A Dynamic Scheduling Method for Low-Altitude High-Density Airspace Based on Spatiotemporal Voxel Fields and Reinforcement Learning

Abstract

This application relates to low-altitude airspace traffic scheduling technology, and particularly to a dynamic scheduling method for low-altitude high-density airspace based on spatiotemporal voxel fields and reinforcement learning. The method includes: fusing multi-source heterogeneous data to discretize a continuous target low-altitude airspace into an anisotropic spatiotemporal voxel field; using a graph neural network based on the spatiotemporal voxel field to model the interaction relationships between aircraft within the target low-altitude airspace and predict the probabilistic pipelines of each aircraft within future time windows; inputting the current state of the spatiotemporal voxel field and the probabilistic pipelines of each aircraft within future time windows into a multi-agent reinforcement learning model to obtain preliminary instructions; using a safety filter based on a control obstacle function to correct the deviation of the preliminary instructions to obtain safety instructions; and issuing the safety instructions to the flight control systems of each aircraft within the target low-altitude airspace to achieve safe scheduling, autonomous obstacle avoidance, and efficient collaboration of a large-scale aircraft cluster in the target low-altitude airspace.

Description

Technical Field

[0001] The embodiments of this application relate to the field of low-altitude airspace traffic scheduling technology, and in particular to a dynamic scheduling method for low-altitude high-density airspace based on spatiotemporal voxel fields and reinforcement learning. Background Technology

[0002] With the booming development of the low-altitude economy and the increasing prevalence of applications such as urban air mobility (UAM), last-mile logistics delivery, emergency rescue, and smart city aerial inspection, low-altitude airspace is gradually evolving from a sparse and isolated independent operation mode to a high-density and integrated large-scale operation mode. In the foreseeable future, urban low-altitude airspace will face the extreme challenge of hundreds or even thousands of aircraft operating simultaneously per square kilometer. However, current low-altitude airspace traffic management technologies have significant technical bottlenecks in dealing with such high-density, highly dynamic, and highly uncertain flight missions.

[0003] Traditional low-altitude airspace traffic management technology mainly follows the "unified control" or "static isolation" approach of civil aviation. This model typically relies on manual approval of flight plans and divides airspace into fixed flight isolation zones or static routes. While this method can ensure safety under low traffic conditions, its drawbacks become glaringly apparent in high-density scenarios. First, statically designated routes waste significant time and space resources. Once a route is occupied or idle, other aircraft urgently needing passage cannot dynamically reuse that airspace, resulting in extremely low airspace utilization. Second, static routes lack the ability to be reconstructed in real time in the face of sudden weather changes (such as instantaneous wind shear between urban buildings) or intrusions by non-cooperative targets, easily leading to safety incidents. Furthermore, with the exponential growth in flight frequency, centralized scheduling based on manual methods or traditional operations research algorithms (such as mixed integer programming) faces the curse of dimensionality. The computation time increases explosively with the number of aircraft, failing to meet the millisecond-level real-time obstacle avoidance requirements.

[0004] To address these issues, some research teams have attempted to introduce rule-based geometric obstacle avoidance algorithms (such as artificial potential field methods and velocity obstacle methods) or artificial intelligence-based deep reinforcement learning algorithms. While geometric obstacle avoidance algorithms are computationally simple, they are prone to getting trapped in local optima in complex multi-aircraft interaction scenarios, causing the aircraft to repeatedly oscillate in deadlock regions and fail to reach its destination. In contrast, deep reinforcement learning, through trial and error with the environment, can learn more complex cooperative strategies and has become a research hotspot in recent years.

[0005] However, current data-driven reinforcement learning-based low-altitude airspace traffic scheduling methods still face the following serious challenges in practical engineering applications.

[0006] First, safety cannot be guaranteed by mathematical certainty. Neural networks are essentially probabilistic models, which are inherently uninterpretable. In extreme conditions not covered by the training set, the model may produce unpredictable and dangerous outputs (such as suddenly turning and crashing into obstacles), which is unacceptable.

[0007] Second, the granularity of environmental modeling is insufficient. Most existing methods simplify low-altitude airspace to a homogeneous Euclidean space, neglecting the impact of environmental vector fields such as communication signal coverage quality (QoS) and micro-meteorological wind fields on aircraft flight. In urban low-altitude airspace, GPS multipath effects, 5G signal blockage, and building wake turbulence are the main causes of aircraft loss of control and crashes. If the scheduling algorithm does not incorporate these environmental physical fields into the state space, the generated trajectory will be unexecutable or accompanied by an extremely high risk of link failure.

[0008] Third, the coordination mechanism is simplistic. Traditional methods often presuppose that all aircraft are subject to unified scheduling, lacking non-cooperative game mechanisms for different mission priorities (such as medical emergency and general logistics) and different operating entities. This makes it difficult to ensure the fairness and urgency of individual missions while maintaining overall operational efficiency. Summary of the Invention

[0009] In view of this, embodiments of this application propose a dynamic scheduling method for low-altitude, high-density airspace based on spatiotemporal voxel fields and reinforcement learning. It integrates spatiotemporal voxel perception, safety constraints of control obstacle functions, and multi-agent deep reinforcement learning. By constructing a high-fidelity digital twin model of the environment and deterministic safety control theory, it realizes safe scheduling, autonomous obstacle avoidance, and efficient collaboration of large-scale aircraft clusters in complex low-altitude airspace.

[0010] To achieve the above objectives, this application proposes a dynamic scheduling method for low-altitude, high-density airspace based on spatiotemporal voxel fields and reinforcement learning. This method includes: fusing lidar point clouds, signal quality mapping from terrestrial 5G-A communication base stations, and real-time wind field data from weather radar to discretize the continuous target low-altitude airspace into an anisotropic spatiotemporal voxel field; each voxel unit in the spatiotemporal voxel field not only records the spatial occupancy state but also dynamically embeds environmental semantic information including wind vectors, communication signal-to-noise ratio, and potential field gradients; based on the spatiotemporal voxel field, a graph neural network is used to model the interaction relationships between various aircraft within the target low-altitude airspace, treating each aircraft as a node in a graph, and utilizing attention... The force mechanism aggregates the state features of neighboring nodes and predicts the probability pipeline of each aircraft within a future time window through historical trajectory sequences. The current state of the spatiotemporal voxel field and the probability pipeline of each aircraft within the future time window are input into a pre-trained multi-agent reinforcement learning model to obtain the initial instructions output by the multi-agent reinforcement learning model. The deviation of the initial instructions is corrected by a safety filter based on the control obstacle function, and the flight safety constraint is transformed into a quadratic programming optimization problem to obtain the safety instructions output by the safety filter. The safety instructions are then issued to the flight control systems of each aircraft in the target low-altitude airspace to achieve safe scheduling, autonomous obstacle avoidance, and efficient coordination of the aircraft cluster in the target low-altitude airspace.

[0011] To achieve the above objectives, this application also proposes a dynamic scheduling system for low-altitude, high-density airspace based on spatiotemporal voxel fields and reinforcement learning. This system includes: a spatiotemporal voxel field construction module, used to fuse lidar point clouds, signal quality mappings from ground-based 5G-A communication base stations, and real-time wind field data from meteorological radar, discretizing the continuous target low-altitude airspace into anisotropic spatiotemporal voxel fields; wherein each voxel unit in the spatiotemporal voxel field not only records the spatial occupancy state but also dynamically embeds environmental semantic information including wind vectors, communication signal-to-noise ratio, and potential field gradients; and an interaction relationship modeling module, used to model the interaction relationships between various aircraft within the target low-altitude airspace based on the spatiotemporal voxel fields and utilizing graph neural networks, treating the aircraft as a node in a graph and using an attention mechanism to aggregate neighbors. The system employs several modules: a near-node state feature module to predict the probability pipeline of each aircraft within a future time window using historical trajectory sequences; a preliminary command generation module to input the current state of the spatiotemporal voxel field and the probability pipeline of each aircraft within a future time window into a pre-trained multi-agent reinforcement learning model to obtain the preliminary commands output by the multi-agent reinforcement learning model; a deviation correction and safety command generation module to correct the deviation of the preliminary commands using a safety filter based on a control obstacle function, transforming flight safety constraints into a quadratic programming optimization problem to obtain the safety commands output by the safety filter; and a scheduling execution module to issue safety commands to the flight control systems of each aircraft within the target low-altitude airspace, enabling safe scheduling, autonomous obstacle avoidance, and efficient collaboration of the aircraft cluster within the target low-altitude airspace.

[0012] To achieve the above objectives, this application also proposes an electronic device, comprising: a processor and a memory, wherein the memory stores instructions executable by the processor, and the processor is configured to execute the instructions such that the electronic device can implement the low-altitude high-density airspace dynamic scheduling method based on spatiotemporal voxel fields and reinforcement learning as described above.

[0013] To achieve the above objectives, this application also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, enables a dynamic scheduling method for low-altitude, high-density airspace based on spatiotemporal voxel fields and reinforcement learning as described above.

[0014] Optionally, spatiotemporal voxel fields It is a size of A three-dimensional cubic elemental mesh, , , They are respectively Length, width, and height; For any time Location index is voxel units It is defined as the following high-dimensional feature tensor. ; express The occupancy status of static obstacles is pre-loaded using a high-precision map. Static obstacles include buildings and no-fly zones. , This indicates that the space is not occupied. Indicates that the space is occupied; express The real-time wind vector, including wind speed and wind direction angle, is obtained by real-time inversion using weather radar or Doppler lidar. express The real-time communication signal connection quality is calculated by interpolation based on the signal quality mapping of terrestrial 5G-A communication base stations, and is used to assess the risk of link failure for the aircraft. , This indicates that the real-time communication signal connection quality is good. This indicates poor quality of real-time communication signal connection; This represents the repulsive potential energy value calculated based on the artificial potential field method. , for The Euclidean distance between the voxel center and the nearest obstacle. The preset threshold for the range of influence. This is the preset potential field gain coefficient; express In the future time window The probability density estimate of the interior being occupied by aircraft.

[0015] Optionally, based on a spatiotemporal voxel field, a graph neural network is used to model the interaction relationships between various aircraft in the target low-altitude airspace. The aircraft is treated as a node in a graph, and an attention mechanism is used to aggregate the state features of neighboring nodes. The probability pipeline for each aircraft within a future time window is predicted using historical trajectory sequences, including: Construct a spatiotemporal attention map for each aircraft in the target low-altitude airspace. , All aircraft as Node set The interaction between each aircraft and all its neighboring aircraft within its communication range is used as... edge set ; exist At that moment, for the first aircraft Its historical trajectory sequence is , ,in, , , They are respectively exist The three-dimensional position and three-dimensional velocity vector at any given moment. , They are respectively exist Azimuth and elevation angles at any given time; The historical trajectory sequences of each aircraft are encoded using LSTM, and the attention coefficients between the aircraft are calculated based on the state features obtained from the encoding of the historical trajectory sequences of each aircraft using LSTM. Record No. One aircraft With the One aircraft The attention coefficient between them is , The calculation process is expressed by the formula: ; in, for The corresponding state characteristics, for The corresponding state characteristics, The weight matrix is ​​a learnable matrix. The attention vector, with the top-right subscript. This indicates taking the transpose. This indicates that a splicing operation is being performed. Representation layer normalization, Represents the Softmax function; For each aircraft, a graph neural network is used to aggregate, update, and decode based on its state features and its attention coefficients with other aircraft to obtain the future... The position distribution sequence within the time step is used to form a probability pipeline with a confidence interval. The probability pipeline clearly defines the spatiotemporal range of the future existence of the spacecraft and is used to identify potential conflicts. Record No. One aircraft The probability pipeline is , This can be expressed by the formula: ; in, , The first time The position mean and position covariance matrix.

[0016] Optionally, the current state of the spatiotemporal voxel field and the probability pipelines of each aircraft within the future time window are input into a pre-trained multi-agent reinforcement learning model to obtain preliminary instructions output by the multi-agent reinforcement learning model, including: All aircraft are pre-defined as multi-agent agents, and a multi-agent proximal policy optimization algorithm is used as the decision-making core to define the state space. Action space and reward space Train a policy network As a pre-trained multi-agent reinforcement learning model; among which, , , , For policy networks Learnable parameters; yes The state at any given time includes the agent's own state, local perceived environmental features, and the probability pipeline of all neighboring agents; where the agent's own state includes position, velocity, and remaining battery power, and the local perceived environmental features are voxel unit encodings. yes The action at any given moment is a series of continuous control commands, including acceleration vector, azimuth rate of change, and pitch rate of change; yes The reward function at time step is designed as a multi-objective weighted sum, expressed by the formula: ; in, for The distance reward at any given moment increases as you get closer to the target point. for A safety reward is given when the distance between the agent and an obstacle or neighboring agent is less than the safe distance. The safety reward is the safety distance penalty. If an agent enters a signal blind zone, the safety reward is the signal blind zone penalty. , for Momentary comfort rewards are used to punish drastic changes in acceleration. The reward for energy consumption is calculated based on wind field vectors; flying with the wind results in a positive value and flying against the wind results in a negative value and a penalty. , , , They are respectively , , , The corresponding reward weighting coefficient; The multi-agent reinforcement learning model calculates the mean and variance of the action distribution based on the current state of the spatiotemporal voxel field and the probability pipeline of each aircraft within the future time window, and obtains preliminary actions through sampling. As preliminary instructions; among which, the preliminary instructions contain unsafe instructions.

[0017] Optionally, a safety filter based on a control obstacle function is used to correct the deviation of the initial command, transforming the flight safety constraints into a quadratic programming optimization problem, to obtain the safety command output by the safety filter, including: Construct a control barrier function This makes the secure set Defined as , For state variables, For dimension space, Defined as the norm of relative distance, it is expressed as: ; in, for Location, The location of obstacles or neighboring intelligent agents; The dynamic equations of the aircraft are defined as affine nonlinear equations. , and They are two different affine functions. As an action variable, For the updated state variables; To ensure forward invariance, The following Lie derivative constraint inequalities must be satisfied: ; , ; in, yes Li Daoshu, yes Li Daoshu, For extension class functions; based on In accordance with the aforementioned Lie derivative constraint inequality, the safety filter transforms the flight safety constraints into a quadratic programming optimization problem, aiming to find a solution that is consistent with... The new action with minimal deviation and meeting safety constraints. , which serves as the safety command output by the safety filter.

[0018] Optionally, the quadratic programming optimization problem can be expressed by the following formula: ; ; in, The lower bound of the action. This is the upper bound of the action.

[0019] Optionally, if the probability pipelines of multiple aircraft are detected to overlap in the same voxel unit, it is confirmed that the multiple aircraft involved are competing for resources. At this time, the competition process will be triggered to achieve dynamic game conflict resolution. Each participating aircraft determines its corresponding right-of-way valuation based on its own state and mission, denoted as the first... One aircraft The value of the right-of-way is , The calculation formula is: ; in, for Task priority, for The remaining battery power, The preset minimum safe power level, for The cumulative delay time of the task , , These are the weighting coefficients; Based on the estimated road rights of each aircraft involved, the following formula is used to find an optimal road rights allocation scheme that maximizes total social welfare. : ; Among them, total social welfare refers to the sum of benefits obtained by all participating aircraft. Represents all possible right-of-way allocation schemes. In the right-of-way allocation scheme Down The result of the right-of-way allocation, i.e. the order of passage, includes priority passage or circling and waiting; To prevent false bidding, the aircraft that obtains right-of-way needs to pay a virtual cost to compensate for the losses caused to the remaining aircraft. Assuming... The virtual cost that one needs to pay to obtain the right-of-way. for: ; in, Indicates no Participate in the competition, Indicates no When participating in the competition, the remaining aircraft are in the optimal right-of-way allocation scheme The total social welfare obtained by the following express After participating in the competition and obtaining the right-of-way, the remaining aircraft will be in the optimal right-of-way allocation scheme. The total social welfare obtained by the public.

[0020] In summary, the low-altitude high-density airspace dynamic scheduling method proposed in this application, based on spatiotemporal voxel fields and reinforcement learning, has the following significant advantages compared with traditional speed regulation methods.

[0021] First, it achieves deterministic assurance of flight safety. By introducing a control obstacle function as the safety foundation for AI decision-making, this application can forcibly correct the flight trajectory based on safety constraints when multi-agent reinforcement learning models make misjudgments or face unknown environmental interference. This fundamentally solves the uninterpretability and safety hazards of traditional end-to-end reinforcement learning models and meets the high reliability requirements of aviation-grade systems.

[0022] Secondly, it significantly improves the throughput and resource utilization of low-altitude airspace. The application of anisotropic spatiotemporal voxel fields allows for a finer granularity of airspace management from the traditional route level to the grid level. Combined with dynamic game-theoretic conflict resolution mechanisms, this enables aircraft to fly closely at smaller intervals while ensuring safety, effectively mitigating congestion problems in high-density scenarios.

[0023] Third, it enhances the robustness of scheduling in complex environments. This application incorporates communication signal quality and meteorological wind field into the cost function of path planning. The generated actions and tracks are not only physically feasible, but can also actively maintain high-quality data link connections, reducing the risk of aircraft loss of control due to signal blockage or strong wind disturbances. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies of this application will be briefly introduced below. Obviously, the following drawings are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The drawings described herein are only used to explain this application and are not intended to limit this application.

[0025] Figure 1 This is a flowchart of a low-altitude high-density airspace dynamic scheduling method based on spatiotemporal voxel fields and reinforcement learning provided in one embodiment of this application; Figure 2 This is a detailed schematic diagram of a low-altitude high-density airspace dynamic scheduling method based on spatiotemporal voxel fields and reinforcement learning provided in one embodiment of this application; Figure 3 This is a schematic diagram of the properties of a voxel unit provided in one embodiment of this application; Figure 4 This is a schematic diagram of a probability calculation pipeline process provided in one embodiment of this application; Figure 5 This is a schematic diagram illustrating the working principle of the multi-agent reinforcement learning model and security filter provided in one embodiment of this application; Figure 6 This is a schematic diagram of a dynamic game conflict resolution mechanism provided in one embodiment of this application; Figure 7 This is a schematic diagram of the structure of a low-altitude high-density airspace dynamic scheduling system based on spatiotemporal voxel fields and reinforcement learning provided in another embodiment of this application; Figure 8 This is a schematic diagram of the structure of an electronic device provided in another embodiment of this application. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. Those skilled in the art will understand that many technical details have been provided in the embodiments of this application to facilitate better understanding. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of this application. The following embodiments can be combined with and referenced by each other without contradiction.

[0027] One embodiment of this application proposes a dynamic scheduling method for low-altitude high-density airspace based on spatiotemporal voxel fields and reinforcement learning. The implementation details of the dynamic scheduling method for low-altitude high-density airspace based on spatiotemporal voxel fields and reinforcement learning proposed in this embodiment are described in detail below. The following implementation details are provided for ease of understanding and are not necessary for implementing this solution.

[0028] The specific process of the low-altitude high-density airspace dynamic scheduling method based on spatiotemporal voxel fields and reinforcement learning proposed in this embodiment can be described as follows: Figure 1 As shown, its implementation details are as follows: Figure 2 As shown, the method includes: S1 integrates lidar point cloud, signal quality mapping of terrestrial 5G-A communication base stations, and real-time wind field data from weather radar to discretize the continuous low-altitude airspace of the target into an anisotropic spatiotemporal voxel field.

[0029] Specifically, the spatiotemporal voxel field is the basis for dynamic scheduling of low-altitude high-density airspace in this embodiment. Therefore, the first step in dynamic scheduling of low-altitude high-density airspace is to integrate the point cloud of lidar, the signal quality mapping of ground 5G-A communication base stations and the real-time wind field data of meteorological radar, and discretize the continuous target low-altitude airspace into anisotropic spatiotemporal voxel fields. Each voxel unit in the spatiotemporal voxel field not only records the spatial occupancy state, but also dynamically embeds environmental semantic information including wind vector, communication signal-to-noise ratio and potential energy field gradient.

[0030] Understandably, in low-altitude, high-density airspace, traditional latitude and longitude coordinates cannot effectively represent the constraints of complex physical fields. Therefore, this embodiment no longer uses the traditional two-dimensional latitude and longitude coordinate system, but instead constructs an anisotropic spatiotemporal voxel field to represent the target low-altitude airspace. This modeling method enables the dynamic scheduling of low-altitude, high-density airspace to perceive the invisible environmental constraints, thereby avoiding potential physical risks and the risk of losing contact at the source.

[0031] In one example, please refer to Figure 3 Spacetime voxel field It is a size of A three-dimensional cubic elemental mesh, , , They are respectively Length, width, and height.

[0032] For any time Location index is voxel units It is defined as the following high-dimensional feature tensor. .

[0033] express The occupancy status of static obstacles is pre-loaded using a high-precision map. Static obstacles include buildings and no-fly zones. , This indicates that the space is not occupied. This indicates that the item is occupied.

[0034] express The real-time wind vector, including wind speed and wind direction angle, is obtained through real-time inversion using weather radar or Doppler lidar.

[0035] express The real-time communication signal connection quality is calculated by interpolation based on the signal quality mapping of terrestrial 5G-A communication base stations, and is used to assess the risk of link failure for the aircraft. , This indicates that the real-time communication signal connection quality is good. This indicates poor quality of real-time communication signal connection.

[0036] This represents the repulsive potential energy value calculated based on the artificial potential field method. , for The Euclidean distance between the voxel center and the nearest obstacle. The preset threshold for the range of influence. This is the preset potential field gain coefficient.

[0037] express In the future time window The probability density estimate of the interior being occupied by aircraft.

[0038] The aforementioned high-dimensional feature tensors are encoded using a graph neural network and converted into low-dimensional embedding vectors. This serves as the input to the environment portion of the state space for subsequent reinforcement learning.

[0039] S2, based on spatiotemporal voxel fields, uses graph neural networks to model the interaction relationships between various aircraft in the target low-altitude airspace. It treats the aircraft as a node in the graph, uses an attention mechanism to aggregate the state features of neighboring nodes, and predicts the probability pipeline of each aircraft in the future time window through historical trajectory sequences.

[0040] Specifically, to address the uncertainty of multi-aircraft game in low-altitude, high-density environments, this embodiment constructs a trajectory prediction mechanism based on graph neural networks. This mechanism uses a spatiotemporal voxel field and a graph neural network to model the interaction relationships between aircraft in the target low-altitude airspace. The aircraft is treated as a node in the graph, and the attention mechanism is used to aggregate the state features of neighboring nodes. The probability pipeline of each aircraft in the future time window is predicted through historical trajectory sequences.

[0041] In one example, please refer to Figure 4 This embodiment constructs a spatiotemporal attention map for each aircraft within the target low-altitude airspace. , All aircraft as Node set The interaction between each aircraft and all its neighboring aircraft within its communication range is used as... edge set .

[0042] exist At that moment, for the first aircraft Its historical trajectory sequence is , ,in, , , They are respectively exist The three-dimensional position and three-dimensional velocity vector at any given moment. , They are respectively exist Azimuth and elevation angles at any given time.

[0043] Based on this, this embodiment uses LSTM to encode the historical trajectory sequences of each aircraft, and calculates the attention coefficients between the aircraft based on the state features obtained from the LSTM encoding of the historical trajectory sequences of each aircraft. Let the first... One aircraft With the One aircraft The attention coefficient between them is ,but The calculation process can be expressed by the formula: ; in, for The corresponding state characteristics, for The corresponding state characteristics, The weight matrix is ​​a learnable matrix. The attention vector, with the top-right subscript. This indicates taking the transpose. This indicates that a splicing operation is being performed. Representation layer normalization, This represents the Softmax function.

[0044] For each aircraft, a graph neural network is used to aggregate, update, and decode based on its state features and its attention coefficients with other aircraft to obtain the future... The position distribution sequence within a time step is used to form a probability pipeline with a confidence interval. The probability pipeline clearly defines the spatiotemporal range of the future existence of the spacecraft and is used to identify potential conflicts.

[0045] Record No. One aircraft The probability pipeline is , This can be expressed by the formula: ; in, , The first time The position mean and position covariance matrix.

[0046] In short, this embodiment introduces a graph neural network to perform topological modeling of the multi-aircraft interaction relationships within the target low-altitude airspace. Each aircraft is treated as a node in the graph, and an attention mechanism is used to aggregate the state features of neighboring nodes. Historical trajectory sequences are then used to predict the probabilistic pipeline of the target aircraft within future time windows. Unlike traditional linear trajectory prediction, this probabilistic pipeline can quantify uncertainty, identify potential soft and hard conflict regions, and provide feedforward information for subsequent decision-making.

[0047] S3 inputs the current state of the spatiotemporal voxel field and the probability pipelines of each aircraft in the future time window into the pre-trained multi-agent reinforcement learning model to obtain the initial instructions output by the multi-agent reinforcement learning model.

[0048] Specifically, after the probability pipeline calculation is completed, the decision-making stage can be entered. The current state of the spatiotemporal voxel field and the probability pipelines of each aircraft in the future time window are input into the pre-trained multi-agent reinforcement learning model to obtain the initial instructions output by the multi-agent reinforcement learning model.

[0049] In one example, please refer to Figure 5 In this embodiment, all aircraft are pre-defined as multi-agents, and the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm is used as the decision-making core to define the state space sequentially. Action space and reward space Train a policy network As a pre-trained multi-agent reinforcement learning model, among which, , , , For policy networks Learnable parameters.

[0050] yes The state at any given time includes the agent's own state, local perceived environmental features, and the probability pipeline of all neighboring agents. The agent's own state includes position, velocity, and remaining battery power; the local perceived environmental features are voxel unit encodings (i.e., the low-dimensional embedding vectors mentioned earlier). ).

[0051] yes The action at any given moment is a series of continuous control commands, including acceleration vector, azimuth rate of change, and pitch rate of change.

[0052] yes The reward function at time step is designed as a multi-objective weighted sum, expressed by the formula: ; in, for The distance reward at any given moment increases as you get closer to the target point. for A safety reward is given when the distance between the agent and an obstacle or neighboring agent is less than the safe distance. The safety reward is the safety distance penalty. If an agent enters a signal blind zone, the safety reward is the signal blind zone penalty. , for Momentary comfort rewards are used to punish drastic changes in acceleration. The reward for energy consumption is calculated based on wind field vectors; flying with the wind results in a positive value and flying against the wind results in a negative value and a penalty. , , , They are respectively , , , The corresponding reward weighting coefficient.

[0053] Building upon this, the multi-agent reinforcement learning model calculates the mean and variance of the action distribution based on the current state of the spatiotemporal voxel field and the probability pipelines of each aircraft within the future time window, and obtains preliminary actions through sampling. This is a preliminary instruction. It should be noted that preliminary instructions may contain unsafe instructions.

[0054] S4 uses a safety filter based on a control obstacle function to correct the deviation of the initial command, transforming the flight safety constraint into a quadratic programming optimization problem, and obtaining the safety command output by the safety filter.

[0055] Specifically, the initial instructions may contain unsafe instructions. In order to solve the randomness and unsafety problems of the initial instructions, this embodiment sets up a safety filter based on the control obstacle function. The deviation of the initial instructions is corrected by this safety filter, and the flight safety constraints are transformed into a quadratic programming optimization problem to obtain the safe instructions output by the safety filter.

[0056] In one example, please refer to Figure 5 Since security filters are built based on control barrier functions, the first step is to construct a control barrier function. This makes the secure set Defined as , For state variables, For dimension The space.

[0057] Defined as the norm of relative distance, it is expressed as: ; in, for Location, The location of obstacles or neighboring intelligent agents. This is a preset safety distance, usually 5m.

[0058] In this embodiment, the dynamic equations of the aircraft are defined as affine nonlinear equations. , and They are two different affine functions. As an action variable, This refers to the updated state variables.

[0059] To ensure forward invariance, The following Lie derivative constraint inequalities must be satisfied: ; , ; in, yes Li Daoshu, yes Li Daoshu, To extend the class function, this Lie derivative constraint inequality guarantees that if the initial state is in the safe set, then the future state will always be in the safe set.

[0060] based on In accordance with the aforementioned Lie derivative constraint inequality, the safety filter transforms the flight safety constraints into a quadratic programming optimization problem (QP problem), aiming to find a solution that is consistent with... The new action with minimal deviation and meeting safety constraints. , which serves as the safety command output by the safety filter.

[0061] The quadratic programming optimization problem is expressed by the following formula: ; ; in, The lower bound of the action. The upper bound of the action. This ensures that each aircraft obeys efficient dispatch commands as much as possible, while the constraints form a red line; once... When attempting to guide an aircraft beyond a safety boundary, the QP optimization problem forces actions to be... Projected onto the tangent plane of the safe and feasible region, since the QP optimization problem is a convex optimization problem, this embodiment uses the OSQP solver to complete the calculation in milliseconds (less than 10ms), which meets the real-time requirements of high-density flight.

[0062] A safety filter and a multi-agent reinforcement learning model together form a multi-agent reinforcement learning framework with a safety filtering mechanism. This framework utilizes a policy network of "centralized training and distributed execution" to output preliminary instructions based on the current voxel field state, aiming to maximize task efficiency. However, the preliminary instructions are not directly issued to the flight control systems of each aircraft, but must pass through a safety filter based on a control obstacle function. The safety filter transforms flight safety constraints into a QP optimization problem, which essentially constructs a safe feasible region within the action space. When the preliminary instructions output by the multi-agent reinforcement learning model are within the safe region, the instructions can be executed directly. If the preliminary instructions may lead to a violation of the minimum safe interval or contact with obstacle boundaries in future time steps, the safety filter will force a minimum deviation correction to the preliminary instructions, thus mapping them back to the boundary of the safe region. This mechanism cuts off the execution path of dangerous actions from the underlying logic of control theory, ensuring the theoretical safety of the scheduling strategy.

[0063] S5 sends safety commands to the flight control systems of each aircraft in the target low-altitude airspace, enabling safe scheduling, autonomous obstacle avoidance, and efficient coordination of the aircraft cluster in the target low-altitude airspace.

[0064] Specifically, upon receiving a safety instruction, the instruction can be sent to the flight control systems of each aircraft within the target low-altitude airspace, thereby enabling safe scheduling, autonomous obstacle avoidance, and efficient coordination of the aircraft cluster within the target low-altitude airspace.

[0065] It is also important to note that in low-altitude, high-density airspace scenarios, simple obstacle avoidance may lead to multi-aircraft deadlock or path oscillations, which, if not handled properly, could cause large-scale aircraft mission failures. To address this issue, this embodiment introduces a non-cooperative game strategy based on the VCG (Vickrey-Clarke-Groves) auction mechanism.

[0066] Please refer to Figure 6In this embodiment, when the probability pipelines of multiple aircraft are detected to overlap in the same voxel unit, it is confirmed that the multiple aircraft involved are competing for resources. At this time, the competition process will be triggered to achieve dynamic game conflict resolution.

[0067] First, each participating aircraft determines its corresponding right-of-way valuation based on its own status and mission. (Note: The last sentence appears to be incomplete and possibly contains errors. It's best left as is.) One aircraft The value of the right-of-way is , The calculation formula is: ; in, for Task priority, for The remaining battery power, The preset minimum safe power level, for The cumulative delay time of the task , , These are the weighting coefficients.

[0068] Each participating aircraft uses its calculated right-of-way valuation as its bid. In this embodiment, based on the right-of-way valuations of each aircraft involved, the following formula is used to find an optimal right-of-way allocation scheme that maximizes total social welfare. : ; Among them, total social welfare refers to the sum of benefits obtained by all participating aircraft. Represents all possible right-of-way allocation schemes. In the right-of-way allocation scheme Down The result of the right-of-way allocation, i.e. the order of passage, includes priority passage or circling and waiting.

[0069] Finally, to prevent false bids, the aircraft that obtains the right-of-way needs to pay a virtual cost to compensate for the losses caused to the remaining aircraft. Assuming... The virtual cost that one needs to pay to obtain the right-of-way. for: ; in, Indicates no Participate in the competition, Indicates no When participating in the competition, the remaining aircraft are in the optimal right-of-way allocation scheme The total social welfare obtained by the following express After participating in the competition and obtaining the right-of-way, the remaining aircraft will be in the optimal right-of-way allocation scheme. The total social welfare obtained by the public.

[0070] Through this mechanism, the system can automatically determine that emergency rescue aircraft or low-battery aircraft will be given priority passage with extremely high valuations, while ordinary mission aircraft will hover and wait in a safe area or detour, thereby achieving a dynamic balance between global efficiency and individual fairness, and realizing globally optimal collaborative scheduling in a non-cooperative environment.

[0071] In summary, the low-altitude high-density airspace dynamic scheduling method based on spatiotemporal voxel fields and reinforcement learning proposed in this embodiment has the following significant advantages compared with traditional speed regulation methods.

[0072] First, it achieves deterministic assurance of flight safety. This embodiment introduces a control obstacle function as a safety foundation for AI decision-making. In the event of misjudgment by the multi-agent reinforcement learning model or interference from unknown environments, it can forcibly correct the flight trajectory based on safety constraints. This fundamentally solves the uninterpretability and safety hazards of traditional end-to-end reinforcement learning models and meets the high reliability requirements of aviation-grade systems.

[0073] Secondly, it significantly improves the throughput and resource utilization of low-altitude airspace. The application of anisotropic spatiotemporal voxel fields allows for a finer granularity of airspace management from the traditional route level to the grid level. Combined with dynamic game-theoretic conflict resolution mechanisms, this enables aircraft to fly closely at smaller intervals while ensuring safety, effectively mitigating congestion problems in high-density scenarios.

[0074] Third, it enhances the robustness of scheduling in complex environments. This embodiment incorporates communication signal quality and meteorological wind field into the cost function of path planning. The generated actions and tracks are not only physically feasible, but can also actively maintain high-quality data link connections, reducing the risk of aircraft loss of control due to signal blockage or strong wind disturbances.

[0075] The steps described above are merely for clarity in describing the technical solution. In actual implementation, they can be combined into one step, or certain steps can be broken down into multiple steps, as long as they involve the same logical relationship, they are all within the scope of protection of this application. Any insignificant modifications or designs added to the algorithm or process, as long as they do not change the core of the algorithm or process, are also within the scope of protection of this application.

[0076] Another embodiment of this application proposes a dynamic scheduling system for low-altitude high-density airspace based on spatiotemporal voxel fields and reinforcement learning. The details of the dynamic scheduling system for low-altitude high-density airspace based on spatiotemporal voxel fields and reinforcement learning proposed in this embodiment are described in detail below. The following content is only for the convenience of understanding and is not necessary for implementing this solution. Figure 7 This is a schematic diagram of the structure of a low-altitude high-density airspace dynamic scheduling system based on spatiotemporal voxel fields and reinforcement learning proposed in this embodiment. The system includes: a spatiotemporal voxel field construction module M1, an interaction relationship modeling module M2, a preliminary instruction generation module M3, a deviation correction and safety instruction generation module M4, and a scheduling execution module M5.

[0077] The spatiotemporal voxel field construction module M1 is used to fuse lidar point clouds, signal quality mapping of terrestrial 5G-A communication base stations, and real-time wind field data from weather radar to discretize the continuous low-altitude airspace of the target into an anisotropic spatiotemporal voxel field. Each voxel unit in the spatiotemporal voxel field not only records the spatial occupancy state but also dynamically embeds environmental semantic information, including wind vectors, communication signal-to-noise ratio, and potential field gradient.

[0078] The interaction relationship modeling module M2 is used to model the interaction relationships between various aircraft in the target low-altitude airspace based on spatiotemporal voxel fields and using graph neural networks. It treats the aircraft as a node in the graph, uses an attention mechanism to aggregate the state features of neighboring nodes, and predicts the probability pipeline of each aircraft in the future time window through historical trajectory sequences.

[0079] The preliminary instruction generation module M3 is used to input the current state of the spatiotemporal voxel field and the probability pipelines of each aircraft in the future time window into the pre-trained multi-agent reinforcement learning model to obtain the preliminary instructions output by the multi-agent reinforcement learning model.

[0080] The deviation correction and safety command generation module M4 is used to correct the deviation of the initial command by using a safety filter based on the control obstacle function, transforming the flight safety constraints into a quadratic programming optimization problem, and obtaining the safety command output by the safety filter.

[0081] The scheduling and execution module M5 is used to issue safety commands to the flight control systems of each aircraft in the target low-altitude airspace, so as to realize the safe scheduling, autonomous obstacle avoidance and efficient coordination of the aircraft cluster in the target low-altitude airspace.

[0082] It is worth noting that all modules involved in this embodiment are logical modules. In practical applications, a logical module can be a physical module, a part of a physical module, or an organic combination of multiple physical modules. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce modules that are not closely related to solving the technical problems proposed in this application. However, this does not mean that other modules are absent from this embodiment.

[0083] It is not difficult to see that this embodiment is a system embodiment corresponding to the above method embodiments, and this embodiment can be implemented in conjunction with the above method embodiments. The relevant technical details and technical effects mentioned in the above method embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above method embodiments.

[0084] Another embodiment of this application provides an electronic device, such as Figure 8 As shown, it includes a processor C1 and a memory C2. The memory C2 stores instructions that the processor C1 can execute. When the processor C1 is configured to execute the instructions, the electronic device can realize a low-altitude high-density airspace dynamic scheduling method based on spatiotemporal voxel fields and reinforcement learning as described in the above method embodiment.

[0085] The memory and processor are connected via a bus, which includes any number of interconnecting buses and bridges. The bus can connect various circuits of one or more processors and memories, as well as other circuits such as peripherals, voltage regulators, and power management circuits—all well-known in the art and therefore not described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single component or multiple components, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which also receives and transmits data to the processor.

[0086] The processor manages the bus and handles general processing, providing various functions, including but not limited to timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory, on the other hand, is used to store data used by the processor during operation.

[0087] Another embodiment of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, can implement a low-altitude, high-density airspace dynamic scheduling method based on spatiotemporal voxel fields and reinforcement learning as described in the above method embodiments.

[0088] That is, those skilled in the art will understand that all or part of the steps in the above method embodiments can be implemented by a program instructing related hardware. The program is stored in a storage medium and includes several instructions to cause a device (such as a microcontroller, chip, etc.) or processor to execute all or part of the steps of the method described in the method embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0089] It will be understood by those skilled in the art that the above embodiments are specific implementations of this application, and various changes in form and detail can be made in practical applications without departing from the spirit and scope of this application. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A dynamic scheduling method for low-altitude, high-density airspace based on spatiotemporal voxel fields and reinforcement learning, characterized in that, The method includes: By integrating lidar point cloud, signal quality mapping of terrestrial 5G-A communication base stations, and real-time wind field data from meteorological radar, the continuous low-altitude airspace of the target is discretized into anisotropic spatiotemporal voxel fields. Each voxel in the spatiotemporal voxel field not only records the spatial occupancy state but also dynamically embeds environmental semantic information, including wind vector, communication signal-to-noise ratio, and potential field gradient. Based on the spatiotemporal voxel field, a graph neural network is used to model the interaction relationship between various aircraft in the target low-altitude airspace. The aircraft is regarded as a node in the graph, and the state features of neighboring nodes are aggregated by the attention mechanism. The probability pipeline of each aircraft in the future time window is predicted by the historical trajectory sequence. The current state of the spatiotemporal voxel field and the probability pipelines of each aircraft in the future time window are input into the pre-trained multi-agent reinforcement learning model to obtain the initial instructions output by the multi-agent reinforcement learning model. By using a safety filter based on a control obstacle function to correct the deviation of the initial command, the flight safety constraints are transformed into a quadratic programming optimization problem, and the safety command output by the safety filter is obtained. Safety commands are sent to the flight control systems of each aircraft in the target low-altitude airspace to enable safe scheduling, autonomous obstacle avoidance and efficient coordination of the aircraft cluster in the target low-altitude airspace. Spacetime voxel field It is a size of A three-dimensional cubic elemental mesh, , , They are respectively Length, width, and height; For any time Location index is voxel units It is defined as the following high-dimensional feature tensor. ; express The occupancy status of static obstacles is pre-loaded using a high-precision map. Static obstacles include buildings and no-fly zones. , This indicates that the space is not occupied. Indicates that the space is occupied; express The real-time wind vector, including wind speed and wind direction angle, is obtained by real-time inversion using weather radar or Doppler lidar. express The real-time communication signal connection quality is calculated by interpolation based on the signal quality mapping of terrestrial 5G-A communication base stations, and is used to assess the risk of link failure for the aircraft. , This indicates that the real-time communication signal connection quality is good. This indicates poor quality of real-time communication signal connection; This represents the repulsive potential energy value calculated based on the artificial potential field method. , for The Euclidean distance between the voxel center and the nearest obstacle. The preset threshold for the range of influence. This is the preset potential field gain coefficient; express In the future time window The probability density estimate of the interior space occupied by aircraft.

2. The method for dynamic scheduling of low-altitude, high-density airspace based on spatiotemporal voxel fields and reinforcement learning according to claim 1, characterized in that, Based on spatiotemporal voxel fields, a graph neural network is used to model the interaction relationships between various aircraft in the target low-altitude airspace. The aircraft is treated as a node in a graph, and an attention mechanism is used to aggregate the state features of neighboring nodes. Historical trajectory sequences are used to predict the probability pipeline of each aircraft within a future time window, including: Construct a spatiotemporal attention map for each aircraft in the target low-altitude airspace. , All aircraft as Node set The interaction between each aircraft and all its neighboring aircraft within its communication range is used as... edge set ; exist At that moment, for the first aircraft Its historical trajectory sequence is , ,in, , , They are respectively exist The three-dimensional position and three-dimensional velocity vector at any given moment. , They are respectively exist Azimuth and elevation angles at any given time; The historical trajectory sequences of each aircraft are encoded using LSTM, and the attention coefficients between the aircraft are calculated based on the state features obtained from the encoding of the historical trajectory sequences of each aircraft using LSTM. Record No. One aircraft With the One aircraft The attention coefficient between them is , The calculation process is expressed by the formula: ； in, for The corresponding state characteristics, for The corresponding state characteristics, The weight matrix is ​​a learnable matrix. The attention vector, with the top-right subscript... This indicates taking the transpose. This indicates that a splicing operation is being performed. Representation layer normalization, Represents the Softmax function; For each aircraft, a graph neural network is used to aggregate, update, and decode based on its state features and its attention coefficients with other aircraft to obtain the future... The position distribution sequence within the time step is used to form a probability pipeline with a confidence interval. The probability pipeline clearly defines the spatiotemporal range of the future existence of the spacecraft and is used to identify potential conflicts. Record No. One aircraft The probability pipeline is , This can be expressed by the formula: ； in, , The first time The position mean and position covariance matrix.

3. The method for dynamic scheduling of low-altitude, high-density airspace based on spatiotemporal voxel fields and reinforcement learning according to claim 2, characterized in that, The current state of the spatiotemporal voxel field and the probability pipelines of each aircraft within the future time window are input into a pre-trained multi-agent reinforcement learning model to obtain preliminary instructions output by the multi-agent reinforcement learning model, including: All aircraft are pre-defined as multi-agent agents, and a multi-agent proximal policy optimization algorithm is used as the decision-making core to define the state space. Action space and reward space Train a policy network As a pre-trained multi-agent reinforcement learning model; among which, , , , For policy networks Learnable parameters; yes The state at any given time includes the agent's own state, local perceived environmental features, and the probability pipeline of all neighboring agents; where the agent's own state includes position, velocity, and remaining battery power, and the local perceived environmental features are voxel unit encodings. yes The action at any given moment is a series of continuous control commands, including acceleration vector, azimuth rate of change, and pitch rate of change; yes The reward function at time step is designed as a multi-objective weighted sum, expressed by the formula: ； in, for The distance reward at any given moment increases as you get closer to the target point. for A safety reward is given when the distance between the agent and an obstacle or neighboring agent is less than the safe distance. The safety reward is the safety distance penalty. If an agent enters a signal blind zone, the safety reward is the signal blind zone penalty. , for Momentary comfort rewards are used to punish drastic changes in acceleration. The reward for energy consumption is calculated based on wind field vectors; flying with the wind results in a positive value and flying against the wind results in a negative value and a penalty. , , , They are respectively , , , The corresponding reward weighting coefficient; The multi-agent reinforcement learning model calculates the mean and variance of the action distribution based on the current state of the spatiotemporal voxel field and the probability pipeline of each aircraft within the future time window, and obtains preliminary actions through sampling. As preliminary instructions; among which, the preliminary instructions contain unsafe instructions.

4. The method for dynamic scheduling of low-altitude, high-density airspace based on spatiotemporal voxel fields and reinforcement learning according to claim 3, characterized in that, By using a safety filter based on a control obstacle function to correct deviations from initial commands, the flight safety constraints are transformed into a quadratic programming optimization problem, yielding the safety commands output by the safety filter, including: Construct a control barrier function This makes the secure set Defined as , For state variables, For dimension space, Defined as the norm of relative distance, it is expressed as: ； in, for Location, The location of obstacles or neighboring intelligent agents; The dynamic equations of the aircraft are defined as affine nonlinear equations. , and They are two different affine functions. As an action variable, For the updated state variables; To ensure forward invariance, The following Lie derivative constraint inequalities must be satisfied: ； ， ； in, yes Li Daoshu, yes Li Daoshu, For extension class functions; based on In accordance with the aforementioned Lie derivative constraint inequality, the safety filter transforms the flight safety constraints into a quadratic programming optimization problem, aiming to find a solution that is consistent with... The new action with minimal deviation and meeting safety constraints. , which serves as the safety command output by the safety filter.

5. The method for dynamic scheduling of low-altitude, high-density airspace based on spatiotemporal voxel fields and reinforcement learning according to claim 4, characterized in that, The quadratic programming optimization problem is expressed by the following formula: ； ； in, The lower bound of the action. This is the upper bound of the action.

6. The method for dynamic scheduling of low-altitude, high-density airspace based on spatiotemporal voxel fields and reinforcement learning according to claim 2, characterized in that, If the probability pipelines of multiple aircraft are detected to overlap in the same voxel unit, it is confirmed that the multiple aircraft involved are competing for resources. At this time, the competition process will be triggered to achieve dynamic game conflict resolution. Each participating aircraft determines its corresponding right-of-way valuation based on its own state and mission, denoted as the first... One aircraft The value of the right-of-way is , The calculation formula is: ； in, for Task priority, for The remaining battery power, The preset minimum safe power level, for The cumulative delay time of the task , , These are the weighting coefficients; Based on the estimated road rights of each aircraft involved, the following formula is used to find an optimal road rights allocation scheme that maximizes total social welfare. : ； Among them, total social welfare refers to the sum of benefits obtained by all participating aircraft. Represents all possible right-of-way allocation schemes. In the right-of-way allocation scheme Down The result of the right-of-way allocation, i.e. the order of passage, includes priority passage or circling and waiting; To prevent false bidding, the aircraft that obtains right-of-way needs to pay a virtual cost to compensate for the losses caused to the remaining aircraft. Assuming... The virtual cost that one needs to pay to obtain the right-of-way. for: ； in, Indicates no Participate in the competition, Indicates no When participating in the competition, the remaining aircraft are in the optimal right-of-way allocation scheme The total social welfare obtained by the following express After participating in the competition and obtaining the right-of-way, the remaining aircraft will be in the optimal right-of-way allocation scheme. The total social welfare obtained by the individual.

7. A low-altitude, high-density airspace dynamic scheduling system based on spatiotemporal voxel fields and reinforcement learning, used to implement the low-altitude, high-density airspace dynamic scheduling method based on spatiotemporal voxel fields and reinforcement learning as described in any one of claims 1 to 6, characterized in that, The system includes: The spatiotemporal voxel field construction module is used to fuse lidar point clouds, signal quality mapping of terrestrial 5G-A communication base stations, and real-time wind field data from weather radar to discretize the continuous low-altitude airspace of the target into anisotropic spatiotemporal voxel fields. Each voxel unit in the spatiotemporal voxel field not only records the spatial occupancy state but also dynamically embeds environmental semantic information, including wind vectors, communication signal-to-noise ratio, and potential field gradient. The interaction relationship modeling module is used to model the interaction relationship between various aircraft in the target low-altitude airspace based on the spatiotemporal voxel field and using graph neural networks. It treats the aircraft as a node in the graph, uses an attention mechanism to aggregate the state features of neighboring nodes, and predicts the probability pipeline of each aircraft in the future time window through historical trajectory sequences. The preliminary instruction generation module is used to input the current state of the spatiotemporal voxel field and the probability pipelines of each aircraft in the future time window into the pre-trained multi-agent reinforcement learning model to obtain the preliminary instructions output by the multi-agent reinforcement learning model. The deviation correction and safety command generation module is used to correct the deviation of the initial command by using a safety filter based on the control obstacle function, transforming the flight safety constraints into a quadratic programming optimization problem, and obtaining the safety command output by the safety filter. The scheduling and execution module is used to issue safety commands to the flight control systems of each aircraft in the target low-altitude airspace, so as to realize the safe scheduling, autonomous obstacle avoidance and efficient coordination of the aircraft cluster in the target low-altitude airspace.

8. An electronic device, characterized in that, include: The processor and memory, wherein the memory stores instructions that the processor can execute, and the processor is configured to, when executing the instructions, enable the electronic device to implement a low-altitude high-density airspace dynamic scheduling method based on spatiotemporal voxel fields and reinforcement learning as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it can implement a low-altitude high-density airspace dynamic scheduling method based on spatiotemporal voxel fields and reinforcement learning as described in any one of claims 1 to 6.

Images