning method based on dynamic directed graph and DQN algorithm as described in claim 4, characterized in that, The usage status of nodes to be used during the aircraft's taxiing process can be planned by two paths simultaneously.

6. The airport route planning method based on dynamic directed graph and DQN algorithm as described in claim 5, characterized in that, The network structure formula of the DQN algorithm model is as follows: in, It is the action-value function, representing the state. Take action below The "value" (usually expected return) that can be obtained at that time; It is a state-value function, representing the state. Its inherent "value". It is the dominance function, used to measure the dominance in a state. Next, specific actions In DQN, the network outputs the degree of good or bad relative to the average level of this state. and Then, by combining the above formula (4.13), we can form... Indicates all possible actions Summing the dominant function; It represents the number of possible actions in the action space.

7. The airport route planning method based on dynamic directed graph and DQN algorithm as described in claim 5, characterized in that, In S3, when calculating the time range of the aircraft occupying each intermediate node, the time when the aircraft's nose enters node a and the time when the aircraft's tail leaves node a are calculated based on the taxiway length of node a and the aircraft's taxiing speed at node a, thus obtaining the time range of the aircraft occupying node a.

8. The airport route planning method based on dynamic directed graph and DQN algorithm as described in claim 7, characterized in that, Node a includes a taxiway node and an intersection node. When an aircraft is at the intersection node, its taxiing speed is 15-30 km / h, and when an aircraft is at the taxiway node, its taxiing speed is 40-60 km / h.

9. The airport route planning method based on dynamic directed graph and DQN algorithm as described in claim 8, characterized in that, When optimizing the later planned path in S4, it is necessary to ensure that the time interval between the earlier and later paths occupying the same node is greater than 9 seconds. That is, the time interval of the first path occupying node a is (16:30:10-16:30:40), and the time interval of the second path occupying node a should be outside the time interval (16:30:01-16:30:49).

10. An airport route planning system based on dynamic directed graphs and the DQN algorithm, characterized in that, include: A data acquisition and preprocessing module is used to receive real-time MLAT data from airport ground aircraft. The DQN path planning client module is used to call the DQN algorithm model and perform path planning based on the real-time updated directed graph node distribution data and node occupancy data. The Web server module is used to update the node distribution data and node occupancy data of the directed graph in real time. It is responsible for receiving the preprocessed data uploaded by the data acquisition and preprocessing module and the path planning data uploaded by the DQN path planning client module, and using the received data to update the node distribution data and node occupancy data of the directed graph in real time.

Images