Intelligent scheduling method and system for low-altitude resources based on deep learning

Abstract

The present application relates to the technical field of low-altitude equipment, and more particularly to a low-altitude resource intelligent scheduling method and system based on deep learning, which collects real-time state and network load data of a low-altitude flight equipment group, and constructs a dynamic operation data set; generates an operation mode feature set through multi-dimensional space situation awareness and analysis, and performs sparse clustering division based on the feature set to form a network resource demand priority mapping table; dynamically calculates resource demand by traversing the mapping table, determines multi-dimensional weight coefficients, performs high-dimensional feature dimension reduction and optimization through a mixed integer nonlinear programming solver, and obtains a resource demand feature vector; constructs a resource scheduling strategy optimization model based on the vector using a deep reinforcement learning algorithm; inputs real-time data into the model to execute resource scheduling decisions, and outputs a dynamic allocation strategy; and simulates the strategy in a digital twin simulation platform, and checks compliance to realize collaborative intelligent scheduling of communication, calculation and spectrum resources.

Description

Technical Field

[0001] This invention relates to the field of low-altitude equipment technology, and in particular to a method and system for intelligent scheduling of low-altitude resources based on deep learning. Background Technology

[0002] Low-altitude equipment networks exhibit significant characteristics of high dynamism, strong heterogeneity, and dense deployment. Traditional static or rule-driven network resource scheduling methods are ill-suited to address core challenges in low-altitude environments, such as frequent communication link switching, limited spectrum resources, unbalanced computational loads, and significant differences in Quality of Service (QoS) requirements among various task types. This leads to problems such as low resource utilization efficiency, high response latency, and insufficient system robustness.

[0003] In existing technologies, some solutions attempt to allocate resources through preset strategies or heuristic algorithms. However, their adaptability is limited by the completeness of prior knowledge, lacking the ability to perceive and predict the dynamic evolution of airspace operational status. Other methods, while introducing machine learning models to predict traffic load, are mostly limited to single-dimensional feature analysis, failing to fully integrate multi-source heterogeneous information such as aircraft motion status, airspace topology changes, service type distribution, and network performance degradation. This makes it difficult to accurately model resource demands in complex low-altitude scenarios. Furthermore, existing scheduling mechanisms generally lack closed-loop optimization capabilities, failing to perform simulation verification and compliance assessments before policy execution. This can lead to risks such as resource conflicts, security breaches, or policy oscillations in actual deployments.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for intelligent scheduling of low-altitude resources based on deep learning. This aims to solve the technical problems of existing low-altitude resource scheduling methods, which are unable to cope with highly dynamic and heterogeneous low-altitude environments, and suffer from low resource utilization efficiency, high response latency, poor adaptability, and lack of closed-loop verification mechanisms, thus failing to achieve collaborative intelligent scheduling of communication, computing, and spectrum resources.

[0006] To achieve the above objectives, this invention provides a low-altitude resource intelligent scheduling method based on deep learning, the method comprising:

[0007] Collect real-time status data and network load data of low-altitude flight equipment groups within the target airspace to construct a dynamic operation dataset for low-altitude equipment;

[0008] Multi-dimensional airspace situation awareness and analysis are performed on the dynamic operation dataset of the low-altitude equipment, and an airspace operation mode feature set is generated accordingly. The dataset is then sparsely clustered based on the airspace operation mode feature set, and a network resource demand priority mapping table is generated based on the partitioning results.

[0009] The network resource demand priority mapping table is traversed to dynamically calculate the network resource demand of the low-altitude equipment group and determine the multi-dimensional dynamic resource demand weight coefficients. The multi-dimensional dynamic resource demand weight coefficients are then input into a mixed-integer nonlinear programming solver for high-dimensional feature reduction and optimization to obtain the dimensionality-reduced resource demand feature vector. Based on the resource demand feature vector, a deep reinforcement learning algorithm is used to construct a resource scheduling strategy optimization model.

[0010] The real-time acquired dynamic operation dataset of the low-altitude equipment is input into the resource scheduling strategy optimization model to execute resource scheduling decisions and output a dynamic allocation strategy for network resources.

[0011] In the low-altitude network digital twin simulation platform, the dynamic allocation strategy of network resources is simulated and its compliance is verified. Based on the verification results, the strategy is optimized through closed-loop feedback to generate the final dynamic allocation optimization strategy of network resources, so as to realize the coordinated intelligent scheduling of communication links, computing resources and spectrum resources of low-altitude equipment networks in the target airspace.

[0012] Optionally, the step of performing multi-dimensional airspace situational awareness and analysis on the low-altitude equipment dynamic operation dataset to generate an airspace operation modal feature set, and then performing sparse clustering on the dataset based on the airspace operation modal feature set, includes:

[0013] Airspace operation features are extracted from the dynamic operation dataset of the low-altitude equipment to obtain multi-dimensional airspace feature parameters;

[0014] A mixed-integer nonlinear programming model is used to perform feature selection and dimensionality reduction on the multidimensional spatial feature parameters under constraints, resulting in a subset of core spatial situation features.

[0015] Based on the aforementioned core airspace situation feature subset, the operational modes of the target airspace are analyzed and defined, and an airspace operational mode feature set is generated.

[0016] Based on the aforementioned airspace operation mode feature set, a sparse subspace clustering algorithm is used to perform cluster analysis on the dynamic operation dataset of the low-altitude equipment to obtain multiple airspace operation cluster centers;

[0017] Based on the multiple airspace operation cluster centers, the low-altitude equipment dynamic operation dataset is divided into several mutually exclusive airspace operation clusters; the resulting airspace operation clusters are output as the result of the sparse clustering partitioning.

[0018] Optionally, the step of using a mixed-integer nonlinear programming model to perform feature selection and dimensionality reduction on the multidimensional spatial feature parameters under constraints to obtain a subset of core spatial situation features includes:

[0019] A mixed-integer nonlinear programming problem model is constructed with the joint optimization objective of maximizing feature discriminative power and minimizing feature dimension, where a binary integer variable is introduced to represent whether a feature is selected.

[0020] Nonlinear constraints are imposed on the planning problem model regarding the correlation between features, information entropy, and the relevance of business requirements;

[0021] The branch and bound algorithm is used to solve the mixed-integer nonlinear programming problem to obtain the optimal feature selection scheme;

[0022] Based on the optimal feature selection scheme, a feature projection matrix is ​​constructed to transform the original multidimensional spatial feature parameters, resulting in a dimensionality-reduced feature space.

[0023] The saliency test and redundancy analysis are performed on the dimensionality-reduced feature space to select the core spatial situation feature subset.

[0024] Optionally, generating a network resource demand priority mapping table based on the partitioning results includes:

[0025] Retrieve historical network resource scheduling records and performance data for the target airspace;

[0026] Based on historical network resource scheduling records and performance data, each of the divided airspace operation clusters is assigned a resource demand attribute label.

[0027] Based on the resource demand attribute tags, the resource consumption patterns and service quality requirements of low-altitude equipment in each airspace operation cluster are analyzed, and cluster-level resource demand indicators are quantified.

[0028] Based on predefined business criticality levels and resource urgency rules, the cluster-level resource demand indicators are prioritized to determine the priority of multi-level resource demands.

[0029] By combining the multi-level resource demand priorities with the cluster-level resource demand indicators, a weighted evaluation model is constructed, and the network resource demand priority mapping table is generated accordingly.

[0030] Optionally, the step of traversing the network resource demand priority mapping table, dynamically calculating the network resource demand of the low-altitude equipment group, and determining the multi-dimensional dynamic resource demand weight coefficient includes:

[0031] Based on the historical network resource scheduling records and performance data, identify key resource nodes and their topological relationships in the target airspace network.

[0032] By combining the network resource demand priority mapping table with the key resource nodes, a dynamic resource demand prediction model is established to calculate the basic resource demand and simultaneously obtain the multi-level resource demand priority.

[0033] Based on the real-time partitioned airspace operation clusters and their dynamic data, the basic resource requirements are calibrated online to obtain a real-time resource requirement sequence;

[0034] The real-time resource demand sequence is subjected to regression analysis and weight fitting using a Bayesian optimization framework, and the weight coefficients of the multidimensional dynamic resource demand are output.

[0035] Optionally, the step of inputting the multidimensional dynamic resource demand weight coefficients into a mixed-integer nonlinear programming solver for high-dimensional feature reduction and optimization yields a dimensionality-reduced resource demand feature vector; based on the resource demand feature vector, a resource scheduling strategy optimization model is constructed using a deep reinforcement learning algorithm, including:

[0036] Based on the multidimensional dynamic resource demand weight coefficients and network topology constraints, feasible resource scheduling paths are enumerated to generate a set of candidate scheduling paths.

[0037] A resource scheduling state space is constructed using resource utilization, allocation latency, and load balancing as state indicators.

[0038] A resource scheduling action space is constructed by taking the action of selecting a path from the set of candidate scheduling paths and allocating resources.

[0039] A reward function integrating multi-objective optimization factors is designed, and a deep deterministic policy gradient algorithm is used to enable the agent to learn the optimal scheduling policy through interactive training in the state space and action space, and finally converge to obtain the resource scheduling policy optimization model.

[0040] Optionally, the simulation and compliance verification of the network resource dynamic allocation strategy in the low-altitude network digital twin simulation platform, and the closed-loop feedback optimization of the strategy based on the verification results, includes:

[0041] Construct a high-fidelity digital twin simulation environment for low-altitude networks;

[0042] The network resource dynamic allocation strategy is loaded and simulated in the simulation environment, and network performance indicators are collected after the strategy is executed.

[0043] The collected network performance indicators are compared with a predefined dynamic compliance threshold range to determine scheduling compliance.

[0044] If the judgment result is compliant, a positive reinforcement signal is generated to monitor the stability of the current strategy and make fine adjustments.

[0045] If the determination result is non-compliant, a negative feedback signal is generated, the root cause analysis process is initiated, the abnormal allocation point in the strategy is located and a diagnostic report is generated, and the diagnostic report is used to guide the iterative correction of the strategy.

[0046] Furthermore, to achieve the above objectives, the present invention also provides a low-altitude resource intelligent scheduling system based on deep learning, the system comprising:

[0047] The data acquisition module is used to collect real-time status data and network load data of low-altitude flight equipment groups within the target airspace, and to construct a dynamic operation dataset of low-altitude equipment.

[0048] The situation clustering module is used to perform multi-dimensional airspace situation perception and analysis on the dynamic operation dataset of the low-altitude equipment, generate an airspace operation mode feature set, and perform sparse clustering on the dataset based on the airspace operation mode feature set, and generate a network resource demand priority mapping table based on the partitioning result.

[0049] The demand modeling module is used to traverse the network resource demand priority mapping table, dynamically calculate the network resource demand of the low-altitude equipment group, and determine the multi-dimensional dynamic resource demand weight coefficients; input the multi-dimensional dynamic resource demand weight coefficients into a mixed integer nonlinear programming solver for high-dimensional feature reduction and optimization to obtain the dimensionality-reduced resource demand feature vector; based on the resource demand feature vector, a deep reinforcement learning algorithm is used to construct a resource scheduling strategy optimization model.

[0050] The strategy decision module is used to input the real-time acquired dynamic operation dataset of the low-altitude equipment into the resource scheduling strategy optimization model, execute resource scheduling decisions, and output a dynamic allocation strategy for network resources.

[0051] The simulation optimization module is used to simulate and verify the compliance of the network resource dynamic allocation strategy in the low-altitude network digital twin simulation platform. Based on the verification results, the strategy is optimized through closed-loop feedback to generate the final network resource dynamic allocation optimization strategy, so as to realize the coordinated intelligent scheduling of communication links, computing resources and spectrum resources of low-altitude equipment networks in the target airspace.

[0052] Furthermore, to achieve the above objectives, the present invention also provides a low-altitude resource intelligent scheduling device based on deep learning. The device includes: a memory, a processor, and a low-altitude resource intelligent scheduling program based on deep learning stored in the memory and executable on the processor. The low-altitude resource intelligent scheduling program based on deep learning is configured to implement the steps of the low-altitude resource intelligent scheduling method based on deep learning as described above.

[0053] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a deep learning-based intelligent scheduling program for low-altitude resources, wherein when the deep learning-based intelligent scheduling program for low-altitude resources is executed by a processor, it implements the steps of the deep learning-based intelligent scheduling method for low-altitude resources as described above.

[0054] This invention provides a deep learning-based intelligent scheduling method for low-altitude resources. The method constructs a multi-dimensional airspace situational awareness and sparse clustering analysis mechanism, which can accurately characterize the dynamic operating modes of low-altitude flight equipment groups and improve the ability to understand complex airspace environments. By generating a network resource demand priority mapping table and combining it with mixed-integer nonlinear programming for high-dimensional feature dimensionality reduction and optimization, it achieves quantitative modeling and key feature extraction of multi-dimensional dynamic resource demands, effectively improving the accuracy and computational efficiency of resource demand prediction. Based on the dimensionality-reduced resource demand feature vector, a deep reinforcement learning scheduling strategy optimization model is constructed, enabling the system to autonomously learn the optimal scheduling in uncertain environments. The ability to implement scheduling strategies significantly enhances the intelligence and adaptability of resource allocation. By introducing a digital twin simulation platform for low-altitude networks to simulate and verify the compliance of scheduling strategies, and implementing closed-loop feedback optimization, the entire lifecycle management of strategy generation, simulation verification, and iterative correction is achieved, effectively avoiding the risk of non-compliant scheduling and improving the security and robustness of the system. In summary, this method achieves coordinated optimization scheduling of communication links, computing resources, and spectrum resources, solving the technical problems of low resource utilization, slow response, and poor adaptability of traditional methods in highly dynamic and heterogeneous low-altitude environments, and significantly improving the overall operational efficiency and intelligent management level of low-altitude equipment networks. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating an embodiment of the intelligent low-altitude resource scheduling method based on deep learning according to the present invention.

[0056] Figure 2 This is a structural block diagram of an embodiment of the low-altitude resource intelligent scheduling system based on deep learning according to the present invention.

[0057] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0058] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0059] Reference Figure 1 , Figure 1This is a flowchart illustrating an embodiment of the intelligent low-altitude resource scheduling method based on deep learning according to the present invention. An embodiment of the intelligent low-altitude resource scheduling method based on deep learning according to the present invention is presented.

[0060] In one embodiment, the deep learning-based intelligent scheduling method for low-altitude resources includes:

[0061] Step S100: Collect real-time status data and network load data of the low-altitude flight equipment group in the target airspace to construct a dynamic operation dataset of low-altitude equipment.

[0062] The target airspace can be a specific three-dimensional low-altitude geographical area of ​​interest to the scheduling system, encompassing the aircraft's operating space and communication coverage. It can be used to define the boundaries of resource scheduling and provide a spatial reference framework for data acquisition, modeling, and strategy execution. The low-altitude flight equipment swarm can be a collection of low-altitude aircraft with communication and computing capabilities, such as UAVs and eVTOLs, operating within the target airspace. It can be used as the object of resource scheduling, and its dynamic behavior directly drives changes in network resource demand. For example, the low-altitude flight equipment swarm can include, but is not limited to, one or more of the following: logistics UAV swarms, urban air traffic eVTOL formations, and emergency inspection multirotor aircraft swarms. Real-time status data can be a multi-dimensional information stream reflecting the current motion and mission status of the low-altitude flight equipment, providing key situational elements such as aircraft position, speed, heading, and mission type. In an exemplary embodiment, real-time status data can be reported in real time through sensors or interfaces such as airborne GNSS, IMU, and mission management systems. Furthermore, real-time status data can include, but is not limited to, one or more of the following: kinematic status data, mission semantic status data, and energy and health status data.

[0063] Network load data can be a set of performance indicators describing the communication links, edge nodes, and spectrum usage within the target airspace. It can be used to characterize the current network resource occupancy and bottleneck status, supporting resource demand modeling. In a specific embodiment, network load data can be collected and reported in real time through ground base stations, airborne relay nodes, or network management systems. For example, network load data can include, but is not limited to, one or more of the following: link throughput data, edge server CPU / memory load data, spectrum occupancy rate, and interference level data. The low-altitude equipment dynamic operation dataset can be a spatiotemporally correlated data set composed of real-time status data and network load data. It can be used as the raw input for subsequent situational awareness and resource modeling, carrying a comprehensive picture of the airspace dynamics. Furthermore, the low-altitude equipment dynamic operation dataset can be formed by aggregating data from multi-source heterogeneous devices after time alignment, coordinate unification, and format standardization.

[0064] Step S200: Perform multi-dimensional airspace situational awareness and analysis on the low-altitude equipment dynamic operation dataset, generate an airspace operation modal feature set, and perform sparse clustering on the dataset based on the airspace operation modal feature set, and generate a network resource demand priority mapping table based on the partitioning results.

[0065] The multi-dimensional airspace situational awareness and analysis module can be an analysis component used to fuse multi-source information such as aircraft motion status, airspace topology, service type, and network performance. It can be used to generate a structured, computable airspace operation modal feature set. In an exemplary embodiment, the multi-dimensional airspace situational awareness and analysis module can use graph neural networks or spatiotemporal attention mechanisms to extract features and perform correlation modeling on the dynamic operation dataset of low-altitude equipment. Furthermore, the multi-dimensional airspace situational awareness and analysis module can receive the dynamic operation dataset of low-altitude equipment and output an airspace operation modal feature set. The airspace operation modal feature set can be a set of high-dimensional feature vectors that abstractly express the overall operation mode of the low-altitude equipment group, and can be used to characterize the macroscopic behavioral characteristics of the equipment group, such as aggregation, mobility, and task density, within a specific time period. In a specific embodiment, the airspace operation modal feature set can be generated by spatiotemporal correlation, semantic fusion, and pattern recognition of multi-source heterogeneous data.

[0066] The sparse clustering partitioning result can be a non-uniform clustering output based on the spatial operational modal feature set to group low-altitude equipment groups. This can be used to identify equipment subgroups with similar resource demand patterns, reducing scheduling complexity. In an exemplary embodiment, the sparse clustering partitioning result can be used to reduce the dimensionality of high-dimensional features through sparse representation theory (such as K-SVD, sparse subspace clustering), preserving significant modal differences. For example, the sparse clustering partitioning result can include, but is not limited to, one or more of the following: high-mobility communication-intensive clustering, low-speed hovering computation-intensive clustering, and high-priority clustering for bursty tasks. The network resource demand priority mapping table can be a structured rule table that maps the sparse clustering partitioning result to resource request priorities. It can be used to allocate differentiated resource scheduling priorities to different equipment groups, responding to QoS differences in various task types. In a specific embodiment, the network resource demand priority mapping table can set priority weights based on clustering labels, service QoS levels (such as latency-sensitive, bandwidth-guaranteed), and security levels.

[0067] Step S300: Traverse the network resource demand priority mapping table, dynamically calculate the network resource demand of the low-altitude equipment group, and determine the multi-dimensional dynamic resource demand weight coefficients. Input these weight coefficients into a mixed-integer nonlinear programming solver for high-dimensional feature reduction and optimization, obtaining the reduced-dimensional resource demand feature vector. Based on this feature vector, a deep reinforcement learning algorithm is used to construct a resource scheduling strategy optimization model.

[0068] The multidimensional dynamic resource demand weight coefficient can quantify the relative importance of resources (communication, computing, spectrum) in the current spatial situation, guiding the construction of the objective function in subsequent high-dimensional optimization processes and reflecting resource allocation preferences. In a specific embodiment, the multidimensional dynamic resource demand weight coefficient can be dynamically assigned based on a network resource demand priority mapping table, combined with real-time load and task urgency. Furthermore, the multidimensional dynamic resource demand weight coefficient can be used as input to a mixed-integer nonlinear programming solver to generate a resource demand feature vector. The mixed-integer nonlinear programming solver can be a mathematical tool for solving optimization problems with integer variables and nonlinear constraints, enabling effective dimensionality reduction and key feature preservation of high-dimensional features while satisfying physical and protocol constraints. In an exemplary embodiment, the mixed-integer nonlinear programming solver can call commercial or open-source optimization libraries (such as commercial optimization solvers or open-source nonlinear programming solvers) to perform constrained optimization of high-dimensional resource demands. Furthermore, the mixed-integer nonlinear programming solver can receive the multidimensional dynamic resource demand weight coefficient and output the dimensionality-reduced resource demand feature vector. The high-dimensional feature reduction and optimization module can be a processing unit integrating a mixed-integer nonlinear programming solver. It can be used to reduce the input dimension of subsequent scheduling models, improving computational efficiency and generalization ability. The resource requirement feature vector can be a low-dimensional vector representation that retains key resource requirement information after dimensionality reduction and optimization. It can be used as the state input of a deep reinforcement learning scheduling strategy optimization model, balancing expressive power and computational overhead.

[0069] In one specific embodiment, the resource demand feature vector can be generated by a mixed-integer nonlinear programming solver by extracting principal components or optimal basis under resource constraints. Furthermore, the resource demand feature vector can drive the policy learning process of the resource scheduling strategy optimization model. Deep reinforcement learning algorithms can be policy search methods combining deep neural networks and reinforcement learning frameworks, enabling the system to autonomously explore resource allocation strategies in uncertain, highly dynamic environments. For example, deep reinforcement learning algorithms can include, but are not limited to, one or more of the following: value function-based deep Q-networks, policy gradient-based proximal policy optimization, and maximum entropy-based soft actor-commentator algorithms. The resource scheduling strategy optimization model can be a deep reinforcement learning agent that takes the resource demand feature vector as input and resource allocation actions as output, and can be used to generate resource scheduling decisions adapted to the current spatial situation in real time. In one specific embodiment, the resource scheduling strategy optimization model can construct the policy network through a combination of offline pre-training and online fine-tuning.

[0070] Traversing the network resource demand priority mapping table allows for querying the corresponding priority item by cluster label and associating it with specific devices, thus achieving the technical effect of quickly assigning device-level resource demand weights. Dynamically calculating the network resource demands of low-altitude equipment clusters can be achieved by combining device task type, QoS requirements, and current load to estimate the required resource amount, thereby quantifying the specific demands of each device in communication, computing, and spectrum dimensions. Inputting the multidimensional dynamic resource demand weight coefficients into a mixed-integer nonlinear programming solver for high-dimensional feature dimensionality reduction and optimization can solve for the optimal feature representation under constraints such as total resource volume and interference threshold. Furthermore, inputting the multidimensional dynamic resource demand weight coefficients into a mixed-integer nonlinear programming solver for high-dimensional feature dimensionality reduction and optimization can be achieved by modeling resource allocation as a convex optimization problem with integer variables, solving it using the branch and bound method, or introducing sparse regularization constraints to force the solution vector to have a low-dimensional effective support set, thus achieving the technical effect of compressing feature dimensions while retaining key demand information. A resource scheduling strategy optimization model can be constructed using deep reinforcement learning algorithms. This can be achieved by using resource demand feature vectors as states and resource allocation actions as actions, and designing a reward function to train the policy network. Furthermore, the resource scheduling strategy optimization model can be implemented by using the PPO algorithm for policy gradient updates in a simulated environment, or by employing double-delay deep deterministic policy gradient (TD3) to process the continuous action space. This can achieve the technical effect of obtaining an intelligent scheduling agent capable of autonomous decision-making in dynamic environments.

[0071] Step S400: Input the real-time acquired dynamic operation dataset of low-altitude equipment into the resource scheduling strategy optimization model, execute resource scheduling decisions, and output a dynamic allocation strategy for network resources.

[0072] The dynamic allocation strategy for network resources can be a preliminary resource allocation scheme output by the resource scheduling strategy optimization model, which can be used to indicate the specific allocation results of each device on communication links, computing nodes, and spectrum channels. Inputting the real-time acquired dynamic operation dataset of low-altitude equipment into the resource scheduling strategy optimization model can be achieved by performing the same preprocessing procedure as in the training phase on the newly acquired data before feeding it into the model. This ensures that the model input distribution is consistent with the training distribution, guaranteeing the effectiveness of the decision. Executing the resource scheduling decision involves the model outputting the optimal resource allocation action based on the current state, thereby generating a preliminary dynamic allocation strategy for network resources.

[0073] In step S500, the dynamic allocation strategy for network resources is simulated and its compliance is verified in the low-altitude network digital twin simulation platform. Based on the verification results, the strategy is optimized through closed-loop feedback to generate the final dynamic allocation optimization strategy for network resources, so as to realize the coordinated intelligent scheduling of communication links, computing resources and spectrum resources of low-altitude equipment networks in the target airspace.

[0074] The low-altitude network digital twin simulation platform can be a simulation system that performs high-fidelity virtual mapping of physical networks and aircraft behavior within a target airspace. It can be used for security, compliance, and performance verification before policy deployment, mitigating risks in actual operation. In one specific embodiment, the low-altitude network digital twin simulation platform can integrate a 3D airspace model, a communication propagation model, an equipment dynamics model, and a network protocol stack. The simulation and deduction module can be a subsystem within the low-altitude network digital twin simulation platform used to perform dynamic simulation of policies, predicting the execution effects and potential conflicts of scheduling policies in a real environment. The compliance verification module can be a component that performs legality checks on scheduling policies based on air traffic control rules, spectrum regulations, and safety protocols, identifying potential violations such as boundary crossings, interference, or resource conflicts within the policies.

[0075] A closed-loop feedback optimization mechanism can be a control loop that feeds simulation verification results back to the policy model to iteratively correct parameters. This can be used to achieve continuous policy optimization and risk convergence, improving system robustness. In an exemplary embodiment, the closed-loop feedback optimization mechanism can adjust model weights through policy gradient updates, experience replay, or meta-learning. Furthermore, the closed-loop feedback optimization mechanism can connect a low-altitude network digital twin simulation platform with a resource scheduling policy optimization model. The network resource dynamic allocation optimization policy can be the final scheduling policy after digital twin simulation verification and closed-loop feedback correction. It can be used as an actual execution command to achieve safe, efficient, and compliant resource collaborative scheduling. Furthermore, the network resource dynamic allocation optimization policy can be the final output of the entire process, affecting communication links, computing resources, and spectrum resources. Communication links can be data transmission channels established between low-altitude equipment and ground stations or airborne relays. They can be used to carry control commands and service data, and their stability directly affects the quality of task completion. Computing resources can include hardware resources such as edge servers and airborne processors that can be used to perform task computations. They can be used to support the local or off-site execution of computationally intensive tasks such as perception, planning, and AI inference. Spectrum resources can be frequency bandwidth resources in licensed or unlicensed frequency bands that can be used for wireless communication. They can be used to determine communication capacity and anti-interference capability, and are a key and scarce resource for low-altitude networks.

[0076] For example, in the scenario of scheduling urban air traffic during peak hours, the deep learning-based intelligent scheduling method for low-altitude resources in this embodiment can be as follows: During the morning rush hour, commuter eVTOLs, delivery drones, and emergency medical aircraft coexist in a low-altitude corridor of a city. The system collects the location, speed, task type, and current 4G / 5G link load of all devices to construct a dynamic operational dataset; through multi-dimensional situational awareness, it identifies that the eVTOL formation is in a high-density cruise mode, the delivery drones are in a dispersed hovering mode, and the medical aircraft are in a high-priority point-to-point flight mode; sparse clustering divides them into three categories, and a priority mapping table is generated to assign the highest weight to the medical task; based on this, multi-dimensional weight coefficients are calculated, and after dimensionality reduction by mixed integer programming, they are input into a deep reinforcement learning model; the model outputs a preliminary scheduling strategy, suggesting reserving a dedicated frequency band for the medical aircraft, allocating edge computing nodes, and establishing a highly reliable link; this strategy is simulated and deduced in a digital twin platform to verify that there is no spectrum conflict and that the 99.9% link availability requirement is met; after closed-loop feedback fine-tuning, the final strategy is issued and executed to achieve the coordinated and efficient operation of the three types of services.

[0077] In one embodiment, multi-dimensional airspace situational awareness and analysis are performed on the dynamic operation dataset of low-altitude equipment to generate an airspace operation modal feature set. Based on this airspace operation modal feature set, the dataset is then sparsely clustered, including:

[0078] Airspace operation features are extracted from the dynamic operation dataset of low-altitude equipment to obtain multidimensional airspace feature parameters.

[0079] The airspace operation feature extraction unit can be a processing module used to extract key features related to airspace operation from the dynamic operation dataset of low-altitude equipment. It can be used to generate multi-dimensional feature parameters that comprehensively reflect the aircraft's behavior patterns in the airspace. In this embodiment, the airspace operation feature extraction unit can extract structured features from the raw state and load data based on signal processing, trajectory analysis, or semantic parsing methods. Furthermore, the airspace operation feature extraction unit can receive the dynamic operation dataset of low-altitude equipment as input and output multi-dimensional airspace feature parameters. These multi-dimensional airspace feature parameters can be a set of multi-dimensional quantitative indicators describing the operational state of low-altitude flight equipment in the airspace, and can be used to provide high-dimensional raw representations for subsequent feature selection and modality recognition. In an exemplary embodiment, the multi-dimensional airspace feature parameters can be generated by encoding raw data such as position, speed, heading, mission type, and link quality through the airspace operation feature extraction unit.

[0080] A mixed-integer nonlinear programming model is used to perform feature selection and dimensionality reduction on multidimensional spatial feature parameters under constraints, resulting in a subset of core spatial situation features.

[0081] The mixed-integer nonlinear programming model can be a mathematical optimization model used to perform feature selection and dimensionality reduction under physical and system constraints. It can be used to select the most discriminative core feature subset that meets practical deployment conditions from multi-dimensional spatial feature parameters. In a specific embodiment, the mixed-integer nonlinear programming model can construct a mixed-integer nonlinear programming problem with feature subset sparsity as the objective and communication distance, spectrum capacity, and computational upper limits as constraints. The constrained feature selection and dimensionality reduction mechanism can be a functional component integrating the mixed-integer nonlinear programming model, achieving constrained high-dimensional feature compression. This ensures that the dimensionality-reduced features retain both spatial situational semantics and meet physical resource boundary conditions. In this embodiment, the constrained feature selection and dimensionality reduction mechanism can connect multi-dimensional spatial feature parameters and the core spatial situational feature subset, forming a feature preprocessing path.

[0082] The core airspace situation feature subset can be a low-dimensional feature set that has strong discriminative ability for airspace operation modes, preserved after constraint optimization. It can be used as the basis for defining operation modes, supporting the accuracy and efficiency of subsequent clustering and resource mapping. For example, the core airspace situation feature subset can be obtained by solving a mixed-integer nonlinear programming model under constraints of sparsity and feasibility. Using a mixed-integer nonlinear programming model to perform feature selection and dimensionality reduction on multi-dimensional airspace feature parameters under constraints, the core airspace situation feature subset can be obtained. This can be achieved by solving for the optimal feature subset under physical constraints such as communication distance, spectrum availability, and computing capacity. Furthermore, this operation can be implemented by introducing L0 norm constraints to achieve sparse feature selection and combining it with nonlinear quality of service constraint modeling, or by using a decompositional solution strategy to break down the large-scale mixed-integer nonlinear programming problem into parallelizable subproblems. This allows for dimensionality reduction while preserving key discriminative information, improving subsequent processing efficiency and generalization ability.

[0083] Based on a subset of core airspace situation features, the operational modes of the target airspace are analyzed and defined, and an airspace operational mode feature set is generated.

[0084] The operational modalities of the target airspace can be typical airspace operational scenario categories identified based on a subset of core airspace situation features. These can be used to provide interpretable airspace state abstractions, such as "dense hovering operations" or "high-speed corridor crossings." In an exemplary embodiment, the operational modalities of the target airspace can be named and defined for typical regions in the feature space using semantic rules or cluster center labels. For example, the operational modalities of the target airspace may include, but are not limited to, dense hovering operation modalities, high-speed corridor crossing modalities, and heterogeneous formation cooperative modalities. Generating the airspace operational modal feature set can be achieved by encoding the defined operational modalities into a standardized set of feature vectors. Further, this operation can be implemented by assigning a prototype vector (generated from the mean of samples within the cluster) to each modality, or by using one-hot encoding or embedding vectors to represent the modality category and concatenating continuous features, thereby providing semantically guided input representations for sparse subspace clustering. Based on the subset of core airspace situation features, parsing and defining the operational modalities of the target airspace can be achieved by assigning semantic labels to typical regions in the feature subspace, forming interpretable operational scenario classifications. Furthermore, this operation can be achieved by matching feature patterns with predefined modal templates through an expert knowledge base, or by automatically naming modalities using clustering results (e.g., mapping "high-density low-speed cluster" to "hovering operation modality"), thereby establishing a mapping from numerical features to semantic modalities and enhancing the cognitive interpretability of the system.

[0085] Based on the airspace operation modal feature set, a sparse subspace clustering algorithm is used to perform cluster analysis on the dynamic operation dataset of low-altitude equipment to obtain multiple airspace operation cluster centers;

[0086] Among them, the sparse subspace clustering algorithm can be an unsupervised clustering method that assumes the data is distributed in multiple low-dimensional linear subspaces. It can be used to accurately identify clustered motion patterns caused by task differences and is suitable for high-dimensional, low-rank low-altitude equipment trajectory data. In a specific embodiment, the sparse subspace clustering algorithm can construct a sparse representation matrix through a self-expressive model and then use spectral clustering to divide the subspace structure. The airspace operation cluster center can be a representative feature point of each operation subspace identified by the sparse subspace clustering algorithm. It can be used as a benchmark for dividing equipment groups and guiding the granularity of subsequent resource demand modeling. In this embodiment, the airspace operation cluster center can be the cluster centroid or typical sample calculated by the clustering algorithm in the core airspace situation feature subspace. Based on the airspace operation modal feature set, the sparse subspace clustering algorithm is used to perform cluster analysis on the low-altitude equipment dynamic operation dataset to obtain multiple airspace operation cluster centers. This can be achieved by identifying multiple low-dimensional subspace structures in the dimensionality-reduced feature space and calculating their representative centers.

[0087] Based on multiple airspace operation cluster centers, the dynamic operation dataset of low-altitude equipment is divided into several mutually exclusive airspace operation clusters;

[0088] The mutually exclusive airspace operation clusters can be sets of non-overlapping low-altitude equipment groups divided according to cluster centers. These clusters can be used to form semantically clear and structurally stable equipment groups, supporting differentiated resource scheduling. In an exemplary embodiment, the mutually exclusive airspace operation clusters can be obtained by assigning each equipment trajectory to the subspace belonging to the nearest airspace operation cluster center. Dividing the low-altitude equipment dynamic operation dataset into mutually exclusive airspace operation clusters based on multiple airspace operation cluster centers can be achieved by assigning each equipment trajectory to the subspace belonging to the most similar cluster center. Furthermore, this operation can be performed through hard allocation based on subspace projection distance, or by using soft allocation to generate a membership matrix to support fuzzy scheduling, thereby forming non-overlapping equipment groups and providing clear boundaries for differentiated resource scheduling.

[0089] The resulting spatial domain clusters are output as the sparse clustering partitioning result.

[0090] The result of sparse clustering can be the final output spatial domain operational cluster partitioning scheme, containing the mapping relationship from devices to clusters, which can be used as a direct basis for generating a network resource demand priority mapping table. In this embodiment, the result of sparse clustering can be used in subsequent steps to construct the association between resource demand and operational modes. Outputting the partitioned spatial domain operational clusters as the result of sparse clustering can be achieved by serializing the mapping relationship from devices to clusters into a structured data format.

[0091] For example, in a scenario where multiple services coexist in urban low-altitude airspace, the low-altitude resource intelligent scheduling method based on deep learning in this embodiment could be: In a smart city demonstration area, delivery drones, aerial taxis, and power line inspection drones coexist in the low-altitude airspace. The system first extracts multi-dimensional airspace feature parameters from the dynamic operation dataset, including flight altitude layer, horizontal speed, mission destination type, and link signal-to-interference-plus-noise ratio. Then, using a mixed-integer nonlinear programming model, under constraints such as the coverage radius of fifth-generation mobile communication base stations, the available bandwidth of the 6 GHz band, and the upper limit of edge server computing power, it selects a subset of core airspace situational features, including "altitude-speed coupling characteristics," "mission urgency," and "link stability." Based on this subset, the system identifies three operational modes: "low-altitude slow-speed logistics mode," "high-altitude high-speed commuting mode," and "fixed-point hovering inspection mode." Next, a sparse subspace clustering algorithm is used to calculate cluster centers in the three low-dimensional subspaces and assign all devices to their corresponding mutually exclusive clusters. The final sparse clustering results clearly distinguish the three service groups, laying the foundation for subsequent allocation of communication slices with different priorities according to mode, calculation of offloading paths, and spectrum resources.

[0092] In one embodiment, a mixed-integer nonlinear programming model is used to perform feature selection and dimensionality reduction on multidimensional spatial feature parameters under constraints, resulting in a subset of core spatial situation features, including:

[0093] A mixed-integer nonlinear programming problem model is constructed with the joint optimization objective of maximizing feature discriminative power and minimizing feature dimension, where a binary integer variable is introduced to represent whether a feature is selected.

[0094] The joint optimization objective can be a multi-objective optimization criterion that simultaneously pursues high discriminative power and low dimensionality during feature selection. It can guide mixed-integer nonlinear programming models to achieve compact representation while preserving discriminative power. Furthermore, the joint optimization objective can include, but is not limited to, one or more of the following: maximizing inter-class divergence, minimizing the number of features, and maximizing information retention. Feature discriminative power can be a measure of the separability of the selected feature subset across different spatial operating modes. It can be used to ensure that the features after dimensionality reduction can still effectively support subsequent clustering and scheduling decisions. In an exemplary embodiment, feature discriminative power can be quantified using metrics such as inter-class distance, Fisher's discriminant ratio, or mutual information. Furthermore, feature discriminative power can be part of the joint optimization objective, forming the optimization function together with the feature dimension. The feature dimension can be the number of selected features, i.e., the length of the feature vector after dimensionality reduction, which can directly affect subsequent computational complexity and model generalization ability.

[0095] Mixed-integer nonlinear programming (MIBN) models can be mathematical programming models that control feature selection with binary integer variables and contain nonlinear objectives and constraints. They can be used to achieve globally optimal or near-optimal feature subset search for scheduling tasks. For example, MIBN models can be obtained by modeling feature selection as an optimization problem with combined discrete decision-making and continuous performance indices. Furthermore, MIBN models can receive multi-dimensional spatial feature parameters and output the optimal feature selection scheme. The binary integer variables can be decision variables with values ​​of 0 or 1, used to indicate whether a feature is selected into the final subset. This can provide an interpretable discrete feature selection mechanism, avoiding ambiguity caused by soft weights. In a specific embodiment, binary integer variables can be embedded in the MIBN model to directly control the construction of the feature projection matrix. Whether a feature is selected can be the feature retention state represented by the binary integer variables, which can be used as the basic unit to construct the optimal feature selection scheme.

[0096] Impose nonlinear constraints on the planning problem model regarding the correlation between features, information entropy, and the relevance of business requirements;

[0097] The nonlinear constraints can be mathematical expressions of the multidimensional semantic and statistical properties that the feature selection results must satisfy. These constraints can be used to ensure that the selected features possess low redundancy, high information content, and business relevance. For example, nonlinear constraints can include, but are not limited to, one or more of the following: feature correlation constraints, information entropy lower bound constraints, and business requirement relevance constraints. Feature correlation can be the degree of statistical dependence between any two selected features, and can be used as one of the nonlinear constraints to suppress the co-occurrence of redundant features. In an exemplary embodiment, feature correlation can be measured by the Pearson correlation coefficient, mutual information, or covariance matrix. Information entropy can be a measure of the uncertainty or information content carried by a feature, and can be used as a constraint to ensure that the selected features have sufficient information richness. Business requirement relevance can be the semantic coupling strength between the feature and the QoS requirements of low-altitude services (such as latency, reliability, and priority), and can be used to ensure that the feature selection results have practical guiding significance for scheduling tasks.

[0098] The branch and bound algorithm is used to solve the mixed-integer nonlinear programming problem and obtain the optimal feature selection scheme.

[0099] The branch-and-bound algorithm can be a global optimization search method for solving mixed-integer programming problems, and it can be used to find near-optimal feature selection schemes in NP-hard problems. The optimal feature selection scheme can be the optimal binary variable configuration that satisfies all constraints, obtained by the branch-and-bound algorithm, and can be used to explicitly indicate which original features should be retained for subsequent processing. Solving mixed-integer nonlinear programming problems using the branch-and-bound algorithm to obtain the optimal feature selection scheme can involve recursively partitioning the variable space and pruning using a relaxed lower bound and a feasible upper bound.

[0100] Based on the optimal feature selection scheme, a feature projection matrix is ​​constructed to transform the original multidimensional spatial feature parameters, resulting in a dimensionality-reduced feature space.

[0101] The feature projection matrix can be a sparse linear transformation matrix constructed based on the optimal feature selection scheme, which can be used to achieve structured dimensionality reduction of the original multidimensional spatial feature parameters. In an exemplary embodiment, the feature projection matrix can be obtained by retaining the rows corresponding to the selected features in the identity matrix and setting the rest to zero. Furthermore, the feature projection matrix can be applied to the multidimensional spatial feature parameters to generate a dimensionality-reduced feature space. The dimensionality-reduced feature space can be a set of low-dimensional feature representations obtained after transformation by the feature projection matrix, which can be used as input for significance testing and redundancy analysis.

[0102] Based on the optimal feature selection scheme, a feature projection matrix is ​​constructed, which can be achieved by arranging the unit basis vectors corresponding to the selected features in order to form a sparse matrix. Furthermore, this operation can be implemented by directly indexing the original feature columns using a Boolean mask, or by generating a sparse CSR format matrix to save storage, thus achieving lossless structured dimensionality reduction while preserving the original feature semantics. The original multidimensional spatial feature parameters are then transformed to obtain the dimensionality-reduced feature space, which can be achieved by left-multiplying the feature projection matrix by the original feature matrix to complete dimensionality compression.

[0103] The saliency test and redundancy analysis were performed on the dimensionality-reduced feature space to select a subset of core airspace situation features.

[0104] The significance test, a statistical method, is used to determine whether a feature significantly contributes to distinguishing operational modes. It can be used to eliminate statistically insignificant features, improving the reliability of the subset. In a specific embodiment, the significance test can assess the distribution differences of features across different modes using t-tests, ANOVA, or permutation tests. Redundancy analysis quantifies and eliminates residual redundancy within the dimensionality-reduced feature subset. It can be used to further refine the feature subset, avoiding implicit redundancy from affecting subsequent model performance. The core spatial situation feature subset can be a refined feature set obtained through optimization, projection transformation, significance testing, and redundancy analysis. It can be used as high-quality input for subsequent spatial operational mode definition and cluster analysis. Performing significance testing and redundancy analysis on the dimensionality-reduced feature space can involve sequentially performing statistical significance tests and feature dependency assessments. Selecting the core spatial situation feature subset can involve retaining features that pass the significance test and have redundancy below a threshold.

[0105] For example, in a scenario where emergency response and regular flight are mixed in airspace, the deep learning-based intelligent scheduling method for low-altitude resources in this embodiment can extract 30-dimensional features from dynamic operational data, including flight altitude change rate, mission emergency label, and control channel retransmission rate; construct a mixed-integer nonlinear programming model with the goal of maximizing the feature separability of fire-fighting and logistics equipment and minimizing the total number of features, and introduce binary variables to control the retention or removal of each feature; impose constraints: the correlation coefficient between any two features is <0.6, the information entropy is >1.2 bits, and at least two features are strongly correlated with "emergency priority"; use a branch and bound algorithm to solve the problem and select eight optimal features; construct an 8×30 projection matrix to map the original data to an 8-dimensional space; then use a t-test to remove one insignificant feature, and use mutual information analysis to merge two redundant features, finally obtaining a 5-dimensional core airspace situation feature subset, which clearly distinguishes three types of operational modes and provides highly discriminative input for subsequent differentiated scheduling.

[0106] In one embodiment, a network resource demand priority mapping table is generated based on the partitioning results, including:

[0107] Retrieve historical network resource scheduling records and performance data for the target airspace.

[0108] The historical network resource scheduling records for the target airspace can be log data of resource allocation decisions and their configuration parameters executed in the past within the target airspace. This data can provide the context of actual scheduling behavior and resource configuration, enabling the inference of resource usage patterns among different operating clusters. In this embodiment, the historical network resource scheduling records for the target airspace can be extracted from the historical database of the network management system, SDN controller, or orchestrator. Performance data can be a set of performance indicators reflecting the actual execution effect of historical scheduling strategies. This data can be used to evaluate the quality of historical scheduling and support the objective assignment of resource demand attribute tags. Furthermore, performance data can be collected through network probes, device telemetry, or user feedback, including task completion rate, latency compliance rate, and resource utilization rate.

[0109] Based on historical network resource scheduling records and performance data, each airspace operation cluster is assigned a resource demand attribute label.

[0110] The resource requirement attribute label can be a semantic identifier attached to a spatial operational cluster, describing its typical resource requirement characteristics. It can be used to map an abstract operational cluster to a scheduling unit with explicit resource semantics. In a specific embodiment, the resource requirement attribute label can be automatically labeled based on historical scheduling records and performance data, through rule matching or cluster center analysis.

[0111] Assigning resource requirement attribute labels to each segmented airspace operational cluster can be achieved by clustering and summarizing the resource usage characteristics of devices belonging to that cluster in historical data and matching them with a predefined label library. Furthermore, assigning resource requirement attribute labels to each segmented airspace operational cluster can be accomplished by automatically labeling them with "high reliability requirement" tags based on the QoS default frequency of devices within the cluster, or by matching tags based on task metadata keywords. This enables a mapping from operational clusters to resource semantics, enhancing interpretability.

[0112] Based on resource demand attribute tags, we analyze the resource consumption patterns and service quality requirements of low-altitude equipment within each airspace operating cluster, and quantify the cluster-level resource demand indicators.

[0113] Resource consumption patterns can be typical resource usage patterns of devices within a specific airspace operating cluster in terms of communication, computing, and spectrum dimensions. These patterns can reveal commonalities in group resource behavior and support cluster-level demand quantification. In this embodiment, resource consumption patterns can be statistically modeled or analyzed using historical resource usage data of devices within the cluster. Quality of Service (QoS) requirements can be the minimum guarantee requirements for network performance of the services carried by the airspace operating cluster. These requirements can define the bottom-line constraints for resource allocation and influence priority determination. Furthermore, QoS requirements can originate from task SLA agreements, industry standards, or user-declared QoS parameters (such as maximum latency and minimum throughput). For example, QoS requirements can include, but are not limited to, latency-sensitive QoS requirements, reliability-sensitive QoS requirements, and throughput-sensitive QoS requirements. Resource consumption patterns and QoS requirements together constitute the input basis for cluster-level resource demand indicators. Cluster-level resource demand indicators can be a quantitative representation of the overall resource demand intensity and QoS requirements of the airspace operating cluster. These indicators can be used as the numerical basis for priority ranking and weighted evaluation, achieving comparability between heterogeneous clusters.

[0114] Based on predefined business criticality levels and resource urgency rules, cluster-level resource demand indicators are prioritized to determine the priority of multi-level resource demands.

[0115] The business criticality level can be a predefined classification system of business type importance, which can be used to assign differentiated scheduling weights to different task types (such as medical emergency vs. commercial delivery). In this embodiment, the business criticality level can be a discrete level based on aviation management regulations, operator strategies, or public safety standards. For example, the business criticality level can include, but is not limited to, emergency public service level, passenger transport level, and general commercial service level. The resource urgency rule can be a judgment criterion that dynamically determines the urgency of resource allocation based on the current network status, which can be used to introduce dynamic risk perception and make the priority environmentally adaptable. Furthermore, the resource urgency rule can be triggered by expert rules or a lightweight learning model based on real-time indicators such as link interruption risk and buffer occupancy rate. The multi-level resource demand priority can be the result of hierarchical ranking of cluster-level indicators by combining business criticality and resource urgency, which can be used to form a structured priority sequence to guide subsequent weighted modeling. In a specific embodiment, the multi-level resource demand priority can use a multi-attribute decision method (such as TOPSIS, hierarchical analysis) to fuse and rank the priorities of each dimension. The multi-level resource demand priority serves as one of the key inputs to the weighted evaluation model.

[0116] Based on predefined business criticality levels and resource urgency rules, cluster-level resource demand indicators are prioritized. This can be achieved by using business level as a static weight and urgency rules as a dynamic adjustment factor, resulting in a comprehensive scoring and ranking. Furthermore, prioritizing cluster-level resource demand indicators based on predefined business criticality levels and resource urgency rules can be achieved using weighted summation or by using a fuzzy logic inference engine to fuse multi-source priority signals, thus enabling the priorities to possess both task semantics and environmental awareness capabilities. Determining multi-level resource demand priorities can be achieved by dividing the ranking results into several discrete priority levels (such as high / medium / low or levels 1-5).

[0117] By combining multi-level resource demand priorities and cluster-level resource demand indicators, a weighted evaluation model is constructed, and a network resource demand priority mapping table is generated accordingly.

[0118] The weighted evaluation model can be a mathematical function that integrates multi-level resource demand priorities and cluster-level resource demand indicators. It can be used to output continuous or discrete priority scores, supporting the construction of a mapping table. In this embodiment, the weighted evaluation model can construct a linear or nonlinear combination model, with weights optimized backward from historical performance or set by experts. The weighted evaluation model directly generates the numerical basis for the network resource demand priority mapping table. Constructing a weighted evaluation model by integrating multi-level resource demand priorities and cluster-level resource demand indicators can be achieved by designing a function that integrates priority levels and quantitative indicators into a unified evaluation score. Furthermore, constructing a weighted evaluation model by integrating multi-level resource demand priorities and cluster-level resource demand indicators can be achieved by constructing a neural network regression model using historical performance as a supervisory signal to train the weights, or by using an interpretable linear model, where the final score of the interpretable linear model = ×Priority coding+ ×Resource indicators, Weighting coefficients, representing the priority of multi-level resource requirements, are used to adjust the importance of priority codes (such as discrete or hierarchical semantic tags like task criticality and security level) in the comprehensive evaluation. Higher weights indicate that the system is more inclined to prioritize the resource requirements of high-priority tasks or equipment. The weighting coefficients, representing cluster-level resource demand indicators, are used to adjust the influence of quantified resource indicators (such as continuous numerical indicators like bandwidth requirements, computational load, and latency constraints) in the comprehensive evaluation. Higher weights indicate that the system is more inclined to allocate resources based on actual resource consumption intensity, thus generating a comprehensive priority representation that balances semantic importance and resource intensity.

[0119] Based on this, a network resource demand priority mapping table is generated, which can establish key-value pairs between each airspace running cluster ID and its corresponding weighted evaluation score or priority level. Furthermore, this network resource demand priority mapping table can be output as an in-memory hash table for use by the online scheduling module, or persisted as a JSON configuration file for synchronous use by the digital twin platform. This forms a quickly queryable scheduling prior knowledge base, supporting real-time resource allocation.

[0120] For example, in a scenario where urban emergency response and routine logistics coexist in mixed airspace, the deep learning-based intelligent scheduling method for low-altitude resources in this embodiment could be as follows: An urban low-altitude airspace simultaneously hosts emergency rescue eVTOLs, delivery drones, and municipal inspection drones. The system retrieves historical scheduling records from the past three months and finds that while medical drones are few in number, 99% of their missions require end-to-end latency <20ms and have never been downgraded; delivery drones frequently experience buffer overflows during peak hours; and inspection drones have a high tolerance for bandwidth fluctuations. Based on this, the system assigns the labels "ultra-low latency critical," "high throughput bursty," and "elastic bandwidth" to the three airspace clusters, respectively. Further analysis shows that the medical cluster uses an average of 50Mbps of bandwidth but has a QoS default rate of 0%, while the delivery cluster uses an average of 30Mbps but has a peak default rate of 15%. Combining predefined rules—that manned emergency services belong to the "emergency public service level" and that the current airspace link vulnerability triggers the "resource urgency rule"—the system prioritizes the medical cluster to the highest level. Ultimately, the weighted evaluation model outputs a score of 0.95 for the medical cluster, 0.65 for express delivery, and 0.40 for inspection. The generated mapping table ensures that medical communication slices and edge computing power are prioritized when resources are scarce, while express delivery tasks can be moderately delayed, and compressed transmission is enabled for inspection.

[0121] In one embodiment, the network resource demand priority mapping table is traversed to dynamically calculate the network resource demand of the low-altitude equipment group and determine the multi-dimensional dynamic resource demand weight coefficients, including:

[0122] Based on historical network resource scheduling records and performance data, identify key resource nodes and their topological relationships in the target airspace network.

[0123] By combining the network resource demand priority mapping table with key resource nodes, a dynamic resource demand prediction model is established to calculate the basic resource demand and simultaneously obtain multi-level resource demand priorities.

[0124] Based on the real-time division of airspace operation clusters and their dynamic data, the basic resource demand is calibrated online to obtain the real-time resource demand sequence;

[0125] A Bayesian optimization framework is used to perform regression analysis and weight fitting on the real-time resource demand sequence, and output multidimensional dynamic resource demand weight coefficients.

[0126] The target airspace network can be a collection of physical and logical network infrastructure within the target airspace, consisting of communication base stations, edge servers, spectrum access points, etc. It can serve as a platform for resource scheduling, and its structural characteristics directly affect the feasibility and efficiency of resource allocation. Key resource nodes can be core facility units in the network with high load, high connectivity, or irreplaceability. They can be used to identify performance bottlenecks or security critical points in the network, guiding resource demand modeling to focus on key areas. Topological relationships can be the connection structure and dependencies between key resource nodes, used to characterize resource flow paths and constraints, providing structural context for demand forecasting.

[0127] Based on historical network resource scheduling records and performance data, key resource nodes and their topological relationships in the target airspace network can be identified. This can be achieved by analyzing historical load, fault records, and connection logs to locate high-influence nodes and reconstruct the network graph. Furthermore, this operation can be implemented by using PageRank or betweenness centrality algorithms to identify key nodes, combining this with SNMP topology discovery to construct the graph structure, or by inferring implicit topological relationships from traffic logs through self-supervised learning of graph neural networks. This allows the abstract network to be visualized as a structured topology, embedding physical constraints into resource prediction.

[0128] A dynamic resource demand prediction model can be a resource demand estimation model that integrates running cluster priorities, network topology, and historical performance. It can be used to generate basic resource demands that take into account network structure constraints, improving prediction accuracy. In an exemplary embodiment, the dynamic resource demand prediction model can employ a graph neural network or a spatiotemporal sequence model to jointly encode the priority mapping table and node topology. Furthermore, the inputs to the dynamic resource demand prediction model include a network resource demand priority mapping table and the topology of critical resource nodes, and the output is the basic resource demand.

[0129] The basic resource requirement can be an estimate of the initial total resources required by each key node under the current airspace operating mode and network structure. This estimate can serve as a benchmark for online calibration, avoiding noise sensitivity caused by relying entirely on real-time data. In one specific embodiment, the basic resource requirement can be derived from a dynamic resource requirement prediction model based on historical patterns and priorities. Furthermore, the basic resource requirement can be used for calibration based on real-time segmentation of airspace operating clusters and their dynamic data.

[0130] Synchronously acquiring multi-level resource demand priorities can be achieved by querying the network resource demand priority mapping table in parallel during the prediction process to obtain the corresponding priorities. Furthermore, this operation can be implemented by using priorities as input feature channels to the model, or by introducing priority weighting terms into the loss function, thus ensuring strict alignment between priority information and demand calculation. Combining the network resource demand priority mapping table with key resource nodes, a dynamic resource demand prediction model can be established. This can be achieved by using priority labels as node attributes, and inputting them into a graph neural network along with the topology graph for demand prediction. Further, this operation can be achieved by constructing a heterogeneous graph (nodes include device clusters and resource nodes, edges represent resource request relationships), or by using an attention mechanism to weight the influence of different priority clusters on key nodes, thereby enabling joint modeling of operational modes and network structure and improving prediction fidelity.

[0131] The calculated basic resource requirements can be the baseline resource requirements of each key node in the current modality, output by the model's forward inference. Furthermore, this operation can be achieved by outputting the independent requirements for each dimension (bandwidth, computing power, spectrum), or by generating a resource requirement covariance matrix to characterize uncertainty, thus providing a stable prediction starting point and reducing the variance of purely real-time estimation. The real-time partitioned spatial domain clusters can be the grouping results of equipment groups obtained through sparse clustering at the current moment, which can be used to reflect the instantaneous spatial domain situation and drive dynamic adjustments to resource requirements. Furthermore, the real-time partitioned spatial domain clusters can be combined with dynamic data to generate real-time resource requirement sequences.

[0132] Dynamic data can be real-time collected information on the status of low-altitude equipment and network load, which can be used to provide signals of current environmental disturbances and support immediate corrections to demand forecasts. Furthermore, dynamic data can be combined with real-time segmented airspace operation clusters to generate real-time resource demand sequences.

[0133] Based on the real-time segmentation of airspace operational clusters and their dynamic data, online calibration of basic resource demand can be performed. This can involve calculating the deviation between the current cluster distribution and historical patterns, and dynamically adjusting the basic demand. Furthermore, this operation can be achieved by using Kalman filtering to fuse basic predictions and real-time observations, or by using a lightweight LSTM to incrementally correct short-term demand trends, thereby enabling rapid response to sudden scenarios (such as dense formations or link interruptions) and improving timeliness. The real-time resource demand sequence can be multi-dimensional time-series data of resource demand that evolves over time after online calibration. It can be used as input to a Bayesian optimization framework to characterize the dynamic trajectory of current resource demand. In a specific embodiment, the real-time resource demand sequence can fuse basic resource demand and dynamic data deviations to form a short-term sliding window sequence. Furthermore, the real-time resource demand sequence can be used by a Bayesian optimization framework for regression analysis and weight fitting.

[0134] Obtaining real-time resource demand sequences can be achieved by organizing the calibrated demand quantities into a sliding sequence within a time window. Furthermore, this operation can be implemented by constructing a fixed-length time series tensor for training a surrogate model, or by using an event-driven approach to update the sequence only when significant changes occur. This provides temporal context for Bayesian optimization, capturing dynamic evolution patterns. The Bayesian optimization framework can be a global optimization method based on a probabilistic surrogate model and a collection function. It can be used to efficiently handle high-dimensional, noisy, and limited-sample weight fitting problems, outputting robust multidimensional dynamic resource demand weight coefficients.

[0135] The regression analysis and weight fitting module can be a functional unit integrating a Bayesian optimization framework, used to transform nonlinear, multidimensional demand sequences into interpretable weight coefficient vectors. Furthermore, this module can receive real-time resource demand sequences and output multidimensional dynamic resource demand weight coefficients. Using a Bayesian optimization framework to perform regression analysis and weight fitting on the real-time resource demand sequence can involve taking the demand sequence as input, optimizing to minimize prediction error, and iteratively fitting the weight coefficients. The output of multidimensional dynamic resource demand weight coefficients can be achieved by extracting the optimal weight vector after Bayesian optimization convergence as the final output. Further, this operation can be achieved by outputting a weight distribution with confidence intervals for subsequent robust optimization, or by normalizing the weights to meet the input requirements of the planning solver, thereby generating high-quality weight inputs that can be directly used in mixed-integer nonlinear programming.

[0136] For example, in a scenario of low-altitude resource rescheduling under a sudden emergency, the deep learning-based intelligent low-altitude resource scheduling method in this embodiment could be as follows: A sudden fire breaks out in a city center, multiple firefighting eVTOL drones are scrambled, while the original logistics drone fleet is still operating. The system first identifies edge servers and 5G micro-stations in the area as key resource nodes based on historical data and constructs their topology map; combined with the existing network resource demand priority mapping table (firefighting tasks are the highest priority), the dynamic resource demand prediction model outputs the basic resource demand: edge computing power 80%, spectrum bandwidth 60%; however, real-time partitioning shows that the number of firefighting clusters suddenly increases by 300%, and dynamic data indicates a sharp drop in link SINR. The system immediately calibrated the basic demand online, generating a real-time resource demand sequence for the past 5 seconds, showing that the computing power demand surged to 95% and fluctuated drastically. The Bayesian optimization framework took this sequence as input and, after 3 rounds of iteration, output multi-dimensional dynamic resource demand weight coefficients: [communication: 0.45, computation: 0.40, spectrum: 0.15], significantly improving the computation weight to support AI analysis tasks at the fire scene. These coefficients were then input into a mixed-integer programming solver to ensure that the edge server prioritized video analysis offloading for firefighting drones, while logistics drones were temporarily reassigned to suboptimal frequency bands.

[0137] In one embodiment, the multidimensional dynamic resource demand weight coefficients are input into a mixed-integer nonlinear programming solver for high-dimensional feature reduction and optimization, resulting in a dimensionality-reduced resource demand feature vector. Based on the resource demand feature vector, a deep reinforcement learning algorithm is used to construct a resource scheduling strategy optimization model, including:

[0138] Based on the multidimensional dynamic resource demand weight coefficients and network topology constraints, feasible resource scheduling paths are enumerated to generate a set of candidate scheduling paths.

[0139] A resource scheduling state space is constructed using resource utilization, allocation latency, and load balancing as state indicators.

[0140] A resource scheduling action space is constructed by taking the action of selecting a path from the set of candidate scheduling paths and allocating resources.

[0141] We design a reward function that integrates multiple objective optimization factors and use a deep deterministic policy gradient algorithm to enable the agent to learn the optimal scheduling policy through interactive training in the state space and action space, and finally converge to obtain the resource scheduling policy optimization model.

[0142] Among these, a feasible resource scheduling path can be a complete resource configuration scheme that allocates communication links, computing nodes, and spectrum resources to low-altitude devices under network topology constraints. This scheme can be used to form the basic unit of the reinforcement learning action space and ensure the physical feasibility of the policy output. In an exemplary embodiment, a feasible resource scheduling path can be generated based on a dimensionality-reduced resource demand feature vector and network topology through a constraint satisfaction problem solver or heuristic search. The candidate scheduling path set can be a finite set of all feasible resource scheduling paths in the current state. This set can be used to discretize the infinite or high-dimensional action space into a manageable scale and improve the training efficiency of reinforcement learning. Resource utilization can be a ratio of allocated resources to total available resources. This ratio can be used to reflect the system's resource utilization efficiency and serve as one of the core dimensions of the state space. Furthermore, resource utilization can include, but is not limited to, one or more of spectrum occupancy, edge server CPU utilization, and backhaul link bandwidth utilization. Allocation latency can be the end-to-end time from the issuance of a resource request to the completion of allocation. This latency can be used to characterize the scheduling response speed and directly affect service continuity in highly dynamic scenarios.

[0143] Load balancing can be a quantitative indicator of the uniformity of load distribution among key resource nodes, and can be used to avoid local overload leading to system bottlenecks and improve overall robustness. In a specific embodiment, load balancing can measure the load difference between nodes using statistics such as variance, entropy, or maximum-minimum ratio. The resource scheduling state space can be a multi-dimensional system state representation composed of indicators such as resource utilization, allocation latency, and load balancing, and can be used to provide agents with comprehensive environmental observation and support policy decisions. In an exemplary embodiment, the resource scheduling state space can fuse the dimensionality-reduced resource demand feature vector with real-time network KPIs to encode a state vector. Furthermore, the resource scheduling state space and the resource scheduling action space jointly define the reinforcement learning problem.

[0144] The resource scheduling action space can be a discrete set of actions that select a path from a set of candidate scheduling paths as a resource allocation instruction. It can be used to structure scheduling decision outputs and ensure that actions are executable and conform to constraints. For example, the resource scheduling action space can map each feasible path to a unique action ID, forming a finite set of actions. In one specific embodiment, the resource scheduling action space is directly defined by the set of candidate scheduling paths. The multi-objective optimization factor can be multiple performance objective weight parameters used to construct the reward function, which can guide the agent to find the Pareto optimal equilibrium point in multi-objective conflicts. The reward function can be a mathematical function that maps system states and actions to scalar feedback signals, which can drive the agent to learn policies and make its behavior tend towards the global optimization objective. For example, the reward function can be designed as a nonlinear function of a weighted combination of multi-objective optimization factors, including immediate rewards and long-term discounted rewards. In one specific embodiment, the reward function guides the policy update direction of the Deep Deterministic Policy Gradient Algorithm (DDPG).

[0145] The formula for the reward function is as follows: In the formula, The immediate reward obtained by the agent (scheduler) after performing an action at time step t is the direct signal learned by the DDPG algorithm. The agent's goal is to maximize the expected cumulative reward in the future by adjusting its policy. , , These are weighting coefficients or preference parameters used to balance the importance and priority of the three optimization objectives in the total reward. Their values ​​are typically set by domain experts based on actual business needs and satisfy the following conditions: + + = 1, The larger the value, the more the system prioritizes resource utilization. The larger the value, the more the system prioritizes low latency. The larger the value, the more the system values ​​load balancing. Let t be the overall system resource utilization rate (the ratio of used resources to total resources), (1 - The lower the utilization rate, the higher the reward for that item. This encourages agents not to over-utilize resources, reserving resources for potential future surges in tasks to avoid system overload. However, utilization cannot be too low, otherwise it will lead to resource waste, so it needs to be optimized in conjunction with other objectives (such as load balancing). The average latency for resource allocation or task transmission at time step t. It is an exponentially decaying function. At that time... The smaller the value, the closer it is to 1 (higher reward); at that time, the delay The larger the value, the closer this value is to 0 (lower reward). This directly encourages the agent to choose the path with lower latency. This is a time delay normalization parameter or temperature parameter used to adjust the sensitivity of the time delay to changes in reward. A smaller τ value indicates a shorter time delay. Even a small increase in τ can lead to a sharp drop in the reward term and a more severe penalty. A larger τ value results in a more gradual response of the reward term to changes in time delay. Its function is to allow the time delay term to... It becomes a dimensionless value, making it easy to add to other rewards. The load balancing efficiency at time step t can be represented by the standard deviation or variance of the utilization of all nodes. The more balanced the load, the better. The smaller the value, the higher the reward. This encourages agents to distribute tasks reasonably across all available nodes, avoiding hotspots where some nodes are overloaded while others are idle, thereby improving the overall reliability and throughput of the system.

[0146] The Critic loss formula for the DDPG optimization objective is as follows:

[0147]

[0148] In the formula, Here is the loss function for the Critic network, with parameters as follows: The difference between the predicted value and the target value is quantified, and the network parameters are updated by minimizing L through gradient descent. For the expected value calculation, a statistical average based on empirical data is used. For the current state of the Critic network Next action Value prediction; To perform the action The immediate rewards obtained afterward, such as resource utilization rewards, latency rewards, etc. As a discount factor, it weighs the immediate rewards against the future rewards. The larger the scale, the more emphasis is placed on long-term future benefits; For the target Critic network, a lagging replica of the main network; To perform the action The next state after that is used as the basis for calculating the valuation of future rewards; For the target Actor network (a lagging copy of the main Actor) based on The output action provides the target Critic with the optimal action candidate for the next moment, which is used to calculate the estimate of future rewards.

[0149] An agent can be a deep reinforcement learning proxy entity that executes resource scheduling decisions, used to perceive the environment in the state space and select the optimal scheduling path in the action space. In an exemplary embodiment, the agent can be implemented using an Actor-Critic neural network architecture and deployed in a scheduling controller. Furthermore, the agent is trained through interaction with a digital twin environment, ultimately outputting an optimal scheduling policy. The optimal scheduling policy can be a state-action mapping function learned by the agent after convergence, used to achieve real-time optimal resource path selection for any input state. For example, the optimal scheduling policy can be obtained by optimizing the policy network parameters through a large amount of interaction experience using the DDPG algorithm. In a specific embodiment, the optimal scheduling policy is encapsulated as the core of a resource scheduling policy optimization model. Based on multi-dimensional dynamic resource demand weight coefficients and network topology constraints, feasible resource scheduling paths are enumerated, which can generate all legal resource configuration combinations while satisfying hard constraints such as link capacity, computational limits, and spectrum availability. Furthermore, this operation can be achieved by using a constraint programming (CP) solver system to generate paths that satisfy all constraints, or by using a graph traversal algorithm to search for end-to-end paths that meet QoS requirements on the resource topology graph, thereby ensuring that subsequent scheduling actions are physically feasible and avoiding ineffective exploration.

[0150] Generating a candidate scheduling path set can be achieved by deduplicating, clustering, or sampling the importance of the enumerated feasible paths to form a compact candidate set. Furthermore, this operation can be implemented through hierarchical clustering based on path similarity and retaining the cluster center path, or by using importance sampling to retain paths with high expected returns and eliminate inefficient options, thereby reducing the dimensionality of the action space and improving the stability and convergence speed of reinforcement learning training. Constructing a resource scheduling state space using resource utilization, allocation latency, and load balancing as state indicators can be achieved by normalizing the three types of indicators, concatenating them into a fixed-length vector, and then fusing the dimensionality-reduced resource demand features. Further, this operation can be achieved by using graph embedding methods to aggregate node-level indicators into a global state vector, or by introducing a sliding window historical state to enhance temporal awareness, thereby forming a comprehensive and compact environmental representation and supporting accurate agent decision-making.

[0151] The resource scheduling action space is constructed by selecting a path from a set of candidate scheduling paths to allocate resources. This can be achieved by mapping each candidate path to a discrete action ID to form a finite action set. Furthermore, this operation can be implemented by using one-hot encoding to represent action selection or by employing action embedding vectors to support generalization to unseen paths, thus transforming the complex resource allocation problem into a standard reinforcement learning action selection problem. A reward function that integrates multi-objective optimization factors is designed. This can be achieved by constructing a weighted combination function that includes resource utilization gain, latency penalty, load balancing reward, etc. Further, this operation can be achieved by designing a piecewise linear reward function (providing positive incentives within a threshold and imposing penalties for exceeding the threshold) or by introducing an adversary learning mechanism to dynamically adjust the weights of each objective to cope with environmental changes. This can guide the agent to learn a balanced strategy in multi-objective conflicts and avoid optimization based on a single metric.

[0152] By employing a deep deterministic policy gradient algorithm, the agent learns the optimal scheduling policy through interactive training in the state and action spaces. This can be achieved using an Actor-Critic architecture (Actors output deterministic actions, Critics evaluate Q-values) and stable training with the target network via experience replay. Furthermore, this operation can be mitigated by using dual-delayed DDPG (TD3) to reduce Q-value overestimation, or by combining Priority Experience Replay (PER) to accelerate key sample learning, thus enabling efficient learning of stable policies in continuous-discrete mixed and high-dimensional nonlinear environments. The final convergence yields an optimized resource scheduling policy model, which can be achieved by saving the trained Actor network as the final model after the policy network loss and reward metrics have stabilized. Further, this operation can be achieved by adversarially robust fine-tuning the convergence policy, or by distilling a large model into a lightweight version suitable for edge deployment, thereby forming a deployable intelligent scheduling engine that supports real-time decision-making.

[0153] Taking multi-service scheduling of low-altitude corridors in high-density cities as an example, the intelligent scheduling method for low-altitude resources based on deep learning in this embodiment can be as follows: During peak air traffic hours in a certain city, the system first inputs the multi-dimensional dynamic resource demand weight coefficients into a mixed-integer nonlinear programming solver to obtain a 10-dimensional reduced feature vector; based on this vector and the topological constraints of 5G base stations-edge servers, 200 feasible scheduling paths are enumerated (such as eVTOL using millimeter-wave direct connection + edge AI inference, logistics drones using Sub-6GHz + cloud offloading, etc.), and 30 representative candidate paths are generated after clustering; the state space consists of the current spectrum utilization rate (78%), the average allocation time, and the average allocation time. The algorithm is composed of metrics such as latency (12ms) and edge node load standard deviation (0.35); the action space is selected from 30 paths; the reward function integrates QoS guarantee (weight 0.5), energy efficiency (0.3) and fairness (0.2); the agent undergoes 100,000 interactive training steps in the digital twin platform, and the DDPG algorithm converges to the optimal strategy: when high load and increased latency are detected, some logistics tasks are automatically switched to suboptimal but idle edge node paths, while reserving high-priority direct connection channels for manned eVTOL; the final model achieves a 98% QoS compliance rate and a 22% improvement in resource utilization in real deployment, with no single point of overload.

[0154] In one embodiment, within a low-altitude network digital twin simulation platform, the dynamic allocation strategy for network resources is simulated and its compliance is verified. Based on the verification results, the strategy is optimized through closed-loop feedback, including:

[0155] Construct a high-fidelity digital twin simulation environment for low-altitude networks;

[0156] The high-fidelity low-altitude network digital twin simulation environment can be a virtual simulation system that performs high-precision modeling of real low-altitude networks at the physical, link, and application layers. It can provide a repeatable, interventionist, and high-fidelity policy verification sandbox, avoiding physical deployment risks. In this embodiment, the high-fidelity low-altitude network digital twin simulation environment can integrate 3D geographic information, wireless channel models (such as 3GPPUMi / UMa), device dynamics models, protocol stack implementation, and QoS policy engine construction. Furthermore, the high-fidelity low-altitude network digital twin simulation environment can serve as a loading and running platform for dynamic network resource allocation policies. In an exemplary embodiment, constructing the high-fidelity low-altitude network digital twin simulation environment can be achieved by integrating geographic information, propagation models, device dynamics, and network protocol stack to build a virtual image. For example, the high-fidelity low-altitude network digital twin simulation environment can be constructed by building a communication-mobility coupling model based on a network simulator and the SUMO co-simulation framework, or by using GPU-accelerated ray tracing to achieve centimeter-level channel simulation, thereby reproducing the multi-layered behavioral characteristics of real low-altitude networks and supporting high-confidence policy verification.

[0157] Load the dynamic allocation strategy for network resources in the simulation environment and simulate its operation, then collect network performance metrics after the strategy is executed.

[0158] The network performance metrics after policy execution can be a set of quantified performance data collected after executing the scheduling policy in a simulation environment. This data can serve as an objective basis for compliance judgment, reflecting the actual effectiveness of the policy. In one specific embodiment, the network performance metrics after policy execution can be recorded in real time through built-in probes or log interfaces, including end-to-end latency, throughput, packet loss rate, and resource utilization KPIs.

[0159] The collected network performance metrics are compared with predefined dynamic compliance threshold ranges to determine scheduling compliance.

[0160] The predefined dynamic compliance threshold range can be a set of compliance boundaries that adaptively adjusts according to task type, airspace status, or time context. This allows for context-aware, fine-grained compliance verification, overcoming the rigidity of static thresholds. In this embodiment, the predefined dynamic compliance threshold range can be generated by air traffic control rules, SLA protocols, and historical safe operation data, supporting condition-triggered updates. Furthermore, the predefined dynamic compliance threshold range can include, but is not limited to, one or more of the following: task-dependent latency thresholds, density-adaptive spectral interference limits, and load-sensitive computational overload boundaries. The scheduling compliance determination module can be a logical unit used to compare performance indicators with compliance thresholds and output determination results. It can be used to generate binary (compliant / non-compliant) or hierarchical determination signals to drive subsequent feedback processes.

[0161] If the judgment result is compliant, a positive reinforcement signal is generated to monitor the stability of the current strategy and make fine adjustments.

[0162] The positive reinforcement signal can be a feedback signal generated when the policy is compliant, used to enhance policy stability. It can guide the policy to fine-tune within the effective region, avoiding the disruption of verified good behavior. In this embodiment, the positive reinforcement signal can be injected into the reinforcement learning experience replay pool in the form of a Boolean flag or a reward value. Furthermore, the positive reinforcement signal can be a conservative optimization mechanism that triggers the stability monitoring and fine-tuning module. The stability monitoring and fine-tuning module can be a control component that performs parameter fine-tuning to improve robustness under the premise of policy compliance, and can be used to improve the policy's resistance to disturbances without changing its core behavior.

[0163] If the judgment result is non-compliant, a negative feedback signal is generated, the root cause analysis process is initiated, the abnormal allocation point in the strategy is located and a diagnostic report is generated, and the diagnostic report is used to guide the iterative correction of the strategy.

[0164] The negative feedback signal can be an anomaly signal generated when a policy is non-compliant, used to trigger a correction process. It can be used to initiate a root cause analysis process to prevent invalid policies from entering the physical network. In this embodiment, the negative feedback signal can include a structured event message containing the violation type, severity level, and associated cluster ID. Furthermore, the negative feedback signal can serve as the trigger input for the root cause analysis process. The root cause analysis process can be an automated diagnostic procedure used to locate the root cause of policy failure, accurately identifying the decision nodes leading to violations and avoiding blind corrections. In a specific embodiment, the root cause analysis process can combine counterfactual reasoning, sensitivity analysis, or graph traversal algorithms to trace the source of abnormal allocation. Further, the root cause analysis process can have a negative feedback signal as input and an abnormal allocation point and a diagnostic report as output. The abnormal allocation point can be a specific resource allocation decision item in the policy that causes performance indicators to exceed limits, used to provide actionable correction targets, such as "over-allocating the 6GHz band to cluster C3". In this embodiment, the abnormal allocation point can be located through contribution analysis or gradient backpropagation to pinpoint the action dimension with the greatest impact on the violation indicators.

[0165] Initiating the root cause analysis process can involve calling a diagnostic engine to perform backtracking analysis on the policy execution trajectory. In one specific embodiment, initiating the root cause analysis process can use Shapley values ​​to quantify the contribution of each allocation action to the violation, or construct a decision dependency graph and prune non-critical paths, thereby locating specific decision errors from apparent violations. Locating anomalous allocation points in the policy can involve identifying resource allocation decisions that play a dominant role in performance metric out-of-bounds behavior. For example, anomalous allocation points in the policy can be located by using gradient backpropagation to locate highly sensitive output dimensions in the policy network, or by comparing differences between compliant and non-compliant operating trajectories to locate divergence points, thereby providing precise correction targets and avoiding global retraining overhead. Generating a diagnostic report can involve formatting the root cause analysis results into a structured technical document. In one exemplary embodiment, generating a diagnostic report can output a Markdown format report containing charts and correction suggestions, or generate a machine-readable JSON-LD semantic description for automatic processing, thereby providing an interpretable and actionable basis for correction. Using the diagnostic report to guide iterative correction of the policy can involve converting diagnostic information into training samples or rule update instructions. In this embodiment, using the diagnostic report to guide iterative correction of the strategy can construct abnormal allocation points as adversarial examples and add them to the reinforcement learning training set, or update the upper limit of the weight of the corresponding cluster in the network resource demand priority mapping table, thereby enabling targeted repair and continuous evolution of the strategy.

[0166] Taking the scheduling verification of eVTOL commuter corridors in densely populated urban areas as an example, the deep learning-based intelligent scheduling method for low-altitude resources in this embodiment can be a new scheduling strategy for the planned deployment of low-altitude corridors in a city. The system loads this strategy in a high-fidelity digital twin environment, simulating a peak commuting scenario with 200 eVTOL aircraft. After the simulation, the peak CPU utilization of the edge server reached 98%, exceeding the dynamic compliance threshold (the threshold is 90% at the current density). The scheduling compliance judgment module outputs "non-compliant," triggering a negative feedback signal. The root cause analysis process locates the abnormal allocation point as: the strategy of "high-speed traversal cluster" excessively allocates real-time video analysis tasks to a single edge node. The diagnostic report recommends limiting the computational offload limit of a single node in this cluster. The strategy iteration and correction mechanism generates a new strategy accordingly. In the next simulation round, the CPU utilization drops to 85%, all indicators are compliant, the system generates a positive reinforcement signal and fine-tunes the strategy to enhance its ability to withstand sudden loads, and finally outputs a safe and reliable optimized strategy.

[0167] In addition, refer to Figure 2 To achieve the above objectives, the present invention also provides a low-altitude resource intelligent scheduling system based on deep learning, the system comprising:

[0168] The data acquisition module 10 is used to collect real-time status data and network load data of the low-altitude flight equipment group in the target airspace, and to construct a dynamic operation dataset of low-altitude equipment.

[0169] The situation clustering module 20 is used to perform multi-dimensional airspace situation perception and analysis on the dynamic operation dataset of the low-altitude equipment, generate an airspace operation mode feature set accordingly, and perform sparse clustering on the dataset based on the airspace operation mode feature set, and generate a network resource demand priority mapping table according to the partitioning result.

[0170] The demand modeling module 30 is used to traverse the network resource demand priority mapping table, dynamically calculate the network resource demand of the low-altitude equipment group, and determine the multi-dimensional dynamic resource demand weight coefficients; input the multi-dimensional dynamic resource demand weight coefficients into a mixed integer nonlinear programming solver for high-dimensional feature dimensionality reduction and optimization to obtain the dimensionality-reduced resource demand feature vector; and construct a resource scheduling strategy optimization model based on the resource demand feature vector using a deep reinforcement learning algorithm.

[0171] The strategy decision module 40 is used to input the real-time acquired dynamic operation dataset of the low-altitude equipment into the resource scheduling strategy optimization model, execute resource scheduling decisions, and output a dynamic allocation strategy for network resources.

[0172] The simulation optimization module 50 is used to simulate and verify the network resource dynamic allocation strategy in the low-altitude network digital twin simulation platform, and to perform closed-loop feedback optimization of the strategy based on the verification results, so as to generate the final network resource dynamic allocation optimization strategy, so as to realize the coordinated intelligent scheduling of communication links, computing resources and spectrum resources of low-altitude equipment networks in the target airspace.

[0173] Other embodiments or specific implementations of the deep learning-based intelligent scheduling system for low-altitude resources described in this invention can be found in the above-described method embodiments, and will not be repeated here.

[0174] Furthermore, to achieve the above objectives, the present invention also provides a low-altitude resource intelligent scheduling device based on deep learning. The device includes: a memory, a processor, and a low-altitude resource intelligent scheduling program based on deep learning stored in the memory and executable on the processor. The low-altitude resource intelligent scheduling program based on deep learning is configured to implement the steps of the low-altitude resource intelligent scheduling method based on deep learning as described above.

[0175] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a deep learning-based intelligent scheduling program for low-altitude resources, wherein when the deep learning-based intelligent scheduling program for low-altitude resources is executed by a processor, it implements the steps of the deep learning-based intelligent scheduling method for low-altitude resources as described above.

[0176] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for intelligent scheduling of low-altitude resources based on deep learning, characterized in that, The method includes: Collect real-time status data and network load data of low-altitude flight equipment groups within the target airspace to construct a dynamic operation dataset for low-altitude equipment; Multi-dimensional airspace situation awareness and analysis are performed on the dynamic operation dataset of the low-altitude equipment, and an airspace operation mode feature set is generated accordingly. The dataset is then sparsely clustered based on the airspace operation mode feature set, and a network resource demand priority mapping table is generated based on the partitioning results. The network resource demand priority mapping table is traversed to dynamically calculate the network resource demand of the low-altitude equipment group and determine the multi-dimensional dynamic resource demand weight coefficients. These multi-dimensional dynamic resource demands include communication links, computing resources, and spectrum resources of the low-altitude equipment network. The multi-dimensional dynamic resource demand weight coefficients are then input into a mixed-integer nonlinear programming solver for high-dimensional feature reduction and optimization to obtain a dimensionality-reduced resource demand feature vector. Based on this resource demand feature vector, a deep reinforcement learning algorithm is used to construct a resource scheduling strategy optimization model. The real-time acquired dynamic operation dataset of the low-altitude equipment is input into the resource scheduling strategy optimization model to execute resource scheduling decisions and output a dynamic allocation strategy for network resources. In the low-altitude network digital twin simulation platform, the dynamic allocation strategy of network resources is simulated and its compliance is verified. Based on the verification results, the strategy is optimized through closed-loop feedback to generate the final dynamic allocation optimization strategy of network resources, so as to realize the coordinated intelligent scheduling of communication links, computing resources and spectrum resources of low-altitude equipment networks in the target airspace.

2. The low-altitude resource intelligent scheduling method based on deep learning as described in claim 1, characterized in that, The process involves performing multi-dimensional airspace situational awareness and analysis on the low-altitude equipment dynamic operation dataset, generating an airspace operation modal feature set, and then performing sparse clustering on the dataset based on the airspace operation modal feature set, including: Airspace operation features are extracted from the dynamic operation dataset of the low-altitude equipment to obtain multi-dimensional airspace feature parameters; A mixed-integer nonlinear programming model is used to perform feature selection and dimensionality reduction on the multidimensional spatial feature parameters under constraints, resulting in a subset of core spatial situation features. Based on the aforementioned core airspace situation feature subset, the operational modes of the target airspace are analyzed and defined, and an airspace operational mode feature set is generated. Based on the aforementioned airspace operation mode feature set, a sparse subspace clustering algorithm is used to perform cluster analysis on the dynamic operation dataset of the low-altitude equipment to obtain multiple airspace operation cluster centers; Based on the multiple airspace operation cluster centers, the low-altitude equipment dynamic operation dataset is divided into several mutually exclusive airspace operation clusters; the resulting airspace operation clusters are output as the result of the sparse clustering partitioning.

3. The low-altitude resource intelligent scheduling method based on deep learning as described in claim 2, characterized in that, The method utilizes a mixed-integer nonlinear programming model to perform feature selection and dimensionality reduction on the multidimensional spatial feature parameters under constraints, resulting in a core spatial situation feature subset, including: A mixed-integer nonlinear programming problem model is constructed with the joint optimization objective of maximizing feature discriminative power and minimizing feature dimension, where a binary integer variable is introduced to represent whether a feature is selected. Nonlinear constraints are imposed on the planning problem model regarding the correlation between features, information entropy, and the relevance of business requirements; The branch and bound algorithm is used to solve the mixed-integer nonlinear programming problem to obtain the optimal feature selection scheme; Based on the optimal feature selection scheme, a feature projection matrix is ​​constructed to transform the original multidimensional spatial feature parameters, resulting in a dimensionality-reduced feature space. The saliency test and redundancy analysis are performed on the dimensionality-reduced feature space to select the core spatial situation feature subset.

4. The low-altitude resource intelligent scheduling method based on deep learning as described in claim 2, characterized in that, The step of generating a network resource demand priority mapping table based on the partitioning results includes: Retrieve historical network resource scheduling records and performance data for the target airspace; Based on historical network resource scheduling records and performance data, each of the divided airspace operation clusters is assigned a resource demand attribute label. Based on the resource demand attribute tags, the resource consumption patterns and service quality requirements of low-altitude equipment in each airspace operation cluster are analyzed, and cluster-level resource demand indicators are quantified. Based on predefined business criticality levels and resource urgency rules, the cluster-level resource demand indicators are prioritized to determine the priority of multi-level resource demands. By combining the multi-level resource demand priorities with the cluster-level resource demand indicators, a weighted evaluation model is constructed, and the network resource demand priority mapping table is generated accordingly.

5. The low-altitude resource intelligent scheduling method based on deep learning as described in claim 4, characterized in that, The process of traversing the network resource demand priority mapping table, dynamically calculating the network resource demand of the low-altitude equipment group, and determining the multi-dimensional dynamic resource demand weight coefficients includes: Based on the historical network resource scheduling records and performance data, identify key resource nodes and their topological relationships in the target airspace network. By combining the network resource demand priority mapping table with the key resource nodes, a dynamic resource demand prediction model is established to calculate the basic resource demand and simultaneously obtain the multi-level resource demand priority. Based on the real-time partitioned airspace operation clusters and their dynamic data, the basic resource requirements are calibrated online to obtain a real-time resource requirement sequence; The real-time resource demand sequence is subjected to regression analysis and weight fitting using a Bayesian optimization framework, and the weight coefficients of the multidimensional dynamic resource demand are output.

6. The low-altitude resource intelligent scheduling method based on deep learning as described in claim 5, characterized in that, The multidimensional dynamic resource demand weight coefficients are input into a mixed integer nonlinear programming solver for high-dimensional feature reduction and optimization to obtain a dimensionality-reduced resource demand feature vector. Based on the resource demand feature vector, a resource scheduling strategy optimization model is constructed using a deep reinforcement learning algorithm, including: The multidimensional dynamic resource demand weight coefficients are input into a mixed integer nonlinear programming solver for high-dimensional feature reduction and optimization to obtain a dimensionality-reduced resource demand feature vector. Based on the resource demand feature vector and network topology constraints, feasible resource scheduling paths are enumerated to generate a set of candidate scheduling paths. A resource scheduling state space is constructed using resource utilization, allocation latency, and load balancing as state indicators. A resource scheduling action space is constructed by taking the action of selecting a path from the set of candidate scheduling paths and allocating resources. A reward function integrating multi-objective optimization factors is designed, and a deep deterministic policy gradient algorithm is used to enable the agent to learn the optimal scheduling policy through interactive training in the state space and action space, and finally converge to obtain the resource scheduling policy optimization model.

7. The low-altitude resource intelligent scheduling method based on deep learning as described in claim 5, characterized in that, In the low-altitude network digital twin simulation platform, the dynamic allocation strategy for network resources is simulated and its compliance is verified. Based on the verification results, the strategy is optimized through closed-loop feedback, including: Construct a high-fidelity digital twin simulation environment for low-altitude networks; The network resource dynamic allocation strategy is loaded and simulated in the simulation environment, and network performance indicators are collected after the strategy is executed. The collected network performance indicators are compared with a predefined dynamic compliance threshold range to determine scheduling compliance. If the judgment result is compliant, a positive reinforcement signal is generated to monitor the stability of the current strategy and make fine adjustments. If the determination result is non-compliant, a negative feedback signal is generated, the root cause analysis process is initiated, the abnormal allocation point in the strategy is located and a diagnostic report is generated, and the diagnostic report is used to guide the iterative correction of the strategy.

8. A low-altitude resource intelligent scheduling system based on deep learning, characterized in that, The system includes: The data acquisition module is used to collect real-time status data and network load data of low-altitude flight equipment groups within the target airspace, and to construct a dynamic operation dataset of low-altitude equipment. The situation clustering module is used to perform multi-dimensional airspace situation perception and analysis on the dynamic operation dataset of the low-altitude equipment, generate an airspace operation mode feature set, and perform sparse clustering on the dataset based on the airspace operation mode feature set, and generate a network resource demand priority mapping table based on the partitioning result. The demand modeling module is used to traverse the network resource demand priority mapping table, dynamically calculate the network resource demand of the low-altitude equipment group, and determine the multi-dimensional dynamic resource demand weight coefficients. These multi-dimensional dynamic resource demands include communication links, computing resources, and spectrum resources of the low-altitude equipment network. The multi-dimensional dynamic resource demand weight coefficients are then input into a mixed-integer nonlinear programming solver for high-dimensional feature reduction and optimization, resulting in a dimensionality-reduced resource demand feature vector. Based on this resource demand feature vector, a deep reinforcement learning algorithm is used to construct a resource scheduling strategy optimization model. The strategy decision module is used to input the real-time acquired dynamic operation dataset of the low-altitude equipment into the resource scheduling strategy optimization model, execute resource scheduling decisions, and output a dynamic allocation strategy for network resources. The simulation optimization module is used to simulate and verify the compliance of the network resource dynamic allocation strategy in the low-altitude network digital twin simulation platform. Based on the verification results, the strategy is optimized through closed-loop feedback to generate the final network resource dynamic allocation optimization strategy, so as to realize the coordinated intelligent scheduling of communication links, computing resources and spectrum resources of low-altitude equipment networks in the target airspace.

9. A low-altitude resource intelligent scheduling device based on deep learning, characterized in that, The device includes: a memory, a processor, and a deep learning-based low-altitude resource intelligent scheduling program stored in the memory and executable on the processor, the deep learning-based low-altitude resource intelligent scheduling program being configured to implement the steps of the deep learning-based low-altitude resource intelligent scheduling method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a deep learning-based intelligent scheduling program for low-altitude resources, which, when executed by a processor, implements the steps of the deep learning-based intelligent scheduling method for low-altitude resources as described in any one of claims 1 to 7.

Images