AI-based low-altitude airspace information center collaborative management and control method

Abstract

The invention relates to the technical field of airspace control, and discloses an AI-based low-altitude airspace information center collaborative control method, which comprises six core steps of multi-source sensing fusion, abnormal behavior discrimination, dynamic risk assessment, environmental constraint re-planning, communication collaborative optimization and collaborative execution decision-making to construct a closed-loop intelligent decision-making mechanism. Microsecond-level time synchronization is realized through a Beidou or GPS satellite time service system, and the target identification precision is improved in combination with a weighted evidence fusion theory; an R-tree and space-time bounding box dual rule engine is combined to accurately detect abnormal flight; the collision risk between the aircrafts is quantified by using a three-layer stacked multi-head map attention network, and trajectory re-planning under environmental constraints is realized through a multi-dimensional environment voxel model; a ground and satellite hybrid networking architecture is adopted for communication collaborative optimization, so that terrain shielding and electromagnetic interference are avoided in management and control communication, global coverage and accurate adjustment are realized, and the accuracy, real-time performance and safety of management and control are ensured.

Description

Technical Field

[0001] This invention relates to the field of airspace control technology, and more specifically, to an AI-based collaborative control method for low-altitude airspace information hubs. Background Technology

[0002] With the deep integration of general aviation, drone technology, and urban air traffic, low-altitude airspace management places higher demands on the ability to accurately identify aircraft, conduct dynamic risk assessments, and optimize collaborative communication. Currently, low-altitude airspace management is gradually evolving from static control based on fixed rules to intelligent collaborative decision-making based on AI. Although traditional low-altitude control methods have made some progress in multimodal sensor data acquisition, basic trajectory monitoring, and simple anomaly alarms, they still have significant shortcomings in the closed-loop decision-making mechanism from sensing data to risk assessment, trajectory replanning and communication optimization, and finally to the collaborative control execution strategy. Therefore, they cannot effectively support real-time collaborative and intelligent control of all elements of low-altitude airspace.

[0003] Specifically, existing solutions largely rely on pre-defined no-fly zone rule bases or fixed airspace classification structures, lacking dynamic identification and adaptive adjustment mechanisms for high-risk aircraft interactions, complex environmental constraints, and changes in communication channel quality. Furthermore, after anomalies trigger emergency responses, actual execution feedback is rarely used to back-optimize multimodal perception fusion weights, risk assessment model parameters, or trajectory replanning strategies. This results in lags in response speed and decision-making accuracy when dealing with complex weather conditions, sudden electromagnetic interference, close-range aircraft conflicts, or illegal intrusions in low-altitude airspace, limiting the depth of intelligence and the breadth of security assurance in low-altitude airspace collaborative management.

[0004] Therefore, how to construct an AI-driven collaborative management and control method for low-altitude airspace information hubs, capable of processing multimodal sensing data, anomaly identification and risk assessment, trajectory replanning and communication optimization, and finally obtaining cross-departmental collaborative execution strategies, and how to continuously mine the evolution patterns of aircraft interaction risks, environmental constraints, and communication channel quality from real-time operational data, and dynamically optimize multi-source sensing fusion, intelligent risk prediction, environmental constraint trajectory replanning, adaptive adjustment of communication parameters, and multi-departmental emergency response strategies accordingly, has become an urgent technical problem to be solved in the field of low-altitude airspace management. Summary of the Invention

[0005] This invention provides an AI-based collaborative management and control method for low-altitude airspace information hubs, which solves the technical problems in existing technologies such as the lack of a closed-loop intelligent decision-making mechanism from multimodal perception to collaborative execution, the inability to dynamically and adaptively identify high-risk aircraft interactions and complex environmental constraints, and the lag in response speed and decision-making accuracy.

[0006] This invention provides an AI-based collaborative management and control method for low-altitude airspace information centers, comprising: Firstly, an AI-based collaborative management method for low-altitude airspace information centers includes: Initial sensing data in the low-altitude airspace is collected using multimodal sensors such as radar, optical sensors, ADS-B receivers, and infrared detectors to obtain a sensing dataset containing timestamps, spatial coordinates, velocity vectors, and signal characteristics. Spatiotemporal calibration is performed using the timestamp information in the perception dataset, and multi-source feature fusion is performed on spatial coordinates, velocity vectors and signal features to obtain an airspace situation dataset containing aircraft position, trajectory and identity features. Extract the identity features and trajectory information from the airspace situation dataset, perform identity association and rule comparison. If the airspace situation dataset at the current airspace location falls into a no-fly zone or is in a trajectory deviation situation, and the deviation value exceeds the preset deviation threshold (horizontal deviation of 50 meters or altitude deviation of 10 meters), then add the corresponding aircraft ID to the identifier list to generate an abnormal aircraft identifier list. The location information and behavioral characteristics of the abnormal aircraft identification list are mapped to a dynamic aircraft spatial distribution map, and the interaction relationship between aircraft is analyzed by a risk propagation model based on graph attention network to conduct dynamic risk assessment, so as to obtain a set of high-risk interaction pairs including risk level, impact range and related aircraft. For the high-risk interaction pairs of associated aircraft and their impact range, a multi-dimensional environmental voxel model and the current position of the aircraft are combined to predict the conflict trajectory through a trajectory prediction network and perform environmental constraint trajectory replanning with the help of a reinforcement learning replanner, so as to generate compliant trajectory adjustment instructions for aircraft that include avoidance strategies, time windows and three-dimensional waypoints. The three-dimensional path points and time window information in the compliant trajectory adjustment instruction are analyzed, the communication coverage of each path point is evaluated by a directional antenna array, and the communication channel quality is evaluated by combining a convolutional autoencoder. The frequency, power and modulation method are adaptively adjusted according to the channel quality to generate communication parameter optimization instructions that are synchronized with the trajectory adjustment. The avoidance strategy in the compliance trajectory adjustment instruction and the modulation method in the communication parameter optimization instruction are jointly encoded and classified. Combined with the risk level in the set, a list of normal and abnormal aircraft statuses is generated. Resource scheduling based on multi-agent game is performed on normal aircraft, and multi-department collaborative abnormal early warning response is triggered for abnormal aircraft, resulting in a low-altitude airspace collaborative control execution strategy that includes scheduling schemes and emergency plans.

[0007] Furthermore, initial sensing data in the low-altitude airspace is acquired using multimodal methods to obtain a sensing dataset, including: Initial sensing data for the low-altitude airspace is collected by multimodal sensors such as millimeter-wave radar, optical cameras, infrared thermal imagers, ADS-B receivers, and RF spectrum detectors deployed at key urban nodes. Configure each multimodal sensing device with a hardware clock that supports the IEEE-1588 precision time protocol or GPS authorization to achieve microsecond-level timestamp alignment of each sensor; It receives raw sensing data from various sensors along with their accompanying timestamps and location metadata, resulting in a sensing dataset containing timestamps, spatial coordinates, velocity vectors, radar echoes, video frames, infrared signals, and spectral features.

[0008] Furthermore, spatiotemporal calibration and multi-source feature fusion are performed on the perceptual dataset, including: The Kalman filter algorithm is used to perform state estimation and interpolation on the asynchronously arriving timestamps and spatial coordinates in the sensing data to generate intermediate data frames with a unified UTC time base and WGS-84 geographic coordinate system. Density clustering is performed on the radar echoes in the intermediate data frames to extract target contours, Gaussian mixture model is applied to the video frames to model the background and separate moving targets, and short-time Fourier transform is performed on the spectral features to extract spectral fingerprints. The preprocessed target contour, moving target, and spectral fingerprint are input into the data fusion layer using the weighted Dempster-Shafer evidence theory, and confidence-weighted fusion is completed at the target level. The resulting airspace situation dataset contains target ID, latitude and longitude, altitude, three-dimensional velocity vector, aircraft type, and confidence weights for each sensor.

[0009] Furthermore, based on the airspace situation dataset, identity association and rule comparison are performed. If the offset value exceeds a preset offset threshold, an abnormal aircraft identifier list is generated, including: Load the no-fly zone rule base stored in GeoJSON format and build an R-tree spatial index structure; load the registered flight plan database containing fields such as takeoff time window, route, altitude layer and operation subject and build a flight plan spatiotemporal bounding box index. Extract the target ID, latitude and longitude, altitude and aircraft type from the airspace situation dataset, perform an R-tree range query on the latitude and longitude coordinates of each aircraft, and determine whether it falls within the geometric boundary of the no-fly zone. If the aircraft has not entered a no-fly zone, the reference trajectory point sequence in the current flight plan is extracted, and the Euclidean distance between the current actual position and the nearest reference trajectory point in the airspace situation dataset is obtained. If the Euclidean distance exceeds a preset offset threshold in the horizontal direction, it is determined that the trajectory deviation exceeds the limit; Write the aircraft ID, event type, occurrence time, and original sensing data snapshot from the airspace situation dataset that meet the conditions of no-fly zone intrusion or excessive trajectory deviation into the abnormal aircraft identifier list. The list of anomalous aircraft identifiers is pushed to the log system for persistent storage and a drive-away command generator is triggered simultaneously to verify whether the anomalous aircraft has a legitimate communication link and identity authentication status. If it does not, it is marked as a high-risk target to generate an anomalous aircraft identifier list containing aircraft ID, deviation amount, and risk level.

[0010] Furthermore, based on the list of anomalous aircraft identifiers, the dynamic spatial distribution map of aircraft, and the risk propagation model based on graph attention networks, a dynamic risk assessment is performed to obtain a set of high-risk interaction pairs, including: Obtain 3D boundary grid data in CityGML format from airspace management geographic information, including building LOD2 models, terrain elevations, and airspace hierarchical structure; The position and velocity vectors of all aircraft in the airspace situation dataset are extracted and projected onto the three-dimensional boundary grid. Combined with the aircraft IDs and risk levels in the abnormal aircraft identifier list, a dynamic spatial distribution map of aircraft is constructed that is updated 10 frames per second. Traverse all pairs of aircraft in the dynamic aircraft spatial distribution map, calculate the Euclidean distance and relative velocity vector, and generate an N×N dimensional aircraft spacing matrix, where N is the total number of aircraft in the airspace; The aircraft spacing matrix, together with the aircraft type, physical size and maximum maneuver overload capability attributes in the airspace situation dataset, are encoded as graph node features and edge weights, and input into a graph attention network model composed of three stacked multi-head graph attention layers. The graph attention network model is trained under supervision based on multiple (5000) real or simulated close-range events in the historical conflict case library, and outputs a collision risk score of 0 to 1 for each pair of aircraft. Aircraft pairs with scores greater than a preset threshold (0.7) are marked as high-risk interaction pairs. Combined with the risk levels in the list of abnormal aircraft identifiers, a set of high-risk interaction pairs containing aircraft ID pairs, collision risk scores, and associated abnormal identifiers is obtained.

[0011] Furthermore, based on the associated aircraft and their impact range in the high-risk interaction set, combined with a multi-dimensional environmental element model and the current position of the aircraft, a trajectory prediction network is used to predict conflict trajectories and perform environmentally constrained trajectory replanning to generate compliant trajectory adjustment instructions, including: It receives echo intensity data from weather radar, interference power spectral density data from electromagnetic spectrum scanner, and three-dimensional static obstacle mesh data in CityGML format; The echo intensity data, interference power spectral density data and three-dimensional static obstacle mesh data are uniformly mapped to a 5×5×5 resolution three-dimensional voxel space through a voxelization engine, generating meteorological disturbance layer, electromagnetic interference layer and obstacle distribution layer respectively. In each layer, voxel datasets are recorded for horizontal wind speed, vertical wind speed, turbulence intensity, signal attenuation index, location of co-frequency interference source and static obstacle voxel occupancy status. Spatiotemporal interpolation is performed on the voxel dataset using a sliding time window. The evolution of the environmental state within a preset time (10 seconds) is extrapolated using linear prediction or LSTM time series models to obtain a multidimensional environmental voxel model. The wind speed gradient threshold, signal attenuation region coordinates, and static obstacle envelope are extracted from the multi-dimensional environmental element model as safety constraint parameters. Based on the associated aircraft IDs and impact ranges in the high-risk interaction set, the current position and speed in the airspace situation dataset are extracted and combined with safety constraint parameters and input into a trajectory prediction network based on the Transformer architecture. The trajectory prediction network encoder receives historical trajectory points within a preset time (10 seconds) in the past, and the decoder predicts multiple (900) trajectory points within a preset time (90 seconds) in the future to form conflict trajectory prediction. The predicted trajectory points in the conflict trajectory prediction are queried point by point to obtain the query results; The process involves querying the environmental parameters of the corresponding voxels in the multidimensional environmental voxel model point by point to determine whether there are any violations of wind speed gradient, signal strength, or obstacle occupancy constraints. If such violations exist, the trajectory replanner based on the DDPG algorithm is invoked to perform environmental constraint trajectory replanning under the constraints of collision risk score and impact range of the high-risk interaction pair set. This generates a new waypoint sequence that meets the environmental constraints, thereby generating an aircraft compliant trajectory adjustment instruction that includes avoidance strategy, time window, and three-dimensional path points.

[0012] Furthermore, based on the compliant trajectory adjustment instruction, a communication channel quality assessment is performed using a directional antenna array and a convolutional autoencoder, and communication parameters are adaptively adjusted to generate a communication parameter optimization instruction, including: Extract the three-dimensional path points and time windows from the aircraft's compliant trajectory adjustment instructions, and calculate the line-of-sight path between each waypoint and the nearest low-altitude private network base station using a three-dimensional city model to determine the optimal communication echo sampling time and the ground station receiving azimuth angle. According to the compliant trajectory adjustment instruction, at the sampling time, the base station antenna is activated by the directional antenna array so that its main lobe is aligned with the azimuth angle, receives the downlink communication signal of the aircraft, and generates the original echo signal stream with microsecond-level time and frequency markers; The original echo signal stream is input into a four-layer 1D convolutional autoencoder through a convolutional autoencoder. Multipath effect, co-channel interference and atmospheric noise features are extracted through reconstruction error to perform communication channel quality assessment and obtain a communication channel quality index vector including CINR, delay spread and Doppler shift. Based on the communication channel quality index vector, the aircraft's transmission power, modulation method, and retransmission strategy are adaptively adjusted according to a preset mapping table. Combined with the avoidance strategy and time window in the compliant trajectory adjustment command, a communication parameter optimization command synchronized with the trajectory adjustment is generated.

[0013] Furthermore, the compliant trajectory adjustment instructions and communication parameter optimization instructions are jointly encoded and classified to generate lists of normal and abnormal aircraft states. Multi-agent game-theoretic scheduling and multi-department collaborative anomaly early warning responses are then executed respectively to obtain a low-altitude airspace collaborative control and execution strategy, including: Extract the waypoint sequence, velocity profile and obstacle avoidance margin from the compliant trajectory adjustment command of the aircraft, as well as the transmit power, modulation mode and retransmission strategy from the communication parameter optimization command. Jointly encode the compliant trajectory adjustment command and the communication parameter optimization command into a 128-dimensional operating state feature vector, where the first 64 dimensions represent trajectory adjustment features and the last 64 dimensions represent communication parameter features. The operational status feature vector is input into an ensemble classifier consisting of three base models: XGBoost (trained gradient boosting decision tree), RandomForest (random forest), and LightGBM (lightweight gradient boosting machine) and classified. The ensemble classifier is trained based on historical normal and abnormal flight records and outputs a binary classification result of the aircraft's operational status through a weighted voting mechanism. Based on the binary classification results, and combined with the collision risk score in the high-risk interaction set and the risk level in the abnormal aircraft identifier list, a normal aircraft status list and an abnormal aircraft status list are generated. For each item in the abnormal aircraft status list, extract the aircraft ID, operating status feature vector and risk level, establish a triple relationship of aircraft-event-instruction through the Neo4j graph database, match the preset emergency response mode library, execute multi-department collaborative abnormal early warning response, and generate an emergency response plan that includes multi-department linkage actions, resource scheduling list and strict time window constraints, including public security, civil aviation, emergency management and other departments. The aircraft ID and operational status feature vector in the normal aircraft status list are input into the multi-agent game scheduling optimizer. Multi-agent game scheduling is performed to solve the globally optimal takeoff time slot, route assignment (selected from the predefined route library), and altitude layer configuration (allocated at 100-foot intervals) using Nash equilibrium to obtain the flight scheduling scheme. The emergency response plan and flight scheduling plan are loaded into the Apache-Airflow distributed task scheduling engine, decomposed into atomic operations, and distributed to the standard RESTful API interfaces of public security, civil aviation and emergency management units through message queues to obtain the low-altitude airspace collaborative control and execution strategy.

[0014] Furthermore, the step of inputting the running state feature vector into an ensemble classifier composed of three base models—XGBoost, RandomForest, and LightGBM—and classifying it includes: Extract a training sample set containing both normal and abnormal flight records from the historical flight record database; For each flight record in the training sample set, 128-dimensional features, including waypoint sequence, velocity profile, obstacle avoidance margin, transmit power, modulation method, and retransmission strategy, are extracted and labeled with category labels to obtain a labeled training dataset. The labeled training dataset is divided into a training set and a validation set in a 7:3 ratio. The training set is used to train three base models: gradient boosting decision tree, random forest, and lightweight gradient boosting machine. Hyperparameters are then tuned on the training set to obtain the optimized three base models. The accuracy, recall, F1 score, and AUC of the three base models are evaluated on the validation set, and the weights of the three base models are determined to be 0.4, 0.3, and 0.3 respectively based on the validation set performance. The operational status feature vector is input into the three base models respectively to obtain the probability prediction value of each base model. The probability prediction values ​​are weighted and summed according to the weighted voting weight to obtain the final classification probability. If the final classification probability is greater than 0.5, it is judged as abnormal operation class; otherwise, it is judged as normal operation class, so as to output the binary classification result of the aircraft's operational status.

[0015] Furthermore, the aircraft ID and operational status feature vector from the normal aircraft status list are input into a multi-agent game scheduling optimizer to perform multi-agent game scheduling, using Nash equilibrium to solve for the globally optimal takeoff slot, route assignment, and altitude layer configuration, including: Extract the aircraft ID, expected takeoff time, destination, range, aircraft performance parameters and operating status feature vector of each aircraft from the normal aircraft status list, construct a multi-agent game model, treat each aircraft as an independent agent and define its policy space, and obtain a policy space including takeoff time slot set, route set and altitude layer set; Define a utility function for each agent in the policy space. The utility function comprehensively considers takeoff time deviation cost, route length cost, altitude layer preference cost and conflict risk cost. Based on the current airspace capacity dynamic quota, set constraints to obtain a game model that includes utility function and constraints. The game model is solved for Nash equilibrium using an iterative optimal response algorithm. Each agent is initialized with a random combination of strategies. In each iteration, each agent selects the optimal strategy that maximizes its own utility function while the strategies of other agents are fixed, and updates it iteratively to obtain the Nash equilibrium strategy. Verify whether the Nash equilibrium strategy satisfies all constraints. If it does, output the globally optimal takeoff time slot, route assignment, and altitude layer configuration for each aircraft. Combine the globally optimal takeoff time slot, route assignment, and altitude layer configuration into a flight scheduling scheme to obtain a detailed flight plan that includes takeoff time, waypoint sequence, cruising altitude, and estimated arrival time.

[0016] Furthermore, for each item in the abnormal aircraft status list, the aircraft ID, operational status feature vector, and risk level are extracted. A triplet relationship between aircraft, events, and commands is established using the Neo4j graph database. This is matched against a pre-defined emergency response model library, and multi-departmental collaborative abnormal early warning responses are executed. This generates an emergency response plan that includes coordinated actions from multiple departments such as public security, civil aviation, and emergency management, a resource scheduling list, and strict time window constraints. Extract the aircraft ID, operational status feature vector, risk level, abnormal event type, and occurrence time for each item from the abnormal aircraft status list; Using the aircraft ID as the aircraft node, the abnormal event type as the event node, and the compliant trajectory adjustment command and communication parameter optimization command as the command node, a triple relationship of aircraft-event-command is established in the Neo4j graph database. The aircraft node and the event node are connected through the occurrence relationship, and the event node and the command node are connected through the trigger relationship. Based on the risk level and the type of abnormal event, the corresponding emergency response mode is matched from the preset emergency response mode library. The emergency response mode library contains response processes classified by risk level and event type, including Level I response mode, Level II response mode and Level III response mode. For Level I response mode, a mild response procedure is generated, which includes air traffic control issuing a warning and the aircraft performing autonomous avoidance. For Level II response mode, a moderate response procedure is generated, which includes air traffic control issuing a mandatory instruction, civil aviation coordinating airspace resources, and the aircraft performing mandatory avoidance. For Level III response mode, a high-altitude response procedure is generated, which includes air traffic control activating the emergency plan, management departments intervening in the control, civil aviation coordinating airspace resources, emergency management departments preparing for emergency rescue, and the aircraft performing an emergency landing. Based on the matched emergency response mode, multi-departmental collaborative abnormal early warning response is executed, generating a list of multi-departmental joint actions. The joint actions of the public security department include dispatching drone interception units and initiating ground control procedures; the joint actions of the civil aviation department include issuing airspace control notices and adjusting flight schedules; and the joint actions of the emergency management department include deploying emergency rescue teams and preparing medical rescue resources. Generate a resource scheduling list, specifying the number of personnel, equipment types, vehicle configurations, and communication channels to be scheduled for each department, and determine the resource deployment locations based on the current location and predicted trajectory of the abnormal aircraft; Strict time window constraints are set according to the urgency of the abnormal event to generate an emergency response plan that includes multi-departmental coordinated actions, resource scheduling lists, and time window constraints.

[0017] Secondly, a computer-readable storage medium is provided for storing computer-readable instructions that, when read by a computer, enable the execution of an AI-based low-altitude airspace information hub collaborative control method.

[0018] The beneficial effects of this invention are as follows: This invention solves the problem of ambiguous identity recognition caused by heterogeneous data asynchrony in traditional systems by using microsecond-level time synchronization of multimodal sensors and weighted evidence fusion theory; it accurately captures no-fly zone intrusions and trajectory deviations using a dual rule engine of R-tree and spatiotemporal bounding boxes, overcoming the omission defects of single rule comparison; the graph attention network models aircraft interactions as a dynamic graph structure, upgrading collision risk assessment from static distance thresholds to multi-dimensional attribute-driven intelligent scoring; the multi-dimensional environmental voxel model unifies meteorological, electromagnetic, and obstacle data into queryable voxel parameters, enabling trajectory replanning to have real-world environmental adaptability; the combined use of directional antenna arrays and convolutional autoencoders ensures strict synchronization between communication parameter adjustments and flight trajectory changes, avoiding communication interruptions; the integration of a classifier and the Neo4j graph database upgrades anomaly response from single-point alarms to structured contingency plans involving multi-department resource linkage; and the multi-agent game scheduler uses Nash equilibrium to solve for global optimality, avoiding chain conflicts caused by local scheduling. Therefore, this invention constructs a high-precision, robust, and executable intelligent control system for low-altitude airspace by employing techniques such as multimodal perception fusion, graph neural risk modeling, voxelized environmental constraints, communication-trajectory joint optimization, and multi-agent collaborative decision-making. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of a low-altitude airspace information central collaborative management method based on AI provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the low-altitude airspace information central collaborative management process provided in an embodiment of the present invention. Detailed Implementation

[0020] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0021] At least one embodiment of the present invention discloses an AI-based collaborative management and control method for low-altitude airspace information centers, comprising: like Figure 1 As shown, an AI-based collaborative management method for low-altitude airspace information centers includes the following steps: Step 1: Collect initial sensing data in the low-altitude airspace using multimodal methods to obtain a sensing dataset; Step 2: Perform spatiotemporal calibration and multi-source feature fusion on the perception dataset to obtain a structured spatial situation dataset; Step 3: Perform identity association and rule comparison based on the aforementioned spatial situation dataset; Step 4: Determine whether the airspace situation dataset at the current airspace location has fallen into a no-fly zone or is in a trajectory deviation situation, and the deviation value exceeds the preset deviation threshold. If so, add the corresponding aircraft ID to the identifier list to generate an abnormal aircraft identifier list; otherwise, return to the data acquisition step to continue acquiring the sensing dataset. Step 5: Based on the list of abnormal aircraft identifiers, the dynamic spatial distribution map of aircraft, and the risk propagation model based on graph attention networks, perform dynamic risk assessment to obtain a set of high-risk interaction pairs; Step 6: Based on the associated aircraft and their impact range in the high-risk interaction set, combined with the multi-dimensional environmental element model and the current position of the aircraft, predict the conflict trajectory through the trajectory prediction network and perform environmental constraint trajectory replanning to generate compliant trajectory adjustment instructions; Step 7: Based on the compliant trajectory adjustment instruction, perform communication channel quality assessment using a directional antenna array and a convolutional autoencoder, and adaptively adjust communication parameters to generate communication parameter optimization instructions; Step 8: Jointly encode and classify the compliant trajectory adjustment instructions and communication parameter optimization instructions to generate normal and abnormal aircraft status lists, and execute multi-agent game scheduling and multi-department collaborative abnormal early warning response respectively to obtain the low-altitude airspace collaborative control and execution strategy.

[0022] The above steps work together to provide an AI-based collaborative management method for low-altitude airspace information centers, the specific implementation of which is described below: like Figure 2As shown, this embodiment deploys a complete multimodal perception and intelligent decision-making infrastructure in a typical urban low-altitude airspace management scenario. First, millimeter-wave radar, optical cameras, infrared thermal imagers, ADS-B receivers, and RF spectrum detectors are installed at key urban nodes such as rooftops of high-rise buildings, communication towers, and traffic monitoring poles. These five types of sensors are physically fixed on the same pan-tilt structure, ensuring that the overlap of the center lines of each sensor's field of view on the horizontal plane is no less than 85%. They are connected to an edge computing gateway via a unified power supply module. The edge computing gateway integrates an IEEE-1588 Precision Time Protocol (PTP) master clock module or a GPS timing module, which is connected via coaxial cables to... The external trigger interface connected to each sensor ensures that all data frames output by the sensors carry microsecond-level timestamps generated by the same UTC time source, with time synchronization errors controlled within ±2 microseconds. Simultaneously, the edge computing gateway receives timing signals through the BeiDou Navigation Satellite System (BDS) or GPS Navigation Satellite System, using the 1PPS (one pulse per second) satellite timing signal as the UTC time reference source to ensure high-precision time synchronization is maintained even when the ground network is interrupted. The satellite timing accuracy reaches ±50 nanoseconds, meeting the time consistency requirements of the distributed multi-node sensing system.

[0023] The raw sensing data collected by the above five types of sensors includes the target point cloud coordinates and radial velocity output by millimeter-wave radar, the 1080P@30fps video stream output by optical camera, the 8-bit temperature matrix output by infrared thermal imager, the ICAO address code and flight status message parsed by ADS-B receiver, and the power spectral density sequence in the range of 9kHz-6GHz output by RF spectrum detector; the raw sensing data is transmitted via gigabit Ethernet to the multi-source sensing fusion unit deployed on the local edge server. The multi-source sensing fusion unit first performs time alignment processing on the asynchronously arriving raw sensing data: using the UTC timestamp as a reference, the data from each sensor are interpolated to a unified time grid with a time granularity of 10 milliseconds; then, a discrete-time Kalman filter is used to estimate the spatial state of each target. The state vector includes six dimensions: longitude, latitude, altitude, eastward velocity, northward velocity, and vertical velocity. The process noise covariance matrix Q is set as a diagonal matrix based on the urban low-altitude turbulence model, while the observation noise covariance R is dynamically adjusted according to the measured accuracy of each sensor; after completing the state estimation, DBSCAN density clustering is performed on the radar point cloud with a neighborhood radius of 3 meters and a minimum number of points of 5; a Gaussian mixture model (GMM) is used to model the background of the optical video frames with a mixture component of 5 and a learning rate of 0.01; and adaptive threshold segmentation is performed on the infrared image to extract the heat source wheel. The RF spectrum data is subjected to a 256-point Short-Time Fourier Transform (STFT) with a window length of 10 milliseconds and an overlap rate of 50% to extract spectral fingerprints such as center frequency, bandwidth, and modulation features. The above four types of feature vectors, together with the aircraft type code parsed by ADS-B, constitute the evidence body, which is input into the weighted Dempster-Shafer evidence theory fusion layer. The weight coefficients are dynamically allocated according to the historical accuracy of the sensors under the current weather conditions. For example, the weight of optical cameras is reduced and the weight of millimeter-wave radar and infrared are increased in rainy and foggy weather. The fusion layer outputs a structured airspace situation dataset. Each record contains the target ID (generated by the first appearance time + sensor combination hash), WGS-84 latitude and longitude, ellipsoidal height, three-dimensional velocity vector, aircraft type (such as multi-rotor, fixed-wing, eVTOL, etc.), confidence weight of each sensor, and total confidence after fusion.

[0024] The structured airspace situation dataset is pushed to the abnormal behavior discrimination unit in real time. This unit pre-loads a no-fly zone rule base stored in GeoJSON format, containing polygonal boundaries for government agencies, airport airspace, and major event venues. It also loads a registered flight plan database synchronized by the Civil Aviation Administration of China's UOM platform. Each plan includes the aircraft ID, takeoff / landing coordinates, planned trajectory point sequence (sampling interval 1 second), permitted altitude layer, and time window. The system uses the PostGIS spatial database engine to construct an R-tree spatial index, establishing a hierarchical bounding box structure for the no-fly zone polygons, achieving a query complexity of O(logn). For the registered flight plan, a spatiotemporal bounding box (STBBox) index is constructed for its trajectory point sequence, with each bounding box covering a horizontal range of ±100 meters, a height of ±20 meters, and a time range of ±5 seconds. For each flight in the structured airspace situation dataset... The aircraft first performs an R-tree range query to determine whether its current position falls within any no-fly zone polygon. If a true value is returned, it is immediately marked as a no-fly zone intrusion. If no intrusion occurs, the corresponding ID's trajectory is retrieved from the registration plan, and the Euclidean distance from the current position to the nearest point on that trajectory is calculated. The horizontal component uses the plane distance under Mercator projection, and the vertical component directly takes the height difference. When the horizontal offset exceeds 50 meters or the height deviation exceeds 10 meters and lasts for more than 3 seconds, it is determined that the trajectory deviation exceeds the limit. For aircraft that meet any of the abnormal conditions, their ID, event type, occurrence UTC time, and original perception snapshot (including the original frames of each sensor and the fusion results) are written into the abnormal aircraft identifier list and pushed to the Elasticsearch log system for persistent storage via the Kafka message queue, with a retention period of 180 days.

[0025] The dynamic risk assessment unit receives the structured airspace situation dataset and the list of anomalous aircraft identifiers. This dynamic risk assessment unit first obtains 3D geographic information data in CityGML-2.0 format from the City Information Modeling (CIM) platform, extracting building LOD2 models (including roof geometry and material properties), digital elevation models (DEM, 1-meter resolution), and airspace layering structures (e.g., 0-120 meters for the ultra-low altitude layer, 120-300 meters for the low altitude layer). Using the OpenSceneGraph engine, it projects the aforementioned static elements and the dynamic aircraft positions from the structured airspace situation dataset onto a unified ENU (East-North-Sky) local coordinate system, constructing a 1-meter × 1-meter × 1-meter resolution 3D grid space. In the lattice space, each aircraft generates a dynamic bounding volume with its current position as the center and its velocity vector as the direction, with an update frequency of 10Hz. The system traverses all N aircraft and constructs an N×N dimensional spacing matrix, where the matrix element (i,j) represents the three-dimensional Euclidean distance between aircraft i and j. At the same time, the system extracts the type code (mapped to a one-hot vector), physical dimensions (length, width, and height, in meters), and maximum maneuver overload capacity (in g) from the aircraft registration information, and encodes these attributes into node feature vectors of a graph neural network with a dimension of 16. The edge weights between two aircraft are determined by the product of their relative velocity, relative azimuth angle, and dimensions, and are normalized and used as edge features. The risk propagation model based on graph attention network is implemented as follows: N aircraft in the low-altitude airspace are modeled as a graph structure G=(V,E), where the node set V represents each aircraft, and the edge set E represents any pair of aircraft whose spatial distance is less than a preset interaction threshold (500 meters in this embodiment); the feature vector hi of each node i consists of the above 16-dimensional attributes; the weight wij of the edge (i,j) is calculated by the following formula: wij=Znorm×|vrel|×(Si×Sj) / dij 2 Where wij is the edge weight between aircraft i and j, Znorm is the normalization factor (satisfying Σwij=1, ensuring the sum of all edge weights is 1), vrel is the relative velocity of aircraft i and j (in m / s), Si and Sj are the physical dimensions of aircraft i and j respectively (taking the geometric mean of length, width and height, in meters), and dij is the three-dimensional Euclidean distance between aircraft i and j (in meters); the specific structure of the graph attention network model composed of three stacked multi-head graph attention layers is as follows: the first layer receives node features hi as input, this layer contains 4 attention heads, each attention head performs a linear transformation on the input features through a learnable weight matrix W1, and then calculates the attention coefficient alphaij of node i to its neighbor node j, the calculation formula is: alphaij=softmaxj(LeakyReLU(a T×[W1×hi||W1×hj]), where alphaij is the attention coefficient of node i to its neighbor node j, a is the attention weight vector (learnable parameter, dimension 2 × hidden layer dimension), W1 is the learnable weight matrix of the first layer (dimensional is hidden layer dimension × input feature dimension), hi and hj are the feature vectors of nodes i and j respectively (dimensional 16), || represents the feature vector concatenation operation (concatenating the two vectors end to end), T represents vector transpose, and softmaxj represents softmax normalization on all neighbor nodes j of node i (making Σ j(alphaij)=1), LeakyReLU is the activation function (slope of negative half axis is 0.2); the first layer output features are obtained by concatenating the outputs of 4 attention heads to obtain 64-dimensional hidden features; the second and third layers adopt the same structure, both containing 4 attention heads and 64-dimensional hidden units, propagating and updating node representations layer by layer; after three layers of propagation, the representation vector of each node integrates the state information of all aircraft in its three-hop neighborhood, thereby realizing the propagation and aggregation of risk on the graph structure; finally, for any pair of aircraft (i,j), its node representation hi output by the third layer is... 3 With hj 3 After concatenation, the data is input into a binary classifier (consisting of two fully connected layers, with a hidden layer of dimension 32 and an output layer activated by Sigmoid), which outputs the collision risk score for the aircraft pair, ranging from [0,1]. During the training phase, the model uses 5000 historical close-range conflict events (defined as horizontal distance < 50 meters and height difference < 10 meters lasting ≥ 2 seconds) and their corresponding safety event labels for supervised learning. The loss function uses FocalLoss to address sample imbalance. FocalLoss is defined as: FL(pt) = -αt × (1-pt)γ × log(pt), where FL(pt) is the FocalLoss loss value, pt is the probability that the model predicts the sample as positive (range [0,1]), and αt is the class weight coefficient (positive class αt = 0.25, negative class αt = 0.75, used to balance conflict and safety). The sample ratio is 1:19, γ is the focusing parameter (γ=2, used to reduce the loss contribution of easily distinguishable samples and enhance the attention to difficult samples), and log is the natural logarithm; training uses the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 200 epochs; during inference, the GAT risk assessment model outputs a collision risk score for each pair of aircraft, ranging from [0,1]; aircraft pairs with scores greater than 0.7 are marked as high-risk interaction pairs, forming a set of high-risk interaction pairs. Each record contains an aircraft ID pair, a risk score, an abnormal identifier (if any) in the associated list of abnormal aircraft identifiers, and the most recent interaction timestamp.

[0026] The environmental constraint replanning unit receives the set of high-risk interaction pairs and simultaneously accesses external environmental data sources: meteorological radar echo intensity data (from X-band phased array radar, resolution 500 meters, update rate 1 minute), electromagnetic spectrum interference monitoring data (from urban radio monitoring stations, frequency band coverage 100MHz-6GHz, power spectral density resolution 1dBm / MHz), and CityGML obstacle mesh (including buildings, high-voltage lines, tower cranes, etc.). The multidimensional environmental element model is implemented as follows: The model divides the environmental factors affecting aircraft safety into three categories: meteorological disturbance factors, electromagnetic interference factors, and obstacle distribution factors. Meteorological disturbance factors include parameters such as wind speed, wind direction, turbulence intensity, precipitation intensity, and visibility, which are obtained by converting the meteorological radar echo intensity data using empirical formulas. For example, wind speed vwind is estimated by multiplying the radar echo intensity Zradar to the power of one-third and an empirical coefficient. Electromagnetic interference factors include power spectral density, signal-to-noise ratio, and number of available channels in each frequency band, which are directly provided by the electromagnetic spectrum interference monitoring data. Obstacle distribution factors include building height, high-voltage line location, and tower crane operating range, which are provided with three-dimensional geometric information by the CityGML obstacle mesh. All three types of factors are labeled with spatial location and timestamp in the WGS-84 geodetic coordinate system. The three types of external environmental data are preprocessed and then input into the voxelization engine, which divides the urban low-altitude area (116.3°-116.5°E, 39.8°-40.0°N, altitude 0-500 meters) into 5×5×5 meter voxel units, totaling approximately 8 million voxels.The multidimensional environmental voxel model is specifically implemented as follows: This model discretizes the continuous spatial data in the aforementioned multidimensional environmental element model into a regular three-dimensional voxel grid structure; each voxel unit (x, y, z) stores the values ​​of three channels: the weather disturbance layer (cweather), the electromagnetic interference layer (cemf), and the obstacle distribution layer (cobstacle); the weather echo intensity is mapped to the center point of each voxel through bilinear interpolation, and then normalized by Min-Max to obtain the weather disturbance layer value cweather=(Zradar-Zmin) / (Zmax-Zmin), ranging from [0,1], where Zradar is the current voxel's radar echo intensity dBZ value. Zmin is the minimum echo intensity threshold (set to 10 dBZ), Zmax is the maximum echo intensity threshold (set to 70 dBZ), and cweather is the normalized meteorological disturbance layer value (0 indicates no disturbance, 1 indicates extreme weather). The power spectral density Pemf of the electromagnetic spectrum interference monitoring data is calculated to the received power Prx at the voxel center point using the free space path loss model FSPL. The signal attenuation factor cemf = (Pemf - Prx) / Pemf is stored in the electromagnetic interference layer, where Pemf is the transmitted power spectral density of the monitoring station (unit: dBm / MHz), Prx is the received power at the voxel center point (unit: dBm / MHz), and cemf is the signal attenuation factor. Attenuation factor (range [0,1]); The CityGML obstacle mesh uses a rasterization algorithm to determine whether the voxel center point is located inside the obstacle geometry. If so, it is marked as 1 in the obstacle distribution layer; otherwise, it is marked as 0. The system uses a sliding time window (length 10 seconds, step size 1 second) combined with an LSTM time series model to predict the evolution of each voxel layer in the next 10 seconds. The input of the LSTM time series model is the three-channel voxel matrix of the past 10 frames (10 seconds of history), with dimensions [10, Hvoxel, Wvoxel, Dvoxel, 3], where Hvoxel, Wvoxel, and Dvoxel are the east-west, north-south, and vertical dimensions of the voxel mesh, respectively. The LSTM unit has a hidden layer dimension of 128, and the network structure contains two stacked ConvLSTM units. Each ConvLSTM unit has a convolution kernel size of 3×3×3, a stride of 1, and padding of 1, maintaining the spatial resolution. The output is the voxel state prediction for the next 10 frames, with dimensions of [10, Hvoxel, Wvoxel, Dvoxel, 3], thereby generating the multidimensional environmental voxel model. During the training phase, the model uses historical meteorological radar sequence data, electromagnetic monitoring sequence data, and obstacle change records (such as tower crane movement) to construct a training set, uses mean squared error (MSE) as the loss function, and trains with the Adam optimizer for 100 epochs.Safety constraint parameters are generated by the multidimensional environment voxel model: the wind speed gradient threshold is taken from the maximum difference between adjacent voxels in the meteorological disturbance layer; the signal attenuation region is defined as a continuous voxel group > 0.6 in the electromagnetic interference layer; the obstacle envelope is extracted from the obstacle distribution layer using the MarchingCubes algorithm; for each aircraft in the high-risk interaction pair set, its 90 historical trajectory points (sampling rate 9Hz) within the past 10 seconds are input into a trajectory prediction network based on the Transformer architecture. The encoder contains 6 layers, each with an 8-head self-attention mechanism, and the decoder outputs 900 future prediction points (10Hz × 90 seconds) to form conflict trajectory predictions; for each prediction point in the conflict trajectory prediction, the voxel where it is located is queried in the multidimensional environment voxel model. If any of the following conditions exist in the three-layer parameters of the model: wind speed gradient > 3m / s / m, signal attenuation > 15dB, or the object is located within the obstacle envelope, it is determined to be a violation of environmental constraints. At this time, a trajectory replanner based on the Deep Deterministic Policy Gradient (DDPG) algorithm is invoked. The state space of the trajectory replanner includes the current position, velocity, the positions of neighboring aircraft in the set of high-risk interaction pairs, and the constraint map of the multi-dimensional environmental voxel model. The action space consists of the heading angle change rate and vertical velocity command. The reward function comprehensively considers obstacle avoidance margin, trajectory smoothness, deviation from the original plan, and the scope of influence (based on the number of high-risk pairs). After 200 online policy iterations, a new waypoint sequence is output, including three-dimensional coordinates, expected arrival time, and velocity profile, constituting a compliant trajectory adjustment command.

[0027] The communication coordination optimization unit receives the compliant trajectory adjustment instruction and parses the waypoint sequence and time window. Simultaneously, it loads the low-altitude private network base station location database (including latitude and longitude, antenna height, and adjustable azimuth range) and satellite communication resource configuration table (including visible satellite ephemeris, beam coverage, and channel capacity) from the 3D city model. For each waypoint in the waypoint sequence, it first calculates its line-of-sight (LoS) path to the three nearest ground base stations, using a ray tracing algorithm to account for building obstruction, retaining only base stations with LoS. If a LoS ground base station exists, it selects the one with the minimum signal propagation loss as the optimal choice. The system establishes the primary communication link and calculates the optimal azimuth angle required for reception by the ground station at the waypoint. If all ground base stations are blocked or the waypoint is located in an area with insufficient ground base station coverage on the edge of the city (signal strength below -95dBm), the system switches to the satellite communication backup link. The system queries the satellite communication resource configuration table, calculates the set of visible satellites (elevation angle greater than 15 degrees) based on the WGS-84 coordinates of the waypoint location and the current UTC time, uses the two row roots (TLE) orbit model to predict the azimuth, elevation, and slant range of each satellite in the next 10 seconds, and selects the satellite with the highest elevation angle and the smallest Doppler shift as the primary satellite communication link. This satellite link operates in the Ku band. The segment (14GHz uplink, 12GHz downlink) uses Time Division Multiple Access (TDMA) with a time slot allocation period of 100ms. At the sampling time specified by the compliant trajectory adjustment command (determined by the time window), if a terrestrial link is selected, the directional antenna array of the base station corresponding to that main communication link is activated, and its main lobe is aligned with the optimal azimuth angle via a servo motor, with a beamwidth of 10 degrees. If a satellite link is selected, the phased array antenna of the aircraft's onboard satellite communication terminal is activated, with the beam pointing electronically to track the main satellite in real time, a beamwidth of 3 degrees, and circular polarization to counteract the Doppler effect. The directional antenna... The linear array or the satellite communication terminal receives the raw echo signal stream from the spacecraft. This raw echo signal stream contains microsecond-level time and frequency markers (generated by the spacecraft-end PTP from clock synchronization; for satellite links, time synchronization is calibrated through the BeiDou Satellite Timing System (BDS) or GPS Timing System, with errors controlled within ±100 nanoseconds). The raw echo signal stream is input to a four-layer 1D convolutional autoencoder deployed on the base station edge server or satellite ground station. The encoder of this convolutional autoencoder has a kernel size of 7 per layer, with channel numbers of 32, 64, 128, and 256 respectively, an activation function of LeakyReLU, and a decoder of symmetric deconvolution.By calculating the mean square error between the input and reconstructed signals, multipath effect characteristics (delay spread >50ns for ground links, amplitude fluctuations >3dB due to ionospheric scintillation for satellite links), co-channel interference characteristics (power surges in specific frequency bands), and atmospheric noise characteristics (broadband white noise enhancement; rain attenuation effects need to be considered additionally for satellite links) are extracted from the residuals for communication channel quality assessment. For satellite links, the communication channel quality assessment also includes satellite link budget calculation: Equivalent Isotropic Radiated Power (EIRP) is 53dBW, and Free Space Path Loss (FSPL) is... The carrier-to-noise ratio (C / Nsat) is calculated based on the slant range Rslant (FSPLsat = 32.45 + 20 × log10(fGHz) + 20 × log10(Rslantkm), where fGHz is the frequency of 12GHz and Rslantkm is the slant range in kilometers). The atmospheric absorption loss Latm is set to 0.5dB, and the rain attenuation loss Lrain is dynamically calculated based on the real-time rainfall rate (using the ITU-RP.618 model). The receiver quality factor G / T is 20dB / K, from which the carrier-to-noise ratio C / Nsat is calculated. This is combined with the signal strength of the original echo signal stream. The system outputs a communication channel quality index vector. For terrestrial links, this vector includes three components: carrier-to-interference-plus-noise ratio (CINR), root mean square delay spread, and maximum Doppler shift. For satellite links, it includes three components: carrier-to-noise ratio (C / N), link margin, and bit error rate (BER). This communication channel quality index vector is matched against a pre-defined communication parameter mapping table (trained from historical channel measurement data, divided into terrestrial link parameter table and satellite link parameter table). For terrestrial links, it outputs the corresponding transmit power (dBm) and modulation scheme (e.g., QPSK, 16Q). Based on the satellite link output transmit power (W), modulation and coding scheme (e.g., 8PSK, 16APSK, 32APSK), forward error correction code rate (e.g., LDPC code rate 1 / 2, 2 / 3, 3 / 4), and time slot occupancy (1-4 time slots), a communication parameter optimization instruction is generated. This instruction includes the link type identifier (terrestrial link or satellite link), the selected base station ID or satellite NORAD number, the optimized communication parameters, and the expected link hold-up duration.

[0028] The collaborative execution decision unit receives the compliant trajectory adjustment instruction and the communication parameter optimization instruction, and jointly encodes them: the trajectory part of the compliant trajectory adjustment instruction includes 10 key waypoints (simplified by Douglas-Peucker), a velocity profile (5-segment linear interpolation), and obstacle avoidance margin (minimum voxel distance); the communication part of the communication parameter optimization instruction includes transmit power, modulation order, and HARQ retransmission count, which are concatenated into a 128-dimensional operational state feature vector. This operational state feature vector is input to an ensemble classifier consisting of three models: XGBoost, RandomForest, and LightGBM, which respectively use 5000 decision trees, 200 random trees, and 1000 gradient boosting trees. The training data consists of 100,000 historical flight records (including normal and abnormal labels). The ensemble classifier uses a weighted voting mechanism of 0.4:0.3:0.3 to output a binary classification result (normal / abnormal). Based on the binary classification result, combined with the high-risk interaction pair set and the abnormal aircraft identifier list, a normal aircraft status list is generated. There are two types of status lists: abnormal flightcraft status lists and abnormal flightcraft status lists. For abnormal flightcraft status list items, a "flightcraft-event-instruction" triple relationship is established through the Neo4j graph database: (:FlightCraft{id})-[:INVOLVES]->(:Event{type,time})-[:TRIGGERS]->(:Instruction{content}), and matched with the emergency response modes in the predefined emergency response template library—Level I response mode (such as suspected terrorist attack) triggers public security air interception, radio suppression, and ground control; Level II response mode (such as risk of uncontrolled fall) links fire, medical, and traffic control; Level III response mode (such as slight deviation) only notifies the operator to correct; Each response mode includes specific multi-department linkage actions, resource scheduling lists (such as the number of drone countermeasure vehicles and police deployment points) and strict time window constraints (such as response within 5 minutes), generating an emergency response plan; For normal flightcraft status list items, they are input into a multi-agent game scheduling optimizer. The multi-agent game model is implemented as follows: The model models M normal aircraft in the low-altitude airspace as M rational agents. Each agent i has an independent policy space Si and utility function Ui. The policy space Si consists of three dimensions: takeoff time slot ti is selected from the predefined time slot set {T0, T0+10 seconds, T0+20 seconds, ..., T0+3600 seconds}, available routes Rk are selected from the predefined route library {R1, R2, ..., Rk} (in this embodiment, k equals 15 predefined routes), and altitude layer hi is selected from {0 meters, 10 meters, 20 meters, ..., 300 meters}. The policy space of agent i is represented as si equal to (ti, Rk, hi).

[0029] The specific implementation of the game model including the utility function and constraints is as follows: The utility function Ui consists of three terms: the flight efficiency reward term Refficiency(si) = -Tflighti / Tstandardi (where Tflighti is the actual flight time of agent i and Tstandardi is the standard flight time), the energy consumption penalty term Penergy(si) = -Etotali / Eratedi (where Etotali is the total energy consumption of agent i and Eratedi is the rated energy consumption), and the conflict penalty term Pconflict(si, s{-i}) = -100×Nconflicti (where Nconflicti is the number of spatio-temporal conflicts between agent i and other agents), where s{-i} is the set of strategies of all other agents except agent i; the complete utility function Ui(si, s{-i}) = ω1×Refficiency(si) + ω2×Penergy(si) + ω3×Pconflict(si, s{-i}), where the weight coefficients ω1 = 0.5, ω2 = 0.2, and ω3 = 0.3; the spatio-temporal conflict determination rule is: if ||pi(t) - pj(t)|| < dsafe (where dsafe = 30 meters), then one conflict is counted, where pi(t) and pj(t) are the three-dimensional position vectors of agent i and j at time t (including the three components of longitude, latitude, and altitude), ||·|| represents the Euclidean distance (L2 norm), and dsafe is the safety interval threshold (set to 30 meters in this embodiment); the constraints include: the takeoff time slots must not conflict (at most 2 aircraft are allowed to take off in the same time slot), the combination of altitude layers and routes needs to satisfy the airspace stratification management rules (logistics UAVs are restricted to heights between 0 and 120 meters, and manned eVTOLs are restricted to heights between 120 and 300 meters), and the total flight time must not exceed the endurance of the aircraft; the multi-agent game scheduling optimizer uses the iterative best response algorithm to solve for the Nash equilibrium: initialize all agent strategies as random strategies, in each round of iteration, fix the strategies of other agents sminusi in turn, let agent i find the best response strategy sistar that maximizes Ui by exhaustively searching its strategy space Si, and update si to sistar; repeat the above iteration until all agent strategies do not change in 5 consecutive rounds of iteration or reach the maximum number of iterations 200 times, at which point it is considered that the Nash equilibrium is reached; output the globally optimal flight scheduling plan, including the takeoff time slot, route number, and altitude layer allocation of each aircraft.Ultimately, the emergency response plan and the flight scheduling plan are loaded into the Apache Airflow task scheduling engine, decomposed into atomic operations (such as sending trajectory instructions to ID123, scheduling countermeasure vehicles to coordinates XYZ), and distributed through RabbitMQ message queues to standard RESTful API interfaces of departments such as public security, civil aviation supervision, emergency management, and telecommunications operators. The interfaces follow the OpenAPI 3.0 specification, and authentication uses OAuth 2.0. For scenarios in remote areas or where ground communication networks are interrupted in emergency situations, the atomic operations send instructions to the aircraft through satellite communication links (using the Tiantong-1 mobile satellite communication system or Beidou short message communication service). The instructions are encrypted with AES-256 and encapsulated into Beidou short messages (maximum 1000 Chinese characters per message) or Tiantong voice + data hybrid frames (data rate 9.6kbps), ensuring that the instructions are delivered to the aircraft and receive a confirmation response within 3 seconds, generating a low-altitude airspace collaborative control execution strategy, and realizing closed-loop control from perception to execution.

[0030] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principles of this invention are further supplemented below with a specific application scenario.

[0031] When deploying the multimodal sensing and intelligent decision-making infrastructure in a certain area of ​​Beijing, millimeter-wave radar (operating frequency band 77 GHz, detection range 0 to 2 km), 1080P resolution optical cameras (field of view 60 degrees), infrared thermal imagers (noise equivalent temperature difference less than 50 milliklvin), ADS-B receivers (supporting 1090 extended monitoring mode and general access transceiver dual mode), and RF spectrum detectors (covering 9 kHz) were installed at 12 key nodes, including building rooftops (coordinates: 116.3215 degrees east longitude, 39.9528 degrees north latitude), the Boya Tower communication tower of Peking University, and traffic monitoring poles on main streets. The five types of sensors (up to 6 GHz) are rigidly fixed to the same gimbal base by an L-shaped bracket. The optical axes of each sensor are calibrated using a laser collimator to ensure that the center line of their field of view is projected onto the horizontal plane with an overlap of 87%. The power and signal lines of the five types of sensors are connected to the edge computing gateway. The edge computing gateway has a built-in u-blox ZED-F9P global navigation satellite system module as a UTC time source. One pulse per second and a 10 MHz reference clock are sent to the external trigger ports of each sensor through coaxial cables. The measured maximum deviation of the output frame timestamp of each sensor is ±1.8 microseconds, which meets the ±2 microsecond synchronization requirement.

[0032] The raw sensing data collected by the five types of sensors are aggregated via gigabit Ethernet to an edge server deployed in the data center. The multi-source sensing fusion unit first uses the UTC timestamp as a reference and employs cubic spline interpolation to uniformly resample the asynchronously arriving radar point cloud (update rate 20Hz), video stream (30fps), infrared matrix (25fps), ADS-B message (1Hz), and spectrum sequence (100Hz) to a 10ms time grid. Subsequently, a six-dimensional state vector containing longitude, latitude, altitude, eastward velocity, northward velocity, and vertical velocity is constructed. The process noise covariance Q is set as a diagonal matrix Q=diag(qlon,qlat,qalt,qve,qvn,qvu)=diag(1×10 -8 1×10 -8 1×10 -6 ,0.1,0.1,0.2), where qlon, qlat, qalt, qve, qvn, and qvu correspond to the process noise variances of longitude, latitude, altitude, eastward velocity, northward velocity, and vertical velocity, respectively, to reflect the stronger disturbance of vertical motion by urban low-altitude turbulence; the observation noise R is dynamically adjusted according to the visibility of the day. When the PM2.5 concentration is greater than 150 micrograms per cubic meter, the observation noise of the optical camera increases to 10 times, while the observation noise of the infrared thermal imaging decreases to 0.5 times; after Kalman filtering output, DBSCAN clustering (neighborhood radius of 3 meters, minimum number of points of 5) is performed on the radar point cloud to form target clusters, and GMM background modeling (Gaussian component number of 5, learning rate of 0.01) is performed on the video frame to extract the foreground motion region, and O is used for the infrared image. TSU adaptive threshold segmentation is used to obtain the heat source contour. A 256-point short-time Fourier transform (using Hanning window and 50% overlap) is performed on the RF data to extract the center frequency, 3 dB bandwidth and modulation entropy as spectral fingerprints. The above four features, together with the aircraft type code parsed by ADS-B (such as the hexadecimal code 4B1A2C corresponding to DJI M300), constitute the evidence body, which is input into the weighted DS fusion layer. The weight is calculated as wi=Acci / Σ(Accj), where wi is the fusion weight of sensor i, Acci is the recognition accuracy of sensor i under the current weather conditions in the past 10 minutes, and Σ(Accj) is the sum of the recognition accuracy of all sensors (j traverses all sensors). Targets with a total confidence of 0.92 or higher after fusion are written into the structured airspace situation dataset.

[0033] After receiving the structured airspace situation dataset, the abnormal behavior discrimination unit loads the no-fly zone GeoJSON file (containing 23 polygons including Zhongnanhai and the runway extension of Capital Airport) provided by the Beijing Air Traffic Management Bureau, constructs an R-tree index using the STCreateFishnet function of PostGIS, and achieves a query response time of <5ms; simultaneously, it synchronizes registered flight plans from the Civil Aviation Administration's UOM platform every 5 seconds, constructs an STBBox index for each planned trajectory point sequence, with each bounding box covering [lon±0.001°, lat±0.001°, alt±20m, t±5s]; when the structured airspace... When the aircraft with ID UAV-20240515-0832 in the situational awareness dataset reports its location (116.3250°, 39.9550°, 85m), the R-tree range query returns that it falls within the polygon of the temporary no-fly zone for major events in the software park. The system immediately marks it as a no-fly zone intrusion and packages the aircraft ID, event type, occurrence UTC time, and the original sensing data snapshot (including the original radar point cloud, video frame screenshot, infrared heat map, and spectrum map) into the abnormal aircraft identifier list. This data is then pushed to the Elasticsearch cluster via Kafka, and the index lifecycle policy is set to retain it for 180 days.

[0034] After receiving the structured airspace situation dataset and the list of anomalous aircraft identifiers, the dynamic risk assessment unit obtains the Urban Geographic Markup Language 2.0 version data from the Beijing Urban Information Modeling Platform, extracts the building detail level 2 model (including glass curtain wall reflectivity parameters), the 1-meter resolution digital elevation model, and airspace layering rules (0 to 120 meters for logistics drones, 120 to 300 meters for manned electric vertical take-off and landing aircraft). Using an open scene graphics engine, the static elements and the dynamic aircraft positions in the structured airspace situation dataset are uniformly converted to a Northeast Celestial Coordinate System with 116.32°E, 39.95°N as the origin, constructing a 1-meter resolution three-dimensional grid space. The system detects drone number 20240515-0832 (multi-rotor type, size 1.2m x 1.2m x 0.0m) in the list of anomalous aircraft identifiers. The two aircraft, identified as EV20240515-0915 (tilt rotor type, dimensions 6m x 2m x 2m, maximum overload 1.8 times gravitational acceleration) and EV20240515-0915 (tilt rotor type, dimensions 6m x 2m x 2m, maximum overload 1.8 times gravitational acceleration), are only 42m apart and 8m apart in height, with a duration of 2.3 seconds. The attributes of the two aircraft are encoded as 16-dimensional node features (including one-hot encoded type code, length, width, height, overload capacity, etc.). The relative speed of 15m / s, azimuth angle of 37 degrees, and the product of dimensions of 7.2 square meters are normalized and used as edge features. These are input into the three-layer graph attention network risk assessment model (4 attention heads per layer, 64-dimensional hidden units). After training on 5000 historical conflict samples, the graph attention network model outputs a collision risk score of 0.78, which exceeds the 0.7 threshold. Therefore, the aircraft pair is added to the high-risk interaction pair set.

[0035] After receiving the high-risk interaction pair set, the environmental constraint replanning unit synchronously accesses Beijing X-band phased array meteorological radar data (echo intensity Z=45dBZ), radio monitoring station electromagnetic data (2.4GHz band power spectral density up to -65dBm / MHz), and CityGML obstacle mesh (including the No. 1 300-meter high-rise building) as external environmental data sources; the voxelization engine divides the study area into 5×5×5 meter voxels, totaling 8.12 million voxels; the meteorological echo intensity is mapped to a meteorological disturbance layer value of 0.45, and the electromagnetic spectrum interference monitoring data is converted into an electromagnetic interference layer value of 0.68 through the Okumura-Hata path loss model. The yGML obstacle mesh marks the voxels occupied by tall buildings in the obstacle distribution layer; the LSTM temporal model (128-dimensional hidden layer) extrapolates the evolution of each voxel layer in the next 10 seconds, generating the multi-dimensional environmental voxel model, predicting that the 2.4GHz interference zone will expand eastward; for UAV-20240515-0832 in the high-risk interaction pair set, its 90 historical trajectory points in the past 10 seconds are input into the Transformer trajectory prediction network (6-layer encoder, 8-head self-attention), and the decoder outputs 900 prediction points for the next 90 seconds to form a conflict trajectory prediction; the 125th prediction point in the conflict trajectory prediction (116.3280°, 39.9560°, The voxel containing the 90m value has an electromagnetic interference layer value of 0.71 > 0.6 in the multidimensional environment voxel model, indicating a violation of environmental constraints. The DDPG trajectory replanner uses the current position, velocity, neighboring eVTOL positions in the high-risk interaction pair set, and the multidimensional environment voxel model constraint map as its state. The action output heading angle change rate ∈ [-15°, 15°] / s and vertical velocity ∈ [-3, 3] m / s. The reward function is Rreward = 0.4 × Cclearance + 0.3 × Ssmooth - 0.2 × Ddeviation - 0.1 × Nrisk, where Rreward is the total reward value, Cclearance is the total reward value, and Ddeviation is the minimum reward value. ce represents the obstacle avoidance margin (in meters, indicating the distance from the trajectory point to the nearest obstacle), Ssmooth represents the trajectory smoothness (calculated by the rate of change of heading angle between adjacent waypoints, in degrees / second), Ddeviation represents the deviation from the original plan (the three-dimensional Euclidean distance between the new trajectory and the original planned trajectory, in meters), and Nrisk represents the number of high-risk pairs (the number of other aircraft that pose a collision risk to this aircraft). After 200 online iterations, the new waypoint sequence is output as: [(116.3260°, 39.9555°, 88m, t+5s), (116.3270°, 39.9565°, 92m, t+10s),…], which constitutes the compliant trajectory adjustment instruction.

[0036] After receiving the compliant trajectory adjustment instruction, the communication coordination optimization unit parses the new waypoint sequence and loads the low-altitude private network base station database (containing 32 base station locations) and satellite communication resource configuration table (containing three types of satellite resources: Tiantong-1, Beidou short message, and Starlink). For the waypoint (116.3270°, 39.9565°, 92m) in the new waypoint sequence, the ray tracing algorithm calculates its line-of-sight path (LoS) with the three nearest base stations (ID: BTS-08, BTS-12, BTS-15). Only BTS-12 has no building obstruction, and its predicted signal strength is -78dBm. Since the signal strength meets the requirements... The system selects ground base station BTS-12 as the primary communication link; calculates the optimal receiving azimuth angle at BTS-12 base station as 127°; activates its directional antenna array at the sampling time t+10s specified by the compliant trajectory adjustment command; the servo motor drives the main lobe to align with the optimal azimuth angle of 127°, with a beamwidth of 10°; for waypoints in the urban edge area (116.4980°, 40.0490°, 120m), the signal strength of all ground base stations is below -95dBm, and the system automatically switches to the satellite communication backup link. The system queries the satellite communication resource configuration table to obtain the currently visible Tiantong-1 satellite (NORAD number 43683). With an elevation angle of 38°, an azimuth angle of 215°, and a slant range of 36,850 km, the onboard Ku-band satellite communication terminal of the spacecraft is activated. The phased array antenna is electronically scanned and pointed at the satellite, establishing a bidirectional link of 14.25 GHz uplink and 12.5 GHz downlink. The original echo signal stream received by the directional antenna array or the satellite communication terminal contains microsecond-level time and frequency markers generated by PTP from a clock or BeiDou satellite timing. For the ground link, the four-layer 1D convolutional autoencoder (kernel size 7, channels 32→64→128→256) reconstructs the original echo signal stream. Residual analysis shows that the power of the 2.4 GHz band increases by 12 dB, and the delay spread reaches 65 dB. The communication channel quality assessment is completed using ns; combined with the original echo signal stream's CINR=18dB, the communication channel quality index vector [18dB, 65ns, 8Hz] is output; for the satellite link, the link budget is calculated as follows: EIRP=53dBW, free space path loss FSPLsat=205.8dB (calculated based on a slant range of 36850km and a frequency of 12.5GHz), rain attenuation Lrain=2.1dB (current rainfall rate 8mm / h), received G / T=20dB / K, resulting in a carrier-to-noise ratio C / Nsat=15.2dB, link margin 3.2dB, and bit error rate BER=1.2×10 -5 Output satellite link communication channel quality index vector [15.2dB, 3.2dB, 1.2×10] -5The communication channel quality index vector lookup table matches the communication parameter mapping table. For the ground link, the optimal parameters are: transmit power 23dBm, modulation scheme 16QAM, HARQ retransmission times 2. For the satellite link, the optimal parameters are: transmit power 5W, modulation and coding scheme 8PSK, LDPC code rate 2 / 3, occupied time slots 2. The communication parameter optimization instruction containing the link type identifier (in this example, waypoints use ground links and edge waypoints use satellite links) is generated.

[0037] After receiving the compliant trajectory adjustment instruction and the communication parameter optimization instruction, the collaborative execution decision unit jointly encodes and concatenates the trajectory portion of the compliant trajectory adjustment instruction (10 Douglas-Peucker simplified waypoints, 5 velocity profiles, minimum voxel distance of 12 meters) and the communication portion of the communication parameter optimization instruction (23dBm, 16QAM, 2 times) into a 128-dimensional operational state feature vector; the ensemble classifier (XGBoost: 5000 trees, RandomForest: 200 trees, LightGBM: 1000 trees) outputs a weighted voting binary classification result as an anomaly; combining the high-risk interaction pair set and the anomaly flight... The list of aircraft identifiers is used to generate the normal aircraft status list and the abnormal aircraft status list; for UAV-20240515-0832 in the abnormal aircraft status list, a triplet relationship of aircraft-event-instruction is established through the Neo4j graph database: (:FlightCraft{id:UAV-20240515-0832})-[:INVOLVES]->(:Event{type:No-fly zone intrusion,time:2024-05-15T08:32:15Z})-[:TRIGGERS]->(:Instruction{content:Initiate Level I response}); The Level I response mode is matched from the emergency response template library: two UAV countermeasure vehicles are dispatched to (116.3255°, 39.9552°) and (116.3265°, 39.9558°), the public security air interception team takes off within 5 minutes, and the radio jamming equipment activates 2.4 / 5.8GHz blocking to generate the emergency response plan; the normal list items in the normal aircraft status list are input into the multi-agent game scheduling optimizer, with takeoff time slots (10-second granularity), predefined routes (north-south corridor), and altitude layers (10-meter granularity). The system uses a layer as the policy space, and iteratively solves the Nash equilibrium to output the globally optimal flight scheduling scheme, allocating UAV-20240515-0832 to a backup flight route. Finally, the emergency response scheme and the flight scheduling scheme are decomposed into atomic tasks by the Apache Airflow task scheduling engine, and pushed to the standard RESTful API interfaces of departments such as public security API and civil aviation supervision API through the RabbitMQ message queue, generating the low-altitude airspace collaborative management and control execution strategy, and completing the closed-loop management from perception to execution.

[0038] Through the above steps, this invention achieves accurate identification of illegal aircraft, quantitative assessment of high-risk interactions, real-time trajectory replanning under environmental constraints, dynamic collaborative optimization of communication parameters, and automated scheduling of cross-departmental emergency resources in real and complex urban scenarios. It effectively solves the technical problem that traditional systems cannot balance identification accuracy, response timeliness, and collaborative efficiency.

[0039] A computer-readable storage medium is provided for storing computer-readable instructions. When these instructions are read and executed by a computer, they enable the aforementioned AI-based collaborative management and control method for low-altitude airspace information hubs. The computer-readable storage medium includes, but is not limited to, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, solid-state drive (SSD), optical disk, USB flash drive, or any other form of non-volatile or volatile storage medium. The computer-readable instructions are organized into executable program code, which, when executed by a processor deployed on an edge server, cloud computing platform, or distributed computing cluster, can drive the entire low-altitude airspace collaborative management and control system to complete the closed-loop management and control process from multimodal sensing data acquisition to cross-departmental emergency resource dispatch.

[0040] Specifically, the computer-readable instructions stored in the computer-readable storage medium include program code for the following functional modules: First, a multi-source sensing fusion module instruction, which, when executed, can receive raw sensing data collected from multimodal sensors such as millimeter-wave radar, optical cameras, infrared thermal imagers, ADS-B receivers, and RF spectrum detectors; perform spatiotemporal calibration processing using the UTC timestamps carried in the raw sensing data; uniformly resample asynchronous data with different update rates to a 10-millisecond time grid using cubic spline interpolation; call the Kalman filter algorithm to construct a six-dimensional state vector containing longitude, latitude, altitude, eastward velocity, northward velocity, and vertical velocity for each target to perform state estimation; and execute D... The BSCAN density clustering algorithm extracts target clusters from radar point clouds, and the Gaussian mixture model (GMM) is used to model the background of video frames to separate moving targets. The Otsu adaptive threshold segmentation algorithm is applied to extract infrared heat source contours, and the 256-point short-time Fourier transform (STFT) is performed to extract RF spectral fingerprints. The weighted Dempster-Shafer evidence theory fusion algorithm is called to dynamically allocate weight coefficients based on the historical accuracy of each sensor under the current meteorological conditions to complete the fusion of multi-source features and output a structured airspace situation dataset. Each record includes target ID, WGS-84 latitude and longitude, ellipsoidal height, three-dimensional velocity vector, aircraft type, confidence weight of each sensor, and total confidence after fusion.

[0041] Second, the abnormal behavior discrimination module, when executed, receives the structured airspace situation dataset, loads the no-fly zone rule base stored in GeoJSON format and the registered flight plan database synchronized by the Civil Aviation Administration's UOM platform; calls the PostGIS spatial database engine to construct an R-tree spatial index, establishes a hierarchical bounding box structure for the no-fly zone polygons, and constructs a spatiotemporal bounding box (STBBox) index for the trajectory point sequence of the registered flight plan; for each aircraft in the structured airspace situation dataset, performs an R-tree range query to determine whether its current position falls within the no-fly zone polygon, calculates the Euclidean distance from the current position to the nearest point of the registered trajectory, and determines that the trajectory deviation exceeds the limit when the horizontal offset exceeds 50 meters or the altitude deviation exceeds 10 meters and the duration exceeds 3 seconds; writes the aircraft ID, event type, occurrence time (UTC), and original perception snapshot of the aircraft that meet the no-fly zone intrusion or trajectory deviation exceeding the limit conditions into the abnormal aircraft identifier list, and pushes it to the Elasticsearch log system for persistent storage through the Kafka message queue.

[0042] Third, the dynamic risk assessment module, when executed, receives the structured airspace situation dataset and the list of anomalous aircraft identifiers, obtains 3D geographic information data in CityGML2.0 format from the City Information Modeling (CIM) platform, calls the OpenSceneGraph engine to project static features and the dynamic aircraft positions in the structured airspace situation dataset onto a unified ENU local coordinate system, constructing a 3D grid space with a resolution of 1m×1m×1m; constructs an N×N dimensional aircraft spacing matrix, extracts type code, physical dimensions, and maximum maneuvering overload capacity from aircraft registration information and encodes them into 16-dimensional node feature vectors; calls a risk assessment model composed of a three-layer stacked multi-head graph attention network (GAT), which is trained based on 5000 historical close-range collision events, and outputs a collision risk score of 0 to 1 for each pair of aircraft; marks aircraft pairs with scores greater than 0.7 as high-risk interaction pairs, forming a set of high-risk interaction pairs, with each record containing an aircraft ID pair, risk score, associated anomalous identifier, and recent interaction timestamp.

[0043] Fourth, the environmental constraint trajectory replanning module instruction, when executed, can receive the set of high-risk interaction pairs and simultaneously access external environmental data sources such as meteorological radar echo intensity data, electromagnetic spectrum interference monitoring data, and CityGML obstacle mesh; it calls the voxelization engine to divide the urban low-altitude area into 5×5×5 meter voxel units, maps the meteorological echo intensity to the meteorological disturbance layer value, converts the electromagnetic power spectral density into a signal attenuation factor through the path loss model and stores it in the electromagnetic interference layer, and marks the voxels occupied by obstacles in the obstacle distribution layer; it calls the LSTM time series model to predict the evolution of each voxel layer in the next 10 seconds to generate a multi-dimensional environmental voxel model, from which wind speed gradient threshold, signal attenuation region, and obstacle envelope surface are extracted as safety constraint parameters; and it applies the voxel replanning module instruction to the set of high-risk interaction pairs. For each aircraft, a trajectory prediction network based on the Transformer architecture is invoked. The encoder receives historical trajectory points from the past 10 seconds, and the decoder outputs 900 predicted points for the next 90 seconds to form a conflict trajectory prediction. The predicted trajectory points are queried point by point in the three-layer parameters of the multi-dimensional environment voxel model to determine whether they violate environmental constraints. If the constraints are violated, a trajectory replanner based on the Deep Deterministic Policy Gradient (DDPG) algorithm is invoked. The state space of the replanner includes the current position, velocity, positions of neighboring aircraft, and voxel constraint map. The action space consists of the heading angle change rate and vertical velocity command. The reward function comprehensively considers obstacle avoidance margin, trajectory smoothness, deviation from the original plan, and the range of influence. After 200 online policy iterations, a new waypoint sequence is output, which constitutes a compliant trajectory adjustment command.

[0044] Fifth, the communication coordination optimization module, when executed, receives the compliant trajectory adjustment instruction, parses the waypoint sequence and time window, and loads the low-altitude private network base station location database and satellite communication resource configuration table. For each waypoint in the waypoint sequence, it calls the ray tracing algorithm to calculate the line-of-sight path (LoS) between the waypoint and the three nearest ground base stations. Considering building obstruction, if there are LoS ground base stations, it selects the one with the least signal propagation loss as the main communication link and calculates the optimal azimuth angle. If all ground base stations are obstructed, it switches to the satellite communication backup link and calculates the set of visible satellites and selects the optimal satellite. At the sampling time specified by the compliant trajectory adjustment instruction, it activates the corresponding ground base station directional antenna array or the phased array antenna of the aircraft's airborne satellite communication terminal, and aligns it with the target through a servo motor or electronic scanning. It receives the raw echo signal stream from the aircraft and calls the four-layer 1D convolutional autoencoder deployed on the base station edge server or satellite ground station. Each layer of the encoder has a convolution kernel size of 7 and a channel... The numbers are 32, 64, 128, and 256 respectively, the activation function is LeakyReLU, and the decoder is symmetric deconvolution. Multipath effect features, co-channel interference features, and atmospheric noise features are extracted from the residuals by calculating the mean square error of the input and reconstructed signals. For satellite links, additional link budget parameters (EIRP, free space path loss, atmospheric absorption loss, rain attenuation loss, and carrier-to-noise ratio) are calculated to evaluate the communication channel quality. A communication channel quality index vector is output based on the original signal strength. For terrestrial links, this includes three components: carrier-to-interference-plus-noise ratio (CINR), root mean square delay spread, and maximum Doppler shift. For satellite links, it includes three components: carrier-to-noise ratio (C / N), link margin, and bit error rate (BER). A lookup table is used to match a preset communication parameter mapping table (divided into terrestrial link parameter table and satellite link parameter table), outputting the corresponding transmit power, modulation and coding scheme, error correction code rate, and retransmission strategy. A communication parameter optimization instruction containing the link type identifier and the selected base station ID or satellite NORAD number is generated.

[0045] Sixth, the collaborative execution decision module commands, when executed, can receive the compliant trajectory adjustment command and the communication parameter optimization command, and jointly encode and concatenate the trajectory part (10 Douglas-Peucker simplified waypoints, 5 velocity profiles, obstacle avoidance margin) and the communication part (transmit power, modulation order, HARQ retransmission count) into a 128-dimensional operational state feature vector; it calls an ensemble classifier composed of three models: XGBoost, RandomForest, and LightGBM, which uses a weighted voting mechanism of 0.4:0.3:0.3 to output a binary classification result (normal / abnormal); based on the binary classification result, combined with the high-risk interaction pair set and the abnormal aircraft identifier list, it generates a normal aircraft status list and an abnormal aircraft status list; for the abnormal list items in the abnormal aircraft status list, it establishes a "aircraft-event-command" triple relationship through the Neo4j graph database, and matches the predetermined... The system utilizes emergency response templates in a unified emergency response framework, including emergency response modes (Level I, Level II, and Level III). Each response mode comprises specific multi-departmental collaborative actions, resource scheduling lists, and strict time window constraints to generate an emergency response plan. For the normal list items in the normal aircraft status list, a multi-agent game scheduling optimizer is invoked. This optimizer models each aircraft as an agent, with a policy space encompassing takeoff time slots, available routes, and altitude layers. The payoff function comprehensively considers flight efficiency, energy consumption, and the probability of conflict with other agents. By iteratively solving for Nash equilibrium, the system outputs the globally optimal flight scheduling plan. Finally, the Apache Airflow task scheduling engine is invoked to decompose the emergency response plan and the flight scheduling plan into atomic operations. These operations are then distributed via RabbitMQ message queues to standard RESTful API interfaces of departments such as public security, civil aviation supervision, emergency management, and telecommunications operators, generating a low-altitude airspace collaborative control execution strategy to achieve closed-loop control from perception to execution.

[0046] Furthermore, the computer-readable instructions also include program code for a data interface management module, a system configuration management module, a log recording module, an exception handling module, a concurrency control module, and a resource scheduling module, ensuring that the above six core functional modules can operate stably in a distributed computing environment. The data interface management module instructions are responsible for data interaction with sensor hardware, external data sources, and third-party APIs; the system configuration management module instructions are responsible for loading and managing configuration files such as the no-fly zone rule base, flight plan database, base station location database, communication parameter mapping table, and emergency response template library; the log recording module instructions are responsible for persistently storing intermediate results and final outputs such as the structured airspace situation dataset, the abnormal aircraft identifier list, the high-risk interaction pair set, the compliant trajectory adjustment instructions, the communication parameter optimization instructions, the emergency response plan, and the flight scheduling plan to database systems such as Elasticsearch and Neo4j; the exception handling module instructions are responsible for capturing and handling sensor failures, network interruptions, etc. The system is designed to handle various anomalies, such as algorithm computation errors and external API call failures, ensuring robustness. The concurrency control module coordinates the concurrent execution and data synchronization among the multi-source perception fusion module, the abnormal behavior discrimination module, the dynamic risk assessment module, the environmental constraint trajectory replanning module, the communication collaboration optimization module, and the collaborative execution decision module, ensuring that each module executes in the correct dependency order. The resource scheduling module manages the computing resources of edge servers, cloud computing platforms, or distributed computing clusters, including CPU, GPU, memory, and network bandwidth, achieving load balancing and dynamic resource allocation. This ensures the system can process massive amounts of multimodal perception data in real time and complete risk assessment, trajectory replanning, and emergency decision-making within millisecond-level response times.

[0047] Through the collaborative operation of the program code of the above functional modules, the computer-readable storage medium can drive the low-altitude airspace information central collaborative management and control system to complete the following technical processes: starting from the acquisition of raw sensing data by multimodal sensors, after spatiotemporal calibration (using BeiDou / GPS satellite timing to ensure time consistency of distributed nodes) and multi-source feature fusion to generate a structured airspace situation dataset, generating an abnormal aircraft identification list through rule comparison, using graph attention networks for dynamic risk assessment to form a high-risk interaction pair set, combining a multi-dimensional environmental voxel model for trajectory prediction and replanning to generate compliant trajectory adjustment instructions, and using directional antenna arrays and convolutional neural networks... The encoder performs communication channel quality assessment (adaptive selection of ground link and satellite link) to generate communication parameter optimization instructions. Finally, the trajectory and communication instructions are jointly encoded and classified to generate normal and abnormal aircraft status lists. Multi-agent game scheduling and multi-department collaborative abnormal early warning response are executed respectively (execution instructions are issued through ground network or satellite communication link) to obtain low-altitude airspace collaborative management and control execution strategy. This enables accurate identification of illegal aircraft, quantitative assessment of high-risk interactions, real-time trajectory replanning under environmental constraints, dynamic collaborative optimization of communication parameters (including ground and satellite hybrid network optimization), and automated scheduling of cross-departmental emergency resources.

[0048] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A low-altitude airspace information central collaborative control method based on AI, characterized in that, include: Initial sensing data in the low-altitude airspace is acquired through multimodal acquisition to obtain a sensing dataset; Spatiotemporal calibration and multi-source feature fusion are performed on the perception dataset to obtain a structured spatial situation dataset; Based on the airspace situation dataset, identity association and rule comparison are performed. If the airspace situation dataset of the current airspace location falls into a no-fly zone or is in a trajectory deviation situation, and the deviation value exceeds the preset deviation threshold, the corresponding aircraft ID is added to the identifier list to generate an abnormal aircraft identifier list. Based on the list of anomalous aircraft identifiers, the dynamic spatial distribution map of aircraft, and the risk propagation model based on graph attention networks, dynamic risk assessment is performed to obtain a set of high-risk interaction pairs. Based on the associated aircraft and their impact range in the high-risk interaction set, combined with a multi-dimensional environmental element model and the current position of the aircraft, the trajectory prediction network is used to predict conflict trajectories and perform environmental constraint trajectory replanning to generate compliant trajectory adjustment instructions. Based on the compliant trajectory adjustment instructions, the communication channel quality is evaluated through a directional antenna array and a convolutional autoencoder, and the communication parameters are adaptively adjusted to generate communication parameter optimization instructions. The compliant trajectory adjustment instructions and communication parameter optimization instructions are jointly encoded and classified to generate normal and abnormal aircraft status lists. Multi-agent game scheduling and multi-department collaborative abnormal early warning response are then executed to obtain a low-altitude airspace collaborative control and execution strategy.

2. The AI-based low-altitude airspace information central collaborative control method according to claim 1, characterized in that, Based on the airspace situation dataset, identity association and rule comparison are performed. If the offset value exceeds a preset offset threshold, an abnormal aircraft identifier list is generated, including: Load the no-fly zone rule base stored in GeoJSON format and build an R-tree spatial index structure; load the registered flight plan database containing fields such as takeoff time window, route, altitude layer and operation subject and build a flight plan spatiotemporal bounding box index. Extract the target ID, latitude and longitude, altitude and aircraft type from the airspace situation dataset, perform an R-tree range query on the latitude and longitude coordinates of each aircraft, and determine whether it falls within the geometric boundary of the no-fly zone. If the aircraft has not entered a no-fly zone, the reference trajectory point sequence in the current flight plan is extracted, and the Euclidean distance between the current actual position and the nearest reference trajectory point in the airspace situation dataset is obtained. If the Euclidean distance exceeds a preset offset threshold in the horizontal direction, it is determined that the trajectory deviation exceeds the limit; Write the aircraft ID, event type, occurrence time, and original sensing data snapshot from the airspace situation dataset that meet the conditions of no-fly zone intrusion or excessive trajectory deviation into the abnormal aircraft identifier list. The list of anomalous aircraft identifiers is pushed to the log system for persistent storage and a drive-away command generator is triggered simultaneously to verify whether the anomalous aircraft has a legitimate communication link and identity authentication status. If it does not, it is marked as a high-risk target to generate an anomalous aircraft identifier list containing aircraft ID, deviation amount, and risk level.

3. The AI-based low-altitude airspace information central collaborative control method according to claim 1, characterized in that, Based on the list of anomalous aircraft identifiers, the dynamic spatial distribution map of aircraft, and the risk propagation model based on graph attention networks, a dynamic risk assessment is performed to obtain a set of high-risk interaction pairs, including: Obtain 3D boundary grid data in CityGML format from airspace management geographic information, including building LOD2 models, terrain elevations, and airspace hierarchical structure; The position and velocity vectors of all aircraft in the airspace situation dataset are extracted and projected onto the three-dimensional boundary grid. Combined with the aircraft IDs and risk levels in the abnormal aircraft identifier list, a dynamic spatial distribution map of aircraft is constructed that is updated 10 frames per second. Traverse all pairs of aircraft in the dynamic aircraft spatial distribution map, calculate the Euclidean distance and relative velocity vector, and generate an N×N dimensional aircraft spacing matrix, where N is the total number of aircraft in the airspace; The aircraft spacing matrix, together with the aircraft type, physical size and maximum maneuver overload capability attributes in the airspace situation dataset, are encoded as graph node features and edge weights, and input into a graph attention network model composed of three stacked multi-head graph attention layers. The graph attention network model is trained under supervision based on multiple real or simulated close-range events in a historical conflict case library, and outputs a collision risk score for each pair of aircraft in the 0 to 1 range. Aircraft pairs with scores greater than a preset threshold are marked as high-risk interaction pairs. Combined with the risk levels in the list of abnormal aircraft identifiers, a set of high-risk interaction pairs containing aircraft ID pairs, collision risk scores, and associated abnormal identifiers is obtained.

4. The AI-based low-altitude airspace information central collaborative control method according to claim 1, characterized in that, Based on the associated aircraft and their impact range in the high-risk interaction set, and combining a multi-dimensional environmental voxel model and the current position of the aircraft, a trajectory prediction network is used to predict conflict trajectories and perform environmentally constrained trajectory replanning to generate compliant trajectory adjustment instructions, including: It receives echo intensity data from weather radar, interference power spectral density data from electromagnetic spectrum scanner, and three-dimensional static obstacle mesh data; The echo intensity data, interference power spectral density data and three-dimensional static obstacle mesh data are uniformly mapped to a 5×5×5 resolution three-dimensional voxel space through a voxelization engine, generating meteorological disturbance layer, electromagnetic interference layer and obstacle distribution layer respectively. In each layer, voxel datasets are recorded for horizontal wind speed, vertical wind speed, turbulence intensity, signal attenuation index, location of co-frequency interference source and static obstacle voxel occupancy status. Spatiotemporal interpolation is performed on the voxel dataset using a sliding time window, and the evolution of the environmental state within a preset time period is extrapolated using linear prediction or LSTM time series models to obtain a multidimensional environmental voxel model. The wind speed gradient threshold, signal attenuation region coordinates, and static obstacle envelope are extracted from the multi-dimensional environmental element model as safety constraint parameters. Based on the associated aircraft IDs and impact ranges in the high-risk interaction set, the current position and velocity in the airspace situation dataset are extracted and input into the trajectory prediction network in combination with safety constraint parameters. The trajectory prediction network encoder receives historical trajectory points within a preset time period, and the decoder predicts multiple trajectory points within a preset time period in the future to form conflict trajectory prediction. The predicted trajectory points in the conflict trajectory prediction are queried point by point to obtain the query results; The process involves querying the environmental parameters of the corresponding voxels in the multidimensional environmental voxel model point by point to determine whether there are any violations of wind speed gradient, signal strength, or obstacle occupancy constraints. If such violations exist, the trajectory replanner is invoked to perform environmental constraint trajectory replanning under the constraints of collision risk score and impact range of the high-risk interaction pair set. This generates a new waypoint sequence that conforms to environmental constraints, thereby generating an aircraft compliant trajectory adjustment instruction that includes avoidance strategy, time window, and three-dimensional path points.

5. The AI-based low-altitude airspace information central collaborative control method according to claim 4, characterized in that, Based on the compliant trajectory adjustment instructions, communication channel quality is assessed using a directional antenna array and a convolutional autoencoder, and communication parameters are adaptively adjusted to generate communication parameter optimization instructions, including: Extract the three-dimensional path points and time windows from the aircraft's compliant trajectory adjustment instructions, and calculate the line-of-sight path between each waypoint and the nearest low-altitude private network base station using a three-dimensional city model to determine the optimal communication echo sampling time and the ground station receiving azimuth angle. According to the compliant trajectory adjustment instruction, at the sampling time, the base station antenna is activated by the directional antenna array so that its main lobe is aligned with the azimuth angle, receives the downlink communication signal of the aircraft, and generates the original echo signal stream with microsecond-level time and frequency markers; The original echo signal stream is input into a four-layer 1D convolutional autoencoder through a convolutional autoencoder. Multipath effect, co-channel interference and atmospheric noise features are extracted through reconstruction error to perform communication channel quality assessment and obtain a communication channel quality index vector including CINR, delay spread and Doppler shift. Based on the communication channel quality index vector, the aircraft's transmission power, modulation method, and retransmission strategy are adaptively adjusted according to a preset mapping table. Combined with the avoidance strategy and time window in the compliant trajectory adjustment command, a communication parameter optimization command synchronized with the trajectory adjustment is generated.

6. The AI-based low-altitude airspace information central collaborative control method according to claim 5, characterized in that, The compliant trajectory adjustment instructions and communication parameter optimization instructions are jointly encoded and classified to generate lists of normal and abnormal aircraft states. Multi-agent game-theoretic scheduling and multi-department collaborative anomaly early warning responses are then executed to obtain a low-altitude airspace collaborative control and execution strategy, including: Extract the waypoint sequence, velocity profile and obstacle avoidance margin from the compliant trajectory adjustment command of the aircraft, as well as the transmit power, modulation mode and retransmission strategy from the communication parameter optimization command. Jointly encode the compliant trajectory adjustment command and the communication parameter optimization command into a 128-dimensional operating state feature vector, where the first 64 dimensions represent trajectory adjustment features and the last 64 dimensions represent communication parameter features. The operational status feature vector is input into an ensemble classifier consisting of three base models: a trained gradient boosting decision tree, a random forest, and a lightweight gradient boosting machine, and then classified. The ensemble classifier is trained based on historical normal and abnormal flight records and outputs a binary classification result of the aircraft's operational status through a weighted voting mechanism. Based on the binary classification results, and combined with the collision risk score in the high-risk interaction set and the risk level in the abnormal aircraft identifier list, a normal aircraft status list and an abnormal aircraft status list are generated. For each item in the abnormal aircraft status list, extract the aircraft ID, operational status feature vector, and risk level. Establish a triple relationship between aircraft, event, and command through the Neo4j graph database, match it with a preset emergency response mode library, execute multi-department collaborative abnormal early warning response, and generate an emergency response plan that includes multi-departmental linkage actions, resource scheduling list, and strict time window constraints from public security, civil aviation, and emergency management. The aircraft ID and operating status feature vector in the normal aircraft status list are input into the multi-agent game scheduling optimizer. Multi-agent game scheduling is performed respectively, and the global optimal takeoff time slot, route assignment and altitude layer configuration are solved by Nash equilibrium to obtain the flight scheduling scheme. The emergency response plan and flight scheduling plan are loaded into the distributed task scheduling engine, decomposed into atomic operations, and distributed to the standard interfaces of public security, civil aviation and emergency management units through message queues to obtain the low-altitude airspace collaborative control and execution strategy.

7. The AI-based low-altitude airspace information central collaborative control method according to claim 6, characterized in that, The running state feature vector is input into an ensemble classifier consisting of three base models: a trained gradient boosting decision tree, a random forest, and a lightweight gradient boosting machine, and classified, including: Extract a training sample set containing both normal and abnormal flight records from the historical flight record database; For each flight record in the training sample set, 128-dimensional features, including waypoint sequence, velocity profile, obstacle avoidance margin, transmit power, modulation method, and retransmission strategy, are extracted and labeled with category labels to obtain a labeled training dataset. The labeled training dataset is divided into a training set and a validation set in a 7:3 ratio. The training set is used to train three base models: gradient boosting decision tree, random forest, and lightweight gradient boosting machine. Hyperparameters are then tuned on the training set to obtain the optimized three base models. The accuracy, recall, F1 score, and AUC of the three base models are evaluated on the validation set, and the weights of the three base models are determined to be 0.4, 0.3, and 0.3 respectively based on the validation set performance. The operational status feature vector is input into the three base models respectively to obtain the probability prediction value of each base model. The probability prediction values ​​are weighted and summed according to the weighted voting weight to obtain the final classification probability. If the final classification probability is greater than the preset threshold, it is determined to be an abnormal operation class; otherwise, it is determined to be a normal operation class, so as to output the binary classification result of the aircraft's operational status.

8. The AI-based low-altitude airspace information central collaborative control method according to claim 6, characterized in that, The aircraft IDs and operational status feature vectors from the normal aircraft status list are input into a multi-agent game scheduling optimizer. Multi-agent game scheduling is then performed to solve for the globally optimal takeoff slots, route assignments, and altitude layer configurations using Nash equilibrium, including: Extract the aircraft ID, expected takeoff time, destination, range, aircraft performance parameters and operating status feature vector of each aircraft from the normal aircraft status list, construct a multi-agent game model, treat each aircraft as an independent agent and define its policy space, and obtain a policy space including takeoff time slot set, route set and altitude layer set; Define the utility function for each agent in the policy space; the utility function comprehensively considers takeoff time deviation cost, route length cost, altitude layer preference cost and conflict risk cost, and sets constraints based on the current airspace capacity dynamic quota, to obtain a game model that includes the utility function and constraints; The game model is solved for Nash equilibrium using an iterative optimal response algorithm. Each agent is initialized with a random combination of strategies. In each iteration, each agent selects the optimal strategy that maximizes its own utility function while the strategies of other agents are fixed, and updates it iteratively to obtain the Nash equilibrium strategy. Verify whether the Nash equilibrium strategy satisfies all constraints. If it does, output the globally optimal takeoff time slot, route assignment, and altitude layer configuration for each aircraft. Combine the globally optimal takeoff time slot, route assignment, and altitude layer configuration into a flight scheduling scheme to obtain a detailed flight plan that includes takeoff time, waypoint sequence, cruising altitude, and estimated arrival time.

9. The AI-based low-altitude airspace information central collaborative control method according to claim 6, characterized in that, For each item in the abnormal aircraft status list, the aircraft ID, operational status feature vector, and risk level are extracted. A triplet relationship between aircraft, events, and commands is established using the Neo4j graph database. This is matched against a pre-defined emergency response model library, and multi-departmental collaborative abnormal early warning responses are executed. This generates an emergency response plan that includes multi-departmental coordinated actions, resource scheduling lists, and strict time window constraints involving public security, civil aviation, and emergency management. Extract the aircraft ID, operational status feature vector, risk level, abnormal event type, and occurrence time for each item from the abnormal aircraft status list; Using the aircraft ID as the aircraft node, the abnormal event type as the event node, and the compliant trajectory adjustment command and communication parameter optimization command as the command node, a triple relationship of aircraft-event-command is established in the Neo4j graph database. The aircraft node and the event node are connected through the occurrence relationship, and the event node and the command node are connected through the trigger relationship. Based on the risk level and the type of abnormal event, the corresponding emergency response mode is matched from the preset emergency response mode library. The emergency response mode library contains response processes classified by risk level and event type, including Level I response mode, Level II response mode and Level III response mode. For Level I response mode, a mild response procedure is generated, which includes air traffic control issuing a warning and the aircraft performing autonomous avoidance. For Level II response mode, a moderate response procedure is generated, which includes air traffic control issuing a mandatory instruction, civil aviation coordinating airspace resources, and the aircraft performing mandatory avoidance. For Level III response mode, a high-altitude response procedure is generated, which includes air traffic control activating the emergency plan, management departments intervening in the control, civil aviation coordinating airspace resources, emergency management departments preparing for emergency rescue, and the aircraft performing an emergency landing. Based on the matched emergency response mode, execute multi-departmental collaborative abnormal early warning response and generate a list of multi-departmental joint actions; Generate a resource scheduling list, specifying the number of personnel, equipment types, vehicle configurations, and communication channels to be scheduled for each department, and determine the resource deployment locations based on the current location and predicted trajectory of the abnormal aircraft; Strict time window constraints are set according to the urgency of the abnormal event to generate an emergency response plan that includes multi-departmental coordinated actions, resource scheduling lists, and time window constraints.

10. A computer-readable storage medium, characterized in that, Used to store computer-readable instructions, which, when read by a computer, enable the execution of an AI-based low-altitude airspace information central collaborative control method as described in any one of claims 1-9.

Images