Ai-based route optimization and flow management method and system

Abstract

The embodiment of the application relates to the technical field of low-altitude unmanned aerial vehicle management, and specifically discloses a route optimization and traffic management method and system based on AI. The embodiment of the application divides a plurality of regular three-dimensional grid units, constructs a double-adjacency matrix as the input of a space-time graph neural network, predicts the traffic situation in a future time window through a cloud server at a macro level, and generates a 4D trajectory route of each unmanned aerial vehicle by using a multi-objective particle swarm optimization algorithm at an edge computing node at a micro level. The edge computing node feeds back an intention, and triggers a re-planning process when an abnormal state is detected by an on-board sensor. The future traffic density of the airspace grid can be accurately predicted, the airspace capacity can be dynamically adjusted based on the prediction, traffic control can be actively performed before congestion occurs, the safety margin of airspace operation can be ensured, and core technical support is provided for realizing a green, efficient and sustainable low-altitude three-dimensional traffic network.

Description

Technical Field

[0001] This invention belongs to the field of low-altitude unmanned aerial vehicle (UAV) management technology, and particularly relates to AI-based route optimization and traffic management methods and systems. Background Technology

[0002] In low-altitude unmanned aerial vehicle (UAV) management, spatiotemporal traffic prediction and capacity management are the core foundations of macro-level scheduling. However, with the increasing frequency of low-altitude flight activities, traditional rule-based static airspace management models are insufficient to cope with sudden large-scale traffic flows. Multi-objective route planning technology aims to solve path optimization problems under complex constraints, especially balancing strategies when multiple conflicting objectives such as energy consumption, noise control, and dynamic weather are involved. Traditional route planning often only pursues the shortest distance or the shortest time, ignoring the impact of UAV flight on the ground acoustic environment and the nonlinear loss of battery life due to severe weather. This patent constructs a comprehensive cost function that includes a noise propagation model, a wind resistance energy consumption model, and a meteorological safety factor, and uses intelligent optimization algorithms to find Pareto optimal solutions among multiple objectives, ensuring that UAVs balance operational efficiency, safety, and environmental friendliness when performing missions.

[0003] Currently, flight path planning and traffic management systems for low-altitude unmanned aerial vehicles (UAVs) mainly employ three typical technical approaches: static management based on fixed airspace structures, single-objective path planning based on geometric rules, and reactive conflict resolution strategies. While these solutions can maintain basic flight order under low-density, simple weather conditions, they still reveal several technical bottlenecks when facing high-frequency, complex weather, and multi-constraint dynamic scenarios. Specifically:

[0004] Static management schemes based on fixed airspace structures rely on pre-defined isolated airspaces or "air corridors" for task allocation. They are typically used for logistics trunk lines with open terrain and simple environments. Although such methods have clear management rules and low deployment costs, they lack the ability to dynamically scale up and down airspace capacity. When a region experiences a sudden large-scale logistics demand or encounters severe local weather, the rigid airspace structure cannot flexibly adjust route resources, which can easily lead to local sector congestion, large-scale flight delays, and severe uneven utilization of airspace resources.

[0005] Single-objective path planning based on geometric rules focuses on finding the shortest geometric path or the fastest time path between the start and end points, and is widely used in A / B algorithm development. Classic algorithms such as Dijkstra's algorithm are computationally efficient in static environments, but they often neglect the crucial multiphysics constraints in low-altitude flight. Most existing algorithms do not include the additional energy consumption caused by dynamic wind fields, noise pollution generated when flying over sensitive areas, and sudden weather conditions in the cost function. As a result, although the generated route has the shortest theoretical distance, it may run out of power due to headwinds or cause complaints due to noise pollution in actual execution. It lacks a comprehensive balance of multidimensional objectives.

[0006] Reactive conflict resolution strategies mainly rely on central dispatch systems or airborne collision avoidance systems to intervene when flight conflicts are about to occur. They typically adopt a "first-come, first-served" or simple horizontal circling avoidance rules. This "passive response" mechanism lacks the ability to predict future traffic conditions. When traffic flow density is high, frequent temporary avoidance not only significantly increases the energy consumption of drones, but may also trigger a chain reaction of congestion, resulting in a significant decrease in the traffic efficiency of the entire low-altitude network. Summary of the Invention

[0007] The purpose of this invention is to provide an AI-based method and system for route optimization and traffic management, aiming to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0009] An AI-based route optimization and traffic management method, the method specifically includes the following steps:

[0010] Three-dimensional voxel modeling is performed to divide the urban low-altitude area into multiple regular three-dimensional grid units, and a dual adjacency matrix is ​​constructed as the input of the spatiotemporal graph neural network, including the physical adjacency matrix and the functional adjacency matrix.

[0011] At the macro level, the spatiotemporal graph neural network is run on the cloud server to acquire and process spatiotemporal data across the entire domain, predict traffic conditions within future time windows, and generate and distribute airspace grid access capacity threshold maps.

[0012] At the micro level, edge computing nodes receive the airspace grid admission capacity threshold map and use a multi-objective particle swarm optimization algorithm to generate a 4D trajectory route for each UAV.

[0013] Edge computing nodes provide intent feedback, transmitting multiple 4D trajectory routes back to the spatiotemporal database of the cloud server in real time, and triggering replanning processing when airborne sensors detect abnormal states.

[0014] As a further limitation of the technical solution of this embodiment of the invention, the step of performing three-dimensional voxel modeling, dividing the urban low-altitude area into multiple regular three-dimensional grid units, and constructing a dual adjacency matrix as the input of the spatiotemporal graph neural network, including a physical adjacency matrix and a functional adjacency matrix, specifically includes the following steps:

[0015] The urban low-altitude area is identified and divided into multiple regular three-dimensional grid units. Within each three-dimensional grid unit, a multi-dimensional feature tensor is maintained.

[0016] A dual adjacency matrix is ​​constructed as the input to the spatiotemporal graph neural network. The dual adjacency matrix includes a physical adjacency matrix and a functional adjacency matrix.

[0017] As a further limitation of the technical solution of the embodiment of the present invention, the multidimensional feature tensor includes several key physical quantities, including flow density, average velocity field and environmental impedance, wherein, flow density is the number of aircraft located in the three-dimensional grid cell at the current moment; average velocity field is the mean value of the velocity vectors of all aircraft in the three-dimensional grid cell, which characterizes the microscopic kinetic energy of traffic flow; environmental impedance is a scalar value that combines wind speed drag, visibility restriction and no-fly zone marking.

[0018] As a further limitation of the technical solution of the present invention, the physical adjacency matrix is ​​constructed based on Euclidean distance and is used to capture the physical diffusion and congestion spread of traffic flow; the functional adjacency matrix is ​​constructed based on historical OD flow statistics, and if two three-dimensional grid cells are physically far apart but have a direct route connection, then a connection is established.

[0019] As a further limitation of the technical solution of this invention, the steps of acquiring and processing full-domain spatiotemporal data, predicting traffic conditions within a future time window, and generating and distributing an airspace grid access capacity threshold map specifically include the following steps:

[0020] In the spatial dimension, historical traffic flow data is acquired, spatial features are extracted, a four-dimensional tensor is constructed, and input into the spatiotemporal graph neural network to output a high-dimensional spatial embedding vector.

[0021] In the time dimension, a multi-head temporal attention mechanism and a gated loop unit are introduced in series to perform time evolution modeling and prediction output, and obtain the congestion probability of multiple three-dimensional grid units.

[0022] Based on the high-dimensional spatial embedding vector and multiple congestion probabilities, proactive flow control is implemented to establish a nonlinear mapping relationship between the airspace grid access capacity threshold and risk factors, and an airspace grid access capacity threshold map is generated and distributed.

[0023] As a further limitation of the technical solution of this embodiment of the invention, the expression for spatial feature extraction is:

[0024] ;

[0025] in, For the first The feature matrix output by the layered graph neural network, For the first The feature matrix input to the layer, This is an activation function used to introduce non-linear expressive power. This is the scaled Laplace matrix. for Chebyshev polynomials For the first The first in the layer The learnable convolutional kernel weight parameters corresponding to the Chebyshev terms;

[0026] The expression for the nonlinear mapping relationship between the spatial grid access capacity threshold and the risk factor is as follows:

[0027] ;

[0028] in, For the index number of the 3D mesh unit, for time The dynamic admission capacity threshold of a 3D mesh cell. For physical limit capacity, for time Weather severity index of three-dimensional grid unit This is the sensitivity coefficient. The weighting coefficients for congestion risk factors. These are the weighting coefficients for meteorological risk factors. for time Predicted congestion probability using 3D grid cells.

[0029] As a further limitation of the technical solution of this embodiment of the invention, the edge computing node receives the airspace grid admission capacity threshold map and uses a multi-objective particle swarm optimization algorithm to generate a 4D trajectory route for each UAV, specifically including the following steps:

[0030] Edge computing nodes receive the spatial grid admission capacity threshold map;

[0031] Substitute the dynamic wind farm energy consumption model to automatically search for the optimal energy efficiency path;

[0032] Substitute the population-weighted noise propagation model to plan vertical lift or horizontal bypass strategies;

[0033] A turning point selection mechanism is introduced to automatically select the equilibrium point as the final execution route.

[0034] Generate a 4D trajectory for each drone.

[0035] As a further limitation of the technical solution of this embodiment of the invention, the expression of the dynamic wind farm energy consumption model is:

[0036] ;

[0037] in, This refers to the total energy consumption during the drone's flight. The total duration of the flight mission. The basic power required for a drone to maintain a near-stationary hovering state. The characteristic power coefficient of the rotor propulsion system, The rotational linear velocity at the rotor tip, The speed of the drone relative to the ground. The angle between the ground speed vector and the wind direction vector. This refers to the real-time wind speed in the current environment. It is a time variable;

[0038] The expression for the population-weighted noise propagation model is:

[0039] ;

[0040] in, This is a weighted index for evaluating noise pollution generated by air routes on the ground. This is the path point index after the route is discretized. This represents the total number of waypoints included in a single route. Weighted by ground population density, path point Horizontal geographic coordinates projected on the ground The preset smoothing factor, The reference sound pressure level for the UAV's onboard sound source. For flight altitude, is the absorption coefficient of air for sound waves. path point The straight-line propagation distance from the affected ground area.

[0041] As a further limitation of the technical solution of this embodiment of the invention, the edge computing node performs intent feedback, transmits multiple 4D trajectory routes back to the spatiotemporal database of the cloud server in real time, and triggers replanning processing when the airborne sensors detect an abnormal state, specifically including the following steps:

[0042] Edge computing nodes provide intent feedback and transmit multiple 4D trajectory routes back to the spatiotemporal database of the cloud server in real time, which are then overlaid on the traffic map as future occupied status.

[0043] When the airborne sensors detect sudden gusts of wind shear or abnormal conditions such as unreported obstacles, a perception fusion algorithm based on a dynamic graph transformer is triggered to perform replanning. The sampling distribution is guided by a reinforcement learning policy network to quickly generate an avoidance path.

[0044] An AI-based route optimization and traffic management method and system for executing the AI-based route optimization and traffic management method described above, the system comprising a cloud server, multiple edge computing nodes, and multiple airborne sensors, wherein:

[0045] The cloud server is used for 3D voxel modeling, which divides the urban low-altitude area into multiple regular three-dimensional grid units and constructs a dual adjacency matrix as the input of the spatiotemporal graph neural network, including a physical adjacency matrix and a functional adjacency matrix. At the macro level, the spatiotemporal graph neural network is run to acquire and process spatiotemporal data across the entire domain, predict traffic conditions within future time windows, and generate and distribute airspace grid access capacity threshold maps.

[0046] Multiple airborne sensors are used for abnormal condition detection;

[0047] Multiple edge computing nodes are used at the micro level to receive the airspace grid access capacity threshold map, generate 4D trajectory routes for each UAV using a multi-objective particle swarm optimization algorithm, provide intent feedback, transmit multiple 4D trajectory routes back to the spatiotemporal database of the cloud server in real time, and trigger replanning processing when the airborne sensors detect an abnormal state.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] (1) This invention achieves significant breakthroughs in flow prediction accuracy, multi-objective decision-making efficiency and system closed-loop response capability through deep coupling of macro-flow control and micro-planning;

[0050] (2) This invention introduces a spatiotemporal graph neural network with dual adjacency matrices, which can not only capture the local diffusion effect of traffic flow through the physical adjacency matrix, but also accurately identify the long-distance dependencies that are not geographically adjacent but have strong OD (origin and destination) associations using the functional adjacency matrix, significantly improving the accuracy of long-term traffic flow prediction for complex urban road networks. Furthermore, the established dynamic capacity adjustment mechanism based on risk factors can transform the predicted congestion probability into the pre-emptive airspace access threshold, achieving a qualitative leap from passive diversion after congestion to proactive load balancing based on prediction.

[0051] (3) This invention constructs a dynamic wind field energy consumption model and a population-weighted noise propagation model. It uses the cubic relationship of wind resistance energy consumption to automatically guide the UAV to take advantage of the downwind airflow and uses the population-weighted acoustic potential field to actively avoid sensitive areas. In addition, it introduces an automatic Pareto inflection point selection strategy in multi-objective particle swarm optimization, which can automatically lock the optimal compromise between energy consumption and noise in milliseconds. It solves the engineering pain point of traditional multi-objective programming where the solution set is good but the decision is difficult. It realizes fully automatic path generation that takes into account flight safety, energy efficiency and social environmental friendliness.

[0052] (4) This invention utilizes the high-performance hardware acceleration capability of edge computing nodes to perform millisecond-level local replanning and transmits the generated deterministic flight path back to the cloud spatiotemporal database in real time. Through the shared feedback mechanism, the accumulation of prediction errors based solely on historical data is eliminated, enabling the macro-prediction model to perceive the future actions of micro-individuals and ensuring the overall robustness and control accuracy of the system under large-scale, highly dynamic and sudden weather interference. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention.

[0054] Figure 1 A flowchart of the method provided by an embodiment of the present invention is shown.

[0055] Figure 2 An application architecture diagram of the system provided in an embodiment of the present invention is shown. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0057] It is understandable that in the existing technology, the route planning and traffic management system for low-altitude UAV operation mainly adopts three typical technical paths: static management based on fixed airspace structure, single-objective path planning based on geometric rules, and reactive conflict resolution strategy. Although these schemes can maintain basic flight order under low-density and simple weather conditions, they still expose several technical bottlenecks when facing dynamic scenarios with high frequency, complex weather and multiple constraints. Specifically, (1) the static management scheme based on fixed airspace structure relies on pre-defined isolated airspace or "air corridors" for task allocation. It is usually used for logistics trunk lines with open terrain and simple environment. Although this method has clear management rules and low deployment costs, it lacks the ability to dynamically scale up and down the airspace capacity. When a large-scale logistics demand suddenly occurs in a certain area or when local severe weather occurs, the rigid airspace structure cannot flexibly adjust the route resources, which can easily lead to local sector congestion, a large number of flight delays, and serious uneven utilization of airspace resources; (2) the single-objective path planning based on geometric rules focuses on finding the shortest geometric path or the fastest time path between the origin and destination. It widely uses A Classic algorithms such as Dijkstra are highly efficient in static environments, but they often ignore the crucial multiphysics constraints in low-altitude flight. Most existing algorithms do not include the additional energy consumption caused by dynamic wind fields, noise pollution caused by flying over sensitive areas, and sudden weather conditions in the cost function. As a result, although the generated route has the shortest theoretical distance, it may run out of power due to headwinds or cause complaints due to noise pollution in actual execution. It lacks a comprehensive balance of multidimensional objectives. (3) Reactive conflict resolution strategy mainly relies on the central scheduling system or airborne collision avoidance system to intervene when a flight conflict is about to occur. It usually adopts the "first-come, first-served" or simple horizontal circling avoidance rules. This "passive response" mechanism lacks the ability to predict future traffic conditions. When the traffic flow density is high, frequent temporary avoidance not only greatly increases the energy consumption of the UAV, but may also trigger a chain of congestion effects, resulting in a significant decrease in the traffic efficiency of the entire low-altitude network.

[0058] To address the aforementioned issues, this invention employs 3D voxel modeling to divide urban low-altitude areas into multiple regular 3D grid units. A dual adjacency matrix, comprising a physical adjacency matrix and a functional adjacency matrix, is constructed as input to a spatiotemporal graph neural network. At the macroscopic level, the spatiotemporal graph neural network, running on a cloud server, acquires and processes global spatiotemporal data, predicts traffic conditions within future time windows, and generates and distributes an airspace grid access capacity threshold map. At the microscopic level, edge computing nodes receive the airspace grid access capacity threshold map and utilize a multi-objective particle swarm optimization algorithm to generate 4D trajectory routes for each UAV. These edge computing nodes provide intent feedback, transmitting multiple 4D trajectory routes back to the spatiotemporal database of the cloud server in real time. Furthermore, when airborne sensors detect abnormal conditions, replanning is triggered. This approach enables accurate prediction of future traffic density within the airspace grid, achieving predictive dynamic adjustment of airspace capacity. This proactively controls traffic flow before congestion occurs, ensuring a safety margin for airspace operation and providing core technological support for realizing a green, efficient, and sustainable low-altitude 3D transportation network.

[0059] Figure 1 A flowchart of the method provided by an embodiment of the present invention is shown.

[0060] Specifically, the AI-based route optimization and traffic management method includes the following steps:

[0061] Step 1: Perform 3D voxel modeling, divide the urban low-altitude area into multiple regular 3D grid units, and construct a dual adjacency matrix as the input to the spatiotemporal graph neural network, including the physical adjacency matrix and the functional adjacency matrix.

[0062] In this embodiment of the invention, a low-altitude urban area (e.g., 50m to 1000m above the ground) is determined and divided into multiple regular three-dimensional grid units. Within each three-dimensional grid unit, a multidimensional feature tensor is maintained. The multidimensional feature tensor contains several key physical quantities, including traffic density, average velocity field, and environmental impedance. Here, traffic density is the number of aircraft currently located in the three-dimensional grid unit; average velocity field is the average of the velocity vectors of all aircraft in the three-dimensional grid unit, representing the microscopic kinetic energy of traffic flow; and environmental impedance is a scalar value that integrates wind speed drag, visibility restrictions, and no-fly zone markings. A dual adjacency matrix is ​​constructed as the input to the spatiotemporal graph neural network. The dual adjacency matrix includes a physical adjacency matrix and a functional adjacency matrix. The physical adjacency matrix is ​​constructed based on Euclidean distance and is used to capture the physical diffusion and congestion spread of traffic flow. The functional adjacency matrix is ​​constructed based on historical OD flow statistics. If two three-dimensional grid units are physically far apart but have a direct flight path connection, a connection is established.

[0063] Understandably, in the process of dividing the 3D grid cells, a hierarchical indexing mechanism is used to map the real-time latitude, longitude, and altitude coordinates (Lon, Lat, Alt) of the UAV into a discretized grid index. .

[0064] It is understandable that the physical adjacency matrix ( ): Constructed based on Euclidean distance. If two grids are spatially congruent (including Moore's neighborhood), then This matrix is ​​used to capture the physical spread of traffic flow and congestion propagation; functional adjacency matrix ( ): Constructed based on historical OD (Origin-Destination) flow statistics. If two grids are physically far apart but have frequent direct route connections (such as a logistics hub and a specific distribution station), a connection is established. Its weight... Determined by historical traffic transfer probabilities, the model can "understand" the start and end logic of flight missions, thereby improving the accuracy of long-term predictions.

[0065] Step 2: At the macro level, the spatiotemporal graph neural network is run on the cloud server to acquire and process spatiotemporal data across the entire domain, predict traffic conditions within the future time window, and generate and distribute an airspace grid access capacity threshold map.

[0066] In this embodiment of the invention, in the spatial dimension, historical traffic flow data is acquired, spatial features are extracted, a four-dimensional tensor is constructed, and input into a spatiotemporal graph neural network to output a high-dimensional spatial embedding vector. In the temporal dimension, a multi-head temporal attention mechanism and a gated recurrent unit are introduced in series to perform temporal evolution modeling and prediction output, obtaining the congestion probability of multiple three-dimensional grid units. Then, based on the high-dimensional spatial embedding vector and multiple congestion probabilities, proactive flow control is executed to establish a nonlinear mapping relationship between the airspace grid access capacity threshold and risk factors, generating and distributing an airspace grid access capacity threshold map. Specifically, the expression for spatial feature extraction is:

[0067] ;

[0068] in, For the first The feature matrix output by the layered graph neural network, For the first The feature matrix of the layer input, This is an activation function used to introduce non-linear expressive power. This is the scaled Laplace matrix. for Chebyshev polynomials For the first The first in the layer The learnable convolutional kernel weight parameters corresponding to the Chebyshev terms;

[0069] The expression for the nonlinear mapping relationship between the airspace grid access capacity threshold and the risk factor is as follows:

[0070] ;

[0071] in, For the index number of the 3D mesh unit, for time The dynamic admission capacity threshold of a 3D mesh cell. For physical limit capacity, for time Weather severity index of three-dimensional grid unit This is the sensitivity coefficient. The weighting coefficients for congestion risk factors. These are the weighting coefficients for meteorological risk factors. for time Predicted congestion probability using 3D grid cells.

[0072] Understandably, in the time dimension, by introducing a series structure of Temporal Multi-Head Attention and Gated Recurrent Unit (GRU), the attention layer is responsible for capturing long-distance time dependencies (such as the lagged impact of sudden weather on subsequent traffic flow) and calculating the dynamic weight matrix of historical moments for the current prediction; the GRU layer is used to fit the short-term fluctuations and periodic trends of traffic flow, and the output layer maps the high-dimensional hidden state to the predicted traffic density values ​​of each grid in the future and calculates the congestion probability of the grid.

[0073] It is understandable that establishing a nonlinear mapping relationship between the airspace grid access capacity threshold and risk factors can drive an exponential decrease in the airspace grid access capacity threshold, thereby "squeezing out" excess demand during the planning stage and realizing a paradigm shift from "passive flow restriction" to "pre-emptive smoothing".

[0074] Step 3: At the micro level, the edge computing node receives the airspace grid access capacity threshold map and uses a multi-objective particle swarm optimization algorithm to generate a 4D trajectory route for each UAV.

[0075] In this embodiment of the invention, the edge computing node receives a spatial grid access capacity threshold map, then substitutes it into a dynamic wind field energy consumption model to automatically search for the energy-efficient path. It also substitutes it into a population-weighted noise propagation model to plan a vertical lift or horizontal bypass strategy. Subsequently, an inflection point selection mechanism is introduced to automatically select an equilibrium point as the final execution route, generating a 4D trajectory for each UAV. Specifically, the expression for the dynamic wind field energy consumption model is:

[0076] ;

[0077] in, This refers to the total energy consumption during the drone's flight. The total duration of the flight mission. The basic power required for a drone to maintain a near-stationary hovering state. The characteristic power coefficient of the rotor propulsion system, The rotational linear velocity at the rotor tip, The speed of the drone relative to the ground. The angle between the ground speed vector and the wind direction vector. This refers to the real-time wind speed in the current environment. It is a time variable;

[0078] The expression for the population-weighted noise propagation model is:

[0079] ;

[0080] in, This is a weighted index for evaluating noise pollution generated by air routes on the ground. This is the path point index after the route is discretized. This represents the total number of waypoints included in a single route. Weighted by ground population density, path point Horizontal geographic coordinates projected on the ground The preset smoothing factor, The reference sound pressure level for the UAV's onboard sound source. For flight altitude, is the absorption coefficient of air for sound waves. path point The straight-line propagation distance from the affected ground area.

[0081] Understandably, the cubic term in the dynamic wind field energy consumption model accurately reflects the cubic relationship between air resistance power and relative speed. This means that the cost of flying against the wind is drastically amplified. The algorithm will use this model to automatically search for "tailwind" paths, guiding the drone to utilize favorable airflow in the urban canyon effect, thereby achieving optimal energy efficiency at the physical level.

[0082] Understandably, population-weighted noise propagation models, combining sound wave geometric diffusion, atmospheric absorption principles, and ground population heat maps, can generate extremely high potential energy above schools, hospitals, or high-density residential areas. This forces particles (flight paths) in the MOPSO algorithm to automatically perform vertical lift (increase noise level) when passing through these areas. To utilize distance decay or horizontal flight strategies.

[0083] Understandably, in the inflection point selection mechanism, since energy consumption and noise targets are usually in conflict, the Pareto inflection point (KneePoint) automated decision-making outputs a set of non-dominated solutions. In order to achieve millisecond-level automated decision-making, an inflection point selection mechanism is introduced to calculate the marginal rate of substitution of each solution on the Pareto front in the normalized target space and automatically select the balance point that "exchanges the minimum energy consumption cost for the maximum noise reduction benefit" as the final execution route, thus avoiding the delay of manual selection.

[0084] Step 4: The edge computing node provides intent feedback, transmitting multiple 4D trajectory routes back to the spatiotemporal database of the cloud server in real time, and triggering replanning processing when the airborne sensors detect an abnormal state.

[0085] In this embodiment of the invention, edge computing nodes provide intent feedback, transmitting multiple 4D trajectory routes back to the spatiotemporal database of the cloud server in real time. These routes are then overlaid on the traffic flow map as future occupied states. Furthermore, when airborne sensors detect sudden gusts of wind shear or abnormal states such as unreported obstacles, a perception fusion algorithm based on a dynamic graph transformer is triggered for replanning. A reinforcement learning strategy network guides the sampling distribution, quickly generating avoidance paths to ensure robustness in extremely dynamic environments.

[0086] Furthermore, Figure 2 An application architecture diagram of the system provided in an embodiment of the present invention is shown.

[0087] In another preferred embodiment of the present invention, the AI-based route optimization and traffic management method and system includes:

[0088] The cloud server is used for 3D voxel modeling, which divides the urban low-altitude area into multiple regular three-dimensional grid units and constructs a dual adjacency matrix as the input of the spatiotemporal graph neural network, including a physical adjacency matrix and a functional adjacency matrix. At the macro level, the spatiotemporal graph neural network is run to acquire and process spatiotemporal data across the entire domain, predict traffic conditions within future time windows, and generate and distribute airspace grid access capacity threshold maps.

[0089] Multiple airborne sensors are used for abnormal condition detection;

[0090] Multiple edge computing nodes are used at the micro level to receive the airspace grid access capacity threshold map, generate 4D trajectory routes for each UAV using a multi-objective particle swarm optimization algorithm, provide intent feedback, transmit multiple 4D trajectory routes back to the spatiotemporal database of the cloud server in real time, and trigger replanning processing when the airborne sensors detect an abnormal state.

[0091] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0092] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0093] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0094] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0095] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI-based route optimization and traffic management method, characterized in that, The method specifically includes the following steps: Three-dimensional voxel modeling is performed to divide the urban low-altitude area into multiple regular three-dimensional grid units, and a dual adjacency matrix is ​​constructed as the input of the spatiotemporal graph neural network, including the physical adjacency matrix and the functional adjacency matrix. At the macro level, the spatiotemporal graph neural network is run on the cloud server to acquire and process spatiotemporal data across the entire domain, predict traffic conditions within future time windows, and generate and distribute airspace grid access capacity threshold maps. At the micro level, edge computing nodes receive the airspace grid admission capacity threshold map and use a multi-objective particle swarm optimization algorithm to generate a 4D trajectory route for each UAV. Edge computing nodes provide intent feedback, transmitting multiple 4D trajectory routes back to the spatiotemporal database of the cloud server in real time, and triggering replanning processing when airborne sensors detect abnormal states. The physical adjacency matrix, constructed based on Euclidean distance, is used to capture the physical diffusion of traffic flow and the spread of congestion; the functional adjacency matrix, constructed based on historical OD flow statistics, establishes a connection if two 3D grid cells are physically far apart but have a direct flight path connection. The process of acquiring and processing full-domain spatiotemporal data, predicting traffic conditions within future time windows, and generating and distributing an airspace grid access capacity threshold map specifically includes the following steps: In the spatial dimension, historical traffic flow data is acquired, spatial features are extracted, a four-dimensional tensor is constructed, and input into the spatiotemporal graph neural network to output a high-dimensional spatial embedding vector. In the time dimension, a multi-head temporal attention mechanism and a gated loop unit are introduced in series to perform time evolution modeling and prediction output, and obtain the congestion probability of multiple three-dimensional grid units. Based on the high-dimensional spatial embedding vector and multiple congestion probabilities, active flow control is performed to establish a nonlinear mapping relationship between the airspace grid access capacity threshold and risk factors, and an airspace grid access capacity threshold map is generated and distributed. The edge computing node receives the airspace grid admission capacity threshold map and uses a multi-objective particle swarm optimization algorithm to generate a 4D trajectory for each UAV, specifically including the following steps: Edge computing nodes receive the spatial grid admission capacity threshold map; Substitute the dynamic wind farm energy consumption model to automatically search for the optimal energy efficiency path; Substitute the population-weighted noise propagation model to plan vertical lift or horizontal bypass strategies; A turning point selection mechanism is introduced to automatically select the equilibrium point as the final execution route. Generate a 4D trajectory for each drone; The expression for the dynamic wind farm energy consumption model is: ； in, This refers to the total energy consumption during the drone's flight. The total duration of the flight mission. The basic power required for a drone to maintain a near-stationary hovering state. The characteristic power coefficient of the rotor propulsion system, The linear velocity of the rotor tip is the rotational velocity. The speed of the drone relative to the ground. The angle between the ground speed vector and the wind direction vector. This refers to the real-time wind speed in the current environment. It is a time variable; The expression for the population-weighted noise propagation model is: ； in, This is a weighted index for evaluating noise pollution generated by air routes on the ground. This is the path point index after the route is discretized. This represents the total number of waypoints included in a single route. Weighted by ground population density, path point Horizontal geographic coordinates projected on the ground The preset smoothing factor, The reference sound pressure level for the UAV's onboard sound source. For flight altitude, is the absorption coefficient of air for sound waves. path point The straight-line propagation distance from the affected ground area.

2. The AI-based route optimization and traffic management method according to claim 1, characterized in that, The process of performing three-dimensional voxel modeling, dividing the urban low-altitude area into multiple regular three-dimensional grid units, and constructing a dual adjacency matrix as input to the spatiotemporal graph neural network, including a physical adjacency matrix and a functional adjacency matrix, specifically includes the following steps: The urban low-altitude area is identified and divided into multiple regular three-dimensional grid units. Within each three-dimensional grid unit, a multi-dimensional feature tensor is maintained. A dual adjacency matrix is ​​constructed as the input to the spatiotemporal graph neural network. The dual adjacency matrix includes a physical adjacency matrix and a functional adjacency matrix.

3. The AI-based route optimization and traffic management method according to claim 2, characterized in that, The multidimensional feature tensor contains several key physical quantities, including flow density, average velocity field, and environmental impedance. Flow density is the number of aircraft currently located in the three-dimensional grid cell; average velocity field is the mean of the velocity vectors of all aircraft within the three-dimensional grid cell, representing the microscopic kinetic energy of the traffic flow; and environmental impedance is a scalar value that combines wind speed drag, visibility limitations, and no-fly zone markings.

4. The AI-based route optimization and traffic management method according to claim 1, characterized in that, The expression for spatial feature extraction is: ； in, For the first The feature matrix output by the layered graph neural network, For the first The feature matrix of the layer input, This is an activation function used to introduce non-linear expressive power. This is the scaled Laplace matrix. for Chebyshev polynomials For the first The first in the layer The learnable convolutional kernel weight parameters corresponding to the Chebyshev terms; The expression for the nonlinear mapping relationship between the spatial grid access capacity threshold and the risk factor is as follows: ； in, For the index number of the 3D mesh unit, for time The dynamic admission capacity threshold of a 3D mesh cell. For physical limit capacity, for time Weather severity index of three-dimensional grid unit This is the sensitivity coefficient. The weighting coefficients for congestion risk factors. These are the weighting coefficients for meteorological risk factors. for time Predicted congestion probability using 3D grid cells.

5. The AI-based route optimization and traffic management method according to claim 1, characterized in that, The edge computing node provides intent feedback, transmitting multiple 4D trajectory routes back to the spatiotemporal database of the cloud server in real time. Furthermore, when the onboard sensors detect an abnormal state, the replanning process is triggered, specifically including the following steps: Edge computing nodes provide intent feedback and transmit multiple 4D trajectory routes back to the spatiotemporal database of the cloud server in real time, which are then overlaid on the traffic map as future occupied status. When the airborne sensors detect sudden gusts of wind shear or abnormal conditions such as unreported obstacles, a perception fusion algorithm based on a dynamic graph transformer is triggered to perform replanning. The sampling distribution is guided by a reinforcement learning policy network to quickly generate an avoidance path.

6. An AI-based route optimization and traffic management method and system for executing the AI-based route optimization and traffic management method as described in any one of claims 1-5, characterized in that, The system includes a cloud server, multiple edge computing nodes, and multiple airborne sensors, wherein: The cloud server is used for 3D voxel modeling, which divides the urban low-altitude area into multiple regular three-dimensional grid units and constructs a dual adjacency matrix as the input of the spatiotemporal graph neural network, including a physical adjacency matrix and a functional adjacency matrix. At the macro level, the spatiotemporal graph neural network is run to acquire and process spatiotemporal data across the entire domain, predict traffic conditions within future time windows, and generate and distribute airspace grid access capacity threshold maps. Multiple airborne sensors are used for abnormal condition detection; Multiple edge computing nodes are used at the micro level to receive the airspace grid access capacity threshold map, generate 4D trajectory routes for each UAV using a multi-objective particle swarm optimization algorithm, provide intent feedback, transmit multiple 4D trajectory routes back to the spatiotemporal database of the cloud server in real time, and trigger replanning processing when the airborne sensors detect an abnormal state.

Images