An airport multi-type unmanned device safety constraint cooperative planning method

Abstract

The application belongs to the technical field of intelligent operation control, and specifically discloses a safety constraint collaborative planning method for multiple types of unmanned equipment in an airport, which comprises the following steps: collecting airport area environment data, aircraft operation data, unmanned equipment state data and tower dispatching data, performing time synchronization and coordinate unification, and then constructing an airport global state; generating an airport safety constraint set based on the airport global state, and performing abnormal data filtering, regional safety level marking and authorized state marking on the airport global state to obtain a compliant environment state matrix; and the application aims to solve the technical problems in the existing collaborative planning process of multiple types of unmanned equipment in an airport, such as non-unified multi-source states, difficult pre-execution of safety constraints, mismatch between action outputs of heterogeneous unmanned equipment and physical boundaries, inability of tower authorization to participate in action locking, and lack of traceable records in the collaborative decision-making process.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation control technology, specifically to a method for collaborative planning of safety constraints for multiple types of unmanned equipment at airports. Background Technology

[0002] In civil aviation operation scenarios such as airport flight areas, aprons, taxiways, and service lanes, this field involves collaborative planning and safety management technologies for various types of unmanned equipment. As unmanned guided vehicles, unmanned cleaning vehicles, unmanned refueling vehicles, unmanned baggage transfer vehicles, unmanned shuttle buses, unmanned lawnmowers, and low-altitude inspection drones gradually participate in airport operations, the path planning objects in airport operations have expanded from single vehicles to a heterogeneous group of unmanned equipment coexisting with ground vehicles and aerial drones. This type of scenario differs from the scheduling of unmanned equipment in ordinary parks or roads; its operating area simultaneously faces constraints such as aircraft taxiing, parking space support, temporary fencing, personnel entry, tower dispatch instructions, and high-risk area authorization. Safety rules need to be recorded and verified during the planning and execution process. Existing collaborative planning methods typically focus on path length, task completion efficiency, or local obstacle avoidance results. They lack a unified time reference and coordinate representation for airport environmental data, aircraft operation data, unmanned equipment status data, and tower scheduling data. This results in subsequent planning being based on data from different points in time or different coordinate systems, affecting action judgment results. While some reinforcement learning collaborative algorithms can output multi-agent actions, they usually treat safety constraints as post-event rewards or penalties, lacking mechanisms for deterministically judging restricted areas, collision avoidance distances, aircraft priority passage, equipment action boundaries, and tower authorization status before action execution. Furthermore, there are differences in speed, steering, and braking between ground-based unmanned vehicles and aerial inspection drones. Significant differences exist in the movement, altitude, and avoidance methods of unmanned equipment. Using the same action space and training parameters can easily lead to a mismatch between strategy output and the physical execution capabilities of the equipment. Furthermore, airport sensor data is susceptible to location drift, momentary radar false detections, and short-term obstruction by dynamic obstacles. Without pre-planning anomaly filtering and area safety level labeling, the algorithm may generate actions based on unreliable conditions. For high-risk areas such as runway perimeter, taxiway perimeter, and takeoff and landing airspace, if planned actions are not machine-matched with tower authorization instructions, it will be difficult for unmanned equipment to meet airport scheduling and control requirements, and it will be difficult to generate decision logs, operational trajectories, interlock records, and reward / penalty data for post-event review. Therefore, collaborative planning for multiple types of unmanned equipment at airports requires integrating multi-source operational states, safety constraints, pre-action judgment, tower authorization interlocks, and operational record storage into a single processing flow. Summary of the Invention

[0003] This invention provides a collaborative planning method for safety constraints of multiple types of unmanned equipment in airports. The invention aims to solve the technical problems in the existing collaborative planning process for multiple types of unmanned equipment in airports, such as inconsistent multi-source states, difficulty in pre-executing safety constraints, mismatch between the action output of heterogeneous unmanned equipment and physical boundaries, inability of tower authorization to participate in action interlocking, and lack of traceable records in the collaborative decision-making process.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A collaborative planning method for safety constraints of multiple types of unmanned equipment at an airport includes: collecting airport area environmental data, aircraft operation data, unmanned equipment status data, and tower scheduling data; synchronizing time and unifying coordinates to construct an airport global state; generating an airport safety constraint set based on the airport global state; filtering abnormal data, marking regional security levels, and marking authorized statuses on the airport global state to obtain a compliant environment state matrix; inputting the compliant environment state matrix into a safety constraint improved MADDPG algorithm to generate original candidate actions for each unmanned equipment; performing pre-safety discrimination on the original candidate actions based on the airport safety constraint set to obtain safe candidate actions; matching the safe candidate actions with the tower scheduling data for scheduling status and authorization; issuing execution control commands after successful matching; and storing corresponding decision logs, operating trajectories, interlock records, and reward / penalty data.

[0005] In one aspect of the invention, the airport area environmental data includes the boundaries of the runway area, taxiway area, aircraft stand area, service lane area, temporary enclosure area, construction closure area, fixed structure locations, and dynamic obstacle locations. The aircraft operation data includes aircraft position, aircraft heading, aircraft speed, current aircraft operating phase, and aircraft passage priority; the unmanned equipment status data includes unmanned equipment identifier, unmanned equipment type, unmanned equipment position, unmanned equipment heading, unmanned equipment speed, unmanned equipment remaining battery power or remaining fuel, unmanned equipment mission status, unmanned equipment executable action boundaries, and unmanned equipment fault status. The tower dispatch data includes temporary control areas, temporary control time windows, aircraft priority passage instructions, unmanned equipment operation priority instructions, high-risk area operation authorization instructions, and unmanned equipment operation suspension instructions.

[0006] In one aspect of the invention, the airport safety constraint set includes restricted area constraints, buffer zone speed limit constraints, collision avoidance distance constraints, aircraft priority passage constraints, tower authorization constraints, and equipment action boundary constraints. The restricted area constraint is used to restrict unmanned equipment from entering the core area of ​​the runway, the take-off and landing airspace, the core circle of aircraft parking positions, and the restricted areas temporarily designated by the tower dispatch platform; the collision avoidance distance constraint is used to limit the minimum safe distance between ground unmanned operation vehicles and aircraft, between ground unmanned operation vehicles, between aerial inspection drones and aircraft fuselages, and between aerial inspection drones. The tower authorization constraints are used to restrict unmanned equipment without valid authorization from entering the area around the runway, taxiway, takeoff and landing airspace, or high-risk areas designated by the tower dispatch platform.

[0007] In one aspect of the present invention, filtering abnormal data in the airport's global status includes: The location data of unmanned equipment is checked for speed constraints. When the estimated speed corresponding to the position change in adjacent decision cycles exceeds the sum of the maximum allowable speed and the speed anomaly judgment tolerance value for this type of unmanned equipment, the current positioning point is marked as an abnormal positioning point, and the previous valid positioning point is combined with the current speed estimate for short-term interpolation update. The dynamic obstacle data is continuously verified. When a dynamic obstacle appears only in a single decision cycle and is not confirmed by other data sources in two or more consecutive decision cycles, it is marked as a transient interference target. When a dynamic obstacle is detected in two or more consecutive decision cycles or confirmed by tower dispatch data, aircraft operation data or local perception data of unmanned equipment, it is marked as a valid dynamic obstacle.

[0008] In one aspect of the present invention, the compliance environment state matrix includes: area coordinates, area type, security level, speed limit value, traffic direction, authorization status, temporary control status, and dynamic obstacle status; When the same area corresponds to multiple security levels, the compliance environment state matrix is ​​generated according to the higher security level. When the tower dispatch data sets a certain area as a temporary control area, the area is written into the compliance environment status matrix as a prohibited or restricted area within the corresponding temporary control time window.

[0009] In one aspect of the present invention, the safety constraint improved MADDPG algorithm includes multiple unmanned equipment agents, each unmanned equipment agent is configured with a local Actor policy network, during the training phase the global airport status, the joint actions of all unmanned equipment agents and the tower scheduling instruction vector are received through a global Critic evaluation network, and during the execution phase each unmanned equipment agent outputs the original candidate action based on its local observation value. Among them, the original candidate actions for ground unmanned vehicles include linear velocity and angular velocity, while the original candidate actions for aerial inspection drones include flight speed, heading angle adjustment, and target flight altitude.

[0010] In one aspect of the present invention, performing a preliminary security assessment of the original candidate action based on the airport security constraint set includes: Predict the next cycle position of the unmanned equipment after it executes the original candidate action, and determine whether the next cycle position or the movement path formed from the current position to the next cycle position enters a restricted area, is less than the corresponding collision avoidance distance threshold, violates the aircraft priority passage rules, involves the tower authorization requirements of high-risk areas, and exceeds the boundary of the equipment action. When any judgment result is non-compliant, the original candidate action is marked as a non-compliant candidate action, and a safe alternative action is generated as the safe candidate action.

[0011] In one aspect of the invention, the safety alternative action is generated based on the type of unmanned equipment and the type of violation; When the unmanned equipment is a ground-based unmanned operating vehicle, the safety alternative actions include deceleration actions generated according to the maximum permissible braking deceleration, stopping and waiting actions, low-speed circling actions, or returning actions; when the unmanned equipment is an aerial inspection drone, the safety alternative actions include hovering actions, maintaining a safe altitude actions, ascending to a preset avoidance altitude actions, horizontal circling actions, or returning actions. The marked violation candidate actions, their corresponding safe alternative actions, and violation types are written into the experience replay pool to reduce the output probability of similar violation candidate actions in subsequent training.

[0012] In one aspect of the present invention, the safety constraint improved MADDPG algorithm is trained using a two-layer weighted reward function, which includes a safety reward item and a job reward item. The safety rewards include positive rewards for avoiding aircraft and effective dynamic obstacles, positive rewards for maintaining compliant collision avoidance distances, penalties for violations of restricted areas, penalties for violations of aircraft priority passage, and penalties for action lockouts. The task reward items include task path advancement rewards, task timing rewards, task completion rewards, and collaborative avoidance rewards; When ground-based unmanned vehicles and aerial inspection drones train their local Actor policy networks, they use different local Actor policy network learning rates but the same global Critic evaluation network learning rate.

[0013] Compared with the prior art, the present invention has the following beneficial effects: This invention unifies airport environmental data, aircraft operation data, unmanned equipment status data, and tower scheduling data to a single time reference and coordinate system. The unified global airport status then generates an airport safety constraint set and a compliance environment status matrix, enabling subsequent collaborative planning to process data inputs based on the same time reference and coordinate system, reducing path judgment errors caused by asynchronous data source times or inconsistent coordinates. Furthermore, the improved MADDPG algorithm for safety constraints does not directly output execution control commands. Instead, it first generates initial candidate actions for each unmanned equipment, then uses constraints such as restricted areas, collision avoidance distances, aircraft priority passage, tower authorization, and equipment action boundaries for pre-emptive safety judgment. Non-compliant actions are converted into safe candidate actions before execution, ensuring that the candidate actions output by reinforcement learning are validated against airport operation rules before execution. Moreover, safe candidate actions need to be matched with tower scheduling data for scheduling status and authorization. Only after successful matching is the execution control command issued, ensuring that the behavior of unmanned equipment in high-risk areas such as runway perimeters, taxiway perimeters, and takeoff and landing airspace is included in tower scheduling control. By synchronously storing decision logs, operational trajectories, interlock records, and reward / penalty data, this solution provides data records for subsequent operational reviews, accountability tracking, model parameter adjustments, and safety log management. Overall, this solution establishes a unified processing chain in collaborative planning for multiple types of unmanned equipment, encompassing state input, safety constraint processing, candidate action discrimination, tower authorization interlocking, and execution data retention. This improves rule verification capabilities, scheduling adaptability, and operational record integrity in collaborative operations of airport unmanned equipment. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0015] Figure 1 This is a flowchart of a collaborative planning method for safety constraints of multiple types of unmanned equipment at airports, according to the present invention.

[0016] Figure 2 This is a flowchart of the sub-steps in step 1 of the safety constraint collaborative planning method for multiple types of unmanned equipment at airports according to the present invention.

[0017] Figure 3 This is a flowchart of a sub-step in step 2 of the safety constraint collaborative planning method for multiple types of unmanned equipment at airports according to the present invention.

[0018] Figure 4This is a flowchart of the sub-step of step 3 in the safety constraint collaborative planning method for multiple types of unmanned equipment at airports according to the present invention.

[0019] Figure 5 This is a flowchart of step 4 in the safety constraint collaborative planning method for multiple types of unmanned equipment at airports according to the present invention.

[0020] Figure 6 This is a framework diagram of a collaborative planning system for safety constraints of multiple types of unmanned equipment at airports, as proposed in this invention.

[0021] Figure 7 This invention presents the overall architecture and high-level data flow diagram of a collaborative planning system for safety constraints of multiple types of unmanned equipment in airports.

[0022] Figure 8 This is the core process timing interaction diagram of the safety constraint collaborative planning system for multiple types of unmanned equipment in airports according to the present invention. Detailed Implementation

[0023] The present invention will be further described below with reference to embodiments. These embodiments are merely illustrative of possible implementations of the invention and are not intended to limit the scope of protection of the invention. Those skilled in the art, based on the content described in this specification, can make equivalent substitutions for the number of unmanned devices, airport area size, sensor models, communication links, edge server configurations, local policy network structures, training batch sizes, log storage formats, and tower scheduling interfaces without departing from the technical concept of the present invention. The following embodiments illustrate possible implementations of a safety constraint collaborative planning method and system for multiple types of unmanned equipment at an airport.

[0024] like Figure 1 As shown in this embodiment, a collaborative planning method for safety constraints of multiple types of unmanned equipment at airports is disclosed, wherein the method specifically includes: S1. Collect airport area environmental data, aircraft operation data, unmanned equipment status data and tower dispatch data, and after time synchronization and coordinate unification, construct the airport global status; S2. Generate an airport security constraint set based on the airport global state, and perform abnormal data filtering, regional security level marking, and authorization status marking on the airport global state to obtain a compliance environment state matrix; S3. Input the compliant environment state matrix into the safety constraint improved MADDPG algorithm to generate the original candidate actions of each unmanned device, and perform preliminary safety discrimination on the original candidate actions based on the airport safety constraint set to obtain safe candidate actions; S4. Match the safety candidate actions with the tower scheduling data for scheduling status and authorization. After successful matching, issue execution control commands and store the corresponding decision logs, running trajectories, interlock records and reward / penalty data.

[0025] In this embodiment, it should be noted that in S1, the system does not simply collect the location of a single unmanned device (UAV), but rather unifies airport environment data, aircraft operation data, UAV status data, and tower dispatch data into the same time base and airport coordinate system. Airport environment data describes UAV-accessible areas, prohibited areas, temporarily closed areas, and dynamic obstacles. Aircraft operation data describes aircraft position, heading, speed, operational phase, and passage priority. UAV status data describes each UAV's device identifier, device type, position, heading, speed, remaining battery or fuel, mission status, executable action boundaries, and fault status. Tower dispatch data describes temporary control areas, control time windows, high-risk area authorization status, and UAV pause operation instructions. Since the above data are not entirely consistent in terms of collection frequency, coordinate source, and latency, time synchronization and coordinate unification are necessary before constructing the overall airport status. If this process is omitted, subsequent safety constraint sets may be based on data from different times or coordinate systems, leading to inconsistent action judgment objects. Therefore, the output airport global state of S1 is the input basis for S2 to generate the airport security constraint set and compliance environment state matrix.

[0026] In S2, the system generates a set of airport safety constraints based on the global airport state and transforms airport operation rules into machine-executable data structures. The airport safety constraint set includes restricted area constraints, buffer zone speed limit constraints, collision avoidance distance constraints, aircraft priority passage constraints, tower authorization constraints, and equipment action boundary constraints. Restricted area constraints limit unmanned equipment (UAVs) from entering the runway core area, takeoff and landing airspace, the core circle of aircraft parking positions, and temporarily designated prohibited areas by the tower. Buffer zone speed limit constraints restrict UAVs to low-speed passage within taxiway side buffer zones, aircraft parking position perimeter buffer zones, and temporary control area perimeter buffer zones. Collision avoidance distance constraints limit the minimum distance between UAVs and aircraft, and between UAVs themselves. Aircraft priority passage constraints prioritize aircraft paths when there is a temporal or spatial conflict between UAV and aircraft paths. Tower authorization constraints verify whether operations in high-risk areas have obtained valid tower authorization. Equipment action boundary constraints prevent the MADDPG algorithm, an improvement to safety constraints, from generating actions exceeding the physical execution capabilities of the equipment. The compliance environment state matrix output by S2 includes not only the area location, but also the area type, security level, speed limit, traffic direction, authorization status, temporary control status, and dynamic obstacle status, which serve as the direct input for the S3 algorithm's decision-making.

[0027] In S3, the safety constraint-improved MADDPG algorithm receives the compliance environment state matrix output from S2. Instead of directly outputting the final execution control command, it first outputs the initial candidate actions for each unmanned device. These initial candidate actions are generated by the local Actor policy network of each unmanned device based on its local observations. For ground-based unmanned vehicles, the initial candidate actions include linear velocity and angular velocity. For aerial inspection drones, the initial candidate actions include flight speed, heading angle adjustment, and target flight altitude. After the initial candidate actions are generated, the system does not allow them to be directly issued to the devices without considering the safety constraint set. Instead, it first performs preliminary safety judgment based on the airport safety constraint set generated in S2. Preliminary safety judgment includes predicting the next cycle position, determining restricted areas, determining collision avoidance distance, determining aircraft priority passage, pre-determining tower authorization requirements, and determining device action boundaries. If the initial candidate action meets all judgment conditions, it is considered a safe candidate action. If any non-compliant judgment result exists, the initial candidate action is marked as a violation candidate action, and a safe alternative action is generated according to the unmanned device type and violation type. This safe alternative action is then used as a safe candidate action. Therefore, the output safety candidate actions of S3 have already undergone pre-safety judgment, but still need to enter S4 for scheduling state matching and authorization matching.

[0028] In S4, the system performs scheduling status and authorization matching between the safety candidate actions output from S3 and the tower dispatch data, and decides whether to issue execution control commands based on the matching results. For safety candidate actions within the regular operation area, the system still reads the tower dispatch data for scheduling status matching. The scheduling status matching includes whether there are any unmanned equipment suspension instructions, temporary control area conflicts, temporary control time window conflicts, and aircraft priority passage conflicts. If no such conflicts exist, the scheduling status matching is considered successful, and execution control commands are generated. For high-risk areas around the runway, taxiway, takeoff and landing airspace, or designated by the tower dispatch platform, the system further performs precise matching of authorized equipment identifier, authorized area, authorized time window, authorized operation type, and authorized status based on the scheduling status matching. The tower authorization command includes at least the authorized equipment identifier, authorized area, authorized time window, authorized operation type, and authorized status. Only when the unmanned equipment identifier, predicted path, current time, and operation type corresponding to the safety candidate action all match the tower authorization command will the equipment execution lockout layer issue execution control commands. If a match fails, the corresponding execution control command will not be issued, and a safe alternative action such as waiting, detouring, braking, hovering, or returning to base will be generated. The system will also write the airport's global status, original candidate actions, safe candidate actions, authorization matching results, interlock records, execution results, and reward / penalty data corresponding to each decision into the log, thereby forming a traceable record of collaborative operations of unmanned equipment at the airport.

[0029] like Figure 2 As shown, in one specific embodiment, S1 includes: S11. Collect environmental data of the airport area; S12. Collect aircraft operation data; S13. Collect status data of unmanned equipment; S14. Collect tower dispatch data; S15. Perform time synchronization and coordinate unification; S16. Construct the global state of the airport.

[0030] In this embodiment, it should be noted that for S11, the airport area environmental data includes the boundaries of the runway area, taxiway area, aircraft stand area, service lane area, temporary enclosure area, construction closure area, fixed structure locations, and dynamic obstacle locations. The runway area boundaries, taxiway area boundaries, aircraft stand area boundaries, and service lane area boundaries are derived from the airport's electronic flight area map and basic data from the airport operations management system. The temporary enclosure area boundaries and construction closure area boundaries are derived from real-time data transmitted from the tower dispatch platform or the airport construction management system. The fixed structure locations are derived from airport facility ledgers and on-site survey data. The dynamic obstacle locations are derived from local perception data from airport apron millimeter-wave radar, vehicle-mounted lidar, and aerial inspection drones. After collecting the above data, the system converts each area boundary into a polygonal boundary or a three-dimensional spatial volume boundary in a unified coordinate system, and records the area source, update time, and effective time window for each area, enabling subsequent safety constraint sets to be determined based on the effective area data.

[0031] For S12, aircraft operation data includes aircraft position, heading, speed, current operational phase, and passage priority. The current operational phase includes parking, pushback, taxiing, takeoff preparation, post-landing taxiing, and temporary avoidance. Aircraft passage priority is obtained from the air traffic control platform and the airport operations management system, derived from airport operations scheduling rules and real-time air traffic control instructions. When collecting aircraft operation data, the system does not treat aircraft merely as ordinary dynamic obstacles; instead, it writes the aircraft path as a high-priority operational path into the airport's global status. If there is a short-term deviation between the aircraft operation data and the apron millimeter-wave radar detection data, the system uses the aircraft operation data from the air traffic control platform as the primary data source and uses the apron millimeter-wave radar detection data as position confirmation data. This setting allows aircraft priority passage constraints to be invoked in subsequent steps.

[0032] For S13, the unmanned equipment status data includes unmanned equipment identification, unmanned equipment type, unmanned equipment location, unmanned equipment heading, unmanned equipment speed, unmanned equipment remaining battery or fuel, unmanned equipment mission status, unmanned equipment executable action boundaries, and unmanned equipment fault status. Unmanned equipment types include ground unmanned operation vehicles and aerial inspection drones. Ground unmanned operation vehicles include unmanned guidance vehicles, unmanned sweeping vehicles, unmanned refueling vehicles, and unmanned baggage transfer vehicles. Aerial inspection drones are used for low-altitude airport inspections, airport area status confirmation, or security patrols. The executable action boundaries of the unmanned equipment are derived from the equipment's factory technical parameters, airport on-site commissioning data, and a limited number of historical calibration data. Specifically, during the equipment access phase, the system reads the maximum speed, maximum angular velocity, maximum braking deceleration, minimum turning radius, maximum flight speed, flight altitude range, and hovering capability parameters provided by the equipment manufacturer, and then calibrates these parameters through low-speed operation tests, braking tests, and heading adjustment tests in the airport's closed test area. The calibrated parameters are written into the unmanned equipment status data for subsequent equipment action boundary constraints.

[0033] For S14, tower dispatch data includes temporary control areas, temporary control time windows, aircraft priority passage instructions, unmanned equipment operation priority instructions, high-risk area operation authorization instructions, and unmanned equipment operation suspension instructions. Temporary control areas and temporary control time windows originate from real-time instructions from the tower dispatch platform or airport operations management system. Aircraft priority passage instructions originate from flight takeoff and landing plans, aircraft pushback plans, taxiway instructions, and temporary tower dispatch instructions. High-risk area operation authorization instructions originate from electronic authorization records on the tower dispatch platform. Unmanned equipment operation suspension instructions originate from the tower dispatch platform, airport safety management system, or unmanned equipment fault monitoring system. After collecting tower dispatch data, the system does not treat it as simple log information but encodes it into data fields that can be accessed in subsequent safety constraint sets, pre-emptive safety judgments, dispatch status matching, and authorization matching processes.

[0034] For S15, time synchronization and coordinate unification are used to address the differences in sampling time and coordinate representation methods among different data sources. The system uses the tower dispatch platform clock or the airport edge server clock as a unified time reference, unifying airport area environmental data, aircraft operation data, unmanned equipment status data, and tower dispatch data into the same decision cycle. The decision cycle can be set to 50ms, which is derived from the historical response time calibration data of the airport edge server and the unmanned equipment controller in a closed test scenario. Specifically, under the test condition of 12 unmanned devices simultaneously connected, the average time for system statistical data collection, state fusion, candidate action output, and execution of interlocks is combined with the minimum refresh cycle of the unmanned equipment controller to determine 50ms as the decision cycle under simulation and small-scale deployment conditions. In actual airport deployment, if the edge server performance, communication link, or equipment control interface are different, the decision cycle can be adjusted to 50ms to 200ms using the same calibration method. The 50ms to 200ms range is derived from the refresh frequency of the equipment control interface and the airport edge server stress test data.

[0035] The decision cycle is determined using a limited number of historical test data. Specifically, within a closed airport test area, at least three different scales of unmanned equipment (UAV) access are selected. Data acquisition time, time synchronization time, coordinate transformation time, compliance environment state matrix update time, original candidate action output time, pre-emptive safety judgment time, and control command issuance time are recorded. The decision cycle is then selected based on the cycle value that covers most of the normal processing latency across multiple consecutive test rounds. For example, with 12 UAVs accessing the system, the average processing time over 2000 test cycles is 38ms, and under higher load, it is 46ms. The UAV controller refresh interval is 50ms, so the decision cycle is set to 50ms. With 35 UAVs accessing the system, the processing time under higher load is close to 118ms, so the decision cycle is adjusted to 150ms or 200ms.

[0036] For S16, the airport global state is the output of S1 and serves as the input to S2. The airport global state is not a single coordinate point, but a data set composed of aircraft operational status, unmanned equipment status, airport area status, tower scheduling status, and operational task status. The airport global state refers to a set of airport operational status data formed under the same time reference and the same airport coordinate system. Its input data includes airport area environmental data, aircraft operational data, unmanned equipment status data, tower scheduling data, and operational task status data. Its generation method involves timestamp alignment, coordinate transformation, data source marking, effective time window marking, and trusted status marking. Its output is structured state data used for generating airport safety constraint sets and updating the compliance environment state matrix. Its data type is a data set with timestamps and coordinate fields. Its usage stages are S2 constraint generation and S3 action discrimination. Its technical function is to enable subsequent safety constraints and collaborative planning to be executed based on the same spatiotemporal state. When constructing the airport global state, the system records the data source, timestamp, coordinate system, effective time window, and trusted status for each data entry. If the data comes from a fixed electronic map, its effective time window is determined according to the airport map version. If the data comes from the air traffic control tower, its effective time window is determined according to the dispatch instructions. If the data comes from sensors, its effective time window is determined according to the sensor refresh cycle and historical error calibration results. Through this processing, S2 can generate a set of airport safety constraints consistent with the current operating conditions based on the airport's global status, avoiding the use of outdated regional data or outdated dispatch instructions for action determination.

[0037] In some optional embodiments, regarding S1, it is assumed that the airport area size is 420m × 360m, which is derived from the area modeling parameters of the simulated airport electronic map. The system simultaneously connects to the airport apron millimeter-wave radar, vehicle-mounted lidar, Beidou positioning terminal, aerial inspection UAV positioning module, and tower dispatch platform. Within a certain decision cycle, the system receives information that an aircraft is in the taxiing phase, 8 ground unmanned operation vehicles are in operation, and 4 aerial inspection UAVs are in inspection mode, while simultaneously receiving a temporary control instruction for the taxiway side issued by the tower dispatch platform. The system converts all data to the airport's unified coordinate system, aligns the timestamps according to a 50ms decision cycle, and writes the aircraft taxiing path, temporary control area, UAV location, and UAV task status into the airport global state. At this time, the airport global state output by S1 includes the area boundary, aircraft path, UAV action boundary, and tower authorization status required for subsequent constraint generation.

[0038] like Figure 3 As shown, in one specific embodiment, S2 includes: S21. Divide the airport into functional zones; S22. Input the graded safety distance threshold and equipment action boundary threshold; S23. Generate a set of airport security constraints; S24. Filter abnormal data in the overall airport status; S25. Mark the regional security level and authorization status of the airport's overall status; S26. Generate a compliant environment state matrix.

[0039] In this embodiment, it should be noted that for S21, the airport functional areas include restricted areas, buffer warning areas, and routine operation areas. Restricted areas include the runway core area, takeoff and landing airspace, the core circle of aircraft parking positions, and temporarily designated prohibited areas by the control tower. The takeoff and landing airspace is a three-dimensional spatial area with horizontal projection boundaries and altitude ranges. The horizontal projection boundaries are derived from the airport's electronic flight area map and airspace control data from the control tower, while the altitude ranges are derived from the airport's low-altitude safety control rules and the control requirements of the control tower for takeoff and landing-related airspace. The restricted areas are obtained from the airport's electronic flight area map, airport operation safety rules, and temporary control instructions from the control tower. Buffer warning areas include taxiway side buffer zones, aircraft parking position perimeter buffer zones, aircraft proximity movement areas, and temporary control area perimeter buffer zones. The taxiway side buffer zone can be set to 3 meters, which is derived from airport on-site safety operation rules, taxiway boundary calibration data, and historical manual avoidance distance calibration data. The buffer zone around the aircraft stand can be set at 5 meters. This 5 meters is derived from airport aircraft stand operation safety management rules, historical traffic trajectories of aircraft stand operation vehicles, and on-site operation risk assessment data. Regular operation areas include service lanes, outer pavement, non-takeoff and landing airspace, and other areas where unmanned equipment is permitted to operate routinely.

[0040] The width of the buffer zone is determined by combining airport area boundary calibration data, historical manual driving trajectories, and on-site risk assessment data. The specific process is as follows: First, the boundaries of taxiways, aircraft stands, and service lanes are read from the airport electronic map. Then, the minimum lateral distances of historical operating vehicles near the corresponding boundaries are calculated, and the buffer width is determined by combining the positioning error and braking control error of unmanned equipment. For example, if the minimum lateral distance between manual vehicles and the boundary in historical operating trajectories along a taxiway side is mainly distributed between 2.4m and 3.6m, and the positioning error of unmanned equipment does not exceed 0.3m, then the buffer zone along the taxiway side is set to 3m. If the historical safe distance between operating vehicles outside an aircraft stand and the boundary of the aircraft support area is mainly distributed between 4.2m and 5.8m, then the buffer zone outside the aircraft stand is set to 5m.

[0041] For S22, graded safety distance thresholds are used for collision avoidance distance constraints. The safety distance threshold between ground-based unmanned vehicles and aircraft can be set to 8m. This 8m is derived from airport safety management rules for vehicles operating close to aircraft, aircraft fuselage outline protection requirements, and historical airport manual operation distance records. The safety distance threshold between ground-based unmanned vehicles and aircraft is jointly determined by the aircraft outline protection distance, unmanned equipment positioning error, braking distance, and airport manual operation records. The specific process is as follows: first, read the aircraft fuselage outline and the boundary of the aircraft stand protection area; then, calculate the braking distance of the unmanned equipment at the maximum permissible speed and the commonly used low-speed operating speed; and then add the positioning error, control response delay distance, and safety margin. For example, the test braking distance of an unmanned baggage transfer vehicle at a speed of 6m / s is 4.6m, the positioning error is calculated as 0.4m, the control response delay is converted to a distance of 0.6m, and after adding a 2.4m airport operation safety margin, 8m is obtained as the safety distance threshold between ground-based unmanned vehicles and aircraft.

[0042] The safe distance threshold between two unmanned ground vehicles can be set to 3m. This 3m is derived from braking test data, low-speed following test data, and airport service lane width conditions. The safe distance threshold between unmanned ground vehicles is determined through low-speed following tests, oncoming vehicle tests, and service lane width conditions. Specifically, within the airport's closed service lane, two unmanned ground vehicles are controlled to operate at their usual operating speed and maximum permissible low speed, respectively. Braking response, obstacle avoidance success rate, and lateral deviation are recorded at different vehicle distances. The minimum distance that can cover positioning errors and control delays is selected as the threshold. For example, in a 2m / s low-speed following test, multiple sudden stops occurred when the distance between the two vehicles was 2m, while a distance of 3m allowed for deceleration and stopping, and the service lane still maintained sufficient space for passage. Therefore, the safe distance threshold between the two unmanned ground vehicles is set to 3m.

[0043] The safe distance threshold between an aerial inspection drone and the aircraft fuselage can be set to 12m. This 12m is derived from the drone's positioning error test data, rotor disturbance safety distance assessment data, and aircraft outline protection requirements. The safe distance threshold between the aerial inspection drone and the aircraft fuselage is determined by the drone's positioning error, hovering drift, rotor disturbance range, sensor imaging distance, and aircraft outline protection requirements. Specifically, during low-altitude inspection tests, the drone's hovering position deviation, position drift under lateral wind disturbance, and return-to-home response distance are recorded. This data is then combined with the aircraft outline protection boundary to determine a conservative safe distance. For example, if an aerial inspection drone has a maximum drift of 1.2m when hovering at an altitude of 8m, a positioning error of 0.5m, a horizontal avoidance response distance of 3.5m, and an aircraft outline protection margin of 6.8m, then the safe distance threshold between the aerial inspection drone and the aircraft fuselage is set to 12m.

[0044] The collision avoidance distance threshold between two aerial inspection drones can be set to 5m. This 5m is derived from drone positioning error tests, hovering drift tests, and low-altitude inspection avoidance test data. The collision avoidance distance threshold between aerial inspection drones is determined through drone positioning error, hovering drift, heading adjustment response distance, and low-altitude inspection formation tests. Specifically, two drones are set up within the same low-altitude inspection area to perform opposite, same-direction, and cross-direction inspection tasks. The minimum distance, avoidance success rate, and heading adjustment time at different intervals are statistically analyzed, and the distance that covers the positioning error and drift of both drones is used as the threshold. For example, if two drones can complete horizontal maneuvering or hovering at a 5m interval, but experience multiple emergency stops at a 3m interval, then the collision avoidance distance threshold between the two aerial inspection drones is set to 5m.

[0045] The maximum speed of the ground-based unmanned vehicle can be set to 6 m / s, which is derived from the technical parameters of the ground-based unmanned vehicle manufacturer and airport closed road test data. The maximum angular velocity of the ground-based unmanned vehicle can be set to 0.3 rad / s, which is derived from the chassis controller parameters of the ground-based unmanned vehicle and airport low-speed turning test data. The maximum flight speed of the aerial inspection drone can be set to 10 m / s, which is derived from the technical parameters of the aerial inspection drone manufacturer and airport low-altitude inspection test data. The operating altitude range of the aerial inspection drone can be set to 5 m to 18 m, which is derived from the airport low-altitude inspection operation plan, sensor imaging distance testing, and takeoff and landing airspace avoidance requirements. The takeoff and landing airspace altitude range of 0 m to 4 m can be set as a prohibited entry altitude range, which is derived from the airport low-altitude safety control rules and the control requirements of the tower dispatch platform for takeoff and landing related airspaces.

[0046] For S23, the airport safety constraint set includes restricted area constraints, buffer zone speed limit constraints, collision avoidance distance constraints, aircraft priority passage constraints, tower authorization constraints, and equipment action boundary constraints. The airport safety constraint set refers to a data set of constraint rules formed by airport operation rules, area boundaries, equipment action boundaries, tower scheduling instructions, and safety distance thresholds. Its input data includes the airport's global status, the airport safety rule configuration table, and tower scheduling data. It is generated by encoding restricted areas, buffer zone speed limits, collision avoidance distances, aircraft priority passage, tower authorization, and equipment action boundaries into rule fields that can be judged by a computer. Its output is a set of rules for use in the S3 pre-safety judgment process. Its data type is a combination of area rules, distance rules, speed rules, authorization rules, and equipment action boundary rules. Its usage stages are original candidate action judgment and execution release review. Its technical function is to transform airport operation safety rules into machine-executable action constraints. Restricted area constraints are used to determine whether the predicted position or predicted path of the unmanned equipment enters a restricted area. Buffer zone speed limit constraints are used to determine whether the speed of the unmanned equipment within the buffer warning area exceeds the speed limit value. The speed limit for unmanned ground vehicles within the buffer warning area can be set to 2 m / s. The 2 m / s is derived from braking distance tests, slow-speed passage rules for service lanes, and low-speed control stability tests of unmanned equipment in airport closed test scenarios.

[0047] The speed limit for the buffer warning zone is determined through braking distance testing, low-speed control stability testing, and airport slow-speed passage rules. The specific process involves controlling unmanned ground vehicles at different speeds within the taxiway side buffer zone and the aircraft stand perimeter buffer zone, recording braking distance, positioning deviation, steering overshoot, and stopping response time, and selecting the speed at which deceleration or stopping can be completed within the buffer zone width as the speed limit. For example, if an unmanned sweeper has a large braking distance and steering overshoot at 3 m / s, but can decelerate and maintain stable control within the buffer zone at 2 m / s, then the speed limit for the buffer warning zone is set to 2 m / s.

[0048] Collision avoidance distance constraints determine whether the distance between unmanned equipment and aircraft, between unmanned equipment and other unmanned equipment, and between unmanned equipment and effective dynamic obstacles is less than a corresponding threshold. Aircraft priority passage constraints prioritize aircraft passage paths in case of path conflicts. Tower authorization constraints restrict operations in high-risk areas without valid authorization. Equipment action boundary constraints limit the algorithm's output actions to not exceeding the equipment's actual execution capabilities.

[0049] For S24, abnormal data filtering includes unmanned equipment position anomaly filtering and dynamic obstacle continuity verification. For unmanned equipment position data, the system compares the position changes between adjacent decision cycles with the equipment's maximum permissible speed to determine if the current location point is abnormal. If the calculated speed corresponding to a certain location point exceeds the sum of the equipment's maximum permissible speed and the speed anomaly judgment tolerance value, then the location point is marked as an abnormal location point. The speed anomaly judgment tolerance value can be set to 0.5 m / s, where 0.5 m / s is derived from static drift tests of Beidou positioning terminals, historical error statistics of vehicle-mounted positioning modules, and repeated test data from airport closed areas.

[0050] The tolerance value for speed anomaly detection is determined through static drift testing of the positioning terminal, repeated low-speed driving testing, and historical positioning error statistics. Specifically, the process involves recording the calculated speed based on position changes over adjacent decision cycles in three states: stationary, uniform low-speed driving, and turning. The deviation range of this calculated speed relative to the feedback speed from the device controller is then statistically analyzed. For example, in a static test, the maximum calculated speed due to positioning drift of a certain Beidou positioning terminal is 0.28 m / s, while in a low-speed driving test, the maximum speed deviation is 0.41 m / s. Considering the influence of different device installation locations and localized obstructions at the airport, the tolerance value for speed anomaly detection is set to 0.5 m / s.

[0051] For dynamic obstacle data, if a dynamic obstacle appears only in a single decision cycle and is not confirmed by other data sources in the subsequent two consecutive decision cycles, it is marked as a transient interference target. The two consecutive decision cycles are derived from historical false detection data from airport apron millimeter-wave radar and the target persistence calibration results under a 50ms decision cycle. If a dynamic obstacle is detected in more than two consecutive decision cycles, or confirmed by tower dispatch data, aircraft operation data, or local perception data from unmanned equipment, it is marked as a valid dynamic obstacle.

[0052] The continuous obstacle detection cycle is determined by millimeter-wave radar false detection records, multi-source perception detection confirmation records, and decision cycle length. Specifically, the detection duration of personnel, vehicles, temporary objects, and rain / fog interference targets is recorded within the airport apron test area. The proportion of targets that appear for only one decision cycle and are not confirmed by other data sources is statistically analyzed, and the impact of two consecutive cycle confirmations versus three consecutive cycle confirmations on obstacle avoidance response latency is compared. For example, with a 50ms decision cycle, single-cycle false detection targets account for the majority of false detection records. Targets detected for two consecutive cycles and confirmed by vehicle-mounted LiDAR or UAV local perception are usually real obstacles; therefore, two consecutive decision cycles are used as a confirmation condition for valid dynamic obstacles.

[0053] For S25, the area safety level marker is used to convert different areas into algorithm-readable status information. Restricted areas are marked with the highest safety level, buffer warning areas with an intermediate safety level, and routine operation areas with a basic safety level. If the same area belongs to multiple area types, it is marked according to the highest safety level. If the tower dispatch platform sets a routine operation area as a temporary control area within a certain time window, that area is marked with the safety level specified by the tower dispatch data within the corresponding time window. The authorization status marker describes whether a certain area or operation has valid authorization. The authorization status comes from the electronic authorization instructions of the tower dispatch platform, which include the authorized device identifier, authorized area, authorized time window, authorized operation type, and authorization status. If a safety candidate action requires entering a high-risk area, subsequent S4 must call the authorization status marker for matching.

[0054] For S26, the compliance environment state matrix is ​​the output of S2 and serves as the input to S3. The compliance environment state matrix refers to the state matrix formed after rasterizing the airport area according to the airport's unified coordinate system. Its input data includes the airport's global state, airport safety constraint set, area security level markers, authorization status markers, and dynamic obstacle status. It is generated by writing the area coordinates, area type, security level, speed limit, direction of travel, authorization status, temporary control status, and dynamic obstacle status corresponding to each grid cell into the corresponding matrix unit. Its output is environmental state data used for local observation extraction, original candidate action generation, and pre-emptive safety judgment. Its data type is a two-dimensional or three-dimensional raster matrix. Its usage stages are S3 algorithm decision-making and pre-emptive safety judgment. Its technical function is to transform the airport operating environment into a structured input that the algorithm can read and execute for safety rule judgment. The compliance environment state matrix includes area coordinates, area type, security level, speed limit, direction of travel, authorization status, temporary control status, and dynamic obstacle status. The compliance environment state matrix is ​​stored in a raster matrix format. A list of continuous area objects serves as an area boundary index table, establishing a correspondence with the raster numbers, but does not replace the compliance environment state matrix itself. In the raster matrix format, each raster records area coordinates, area type, security level, speed limit, direction of travel, authorization status, temporary control status, and dynamic obstacle status. In the area boundary index table, each area object records boundary coordinates, area type, security level, and effective time window, used to assist in updating the status fields of the corresponding raster. The system can select the raster resolution based on the computing power of the airport edge server and the accuracy of the airport area map. The compliance environment state matrix must be readable by the S3 security constraint improvement MADDPG algorithm and used for preliminary security judgment of subsequent original candidate actions.

[0055] In some optional embodiments, regarding S2, assuming the system receives a temporary control instruction from the control tower, this instruction sets a 3m area on the side of a taxiway as a buffer warning zone, with a control time window of 10 minutes. This 10 minutes originates from the temporary control instruction issued by the control tower platform for the taxiing phase of that flight. The system writes this area into the airport safety constraint set and increases the corresponding area's safety level in the compliance environment state matrix, while also writing a 2m / s speed limit value. This 2m / s value originates from the low-speed passage test data of the buffer warning zone and the airport's slow-speed passage rules. If the original candidate action of an unmanned cleaning vehicle within this area corresponds to a speed of 3m / s, then during the preliminary safety judgment in S3, this original candidate action will be marked as a speed limit violation action in the buffer zone, and a low-speed safety candidate action will be generated.

[0056] like Figure 4 As shown, in one specific embodiment, S3 includes: S31. Construct a set of agents for the safety-constrained improved MADDPG algorithm; S32. Generate local observation values ​​for each unmanned device; S33. Output the original candidate actions through the local Actor policy network; S34. Perform a preliminary safety assessment on the original candidate actions; S35. Generate safety candidate actions based on the results of the prior safety assessment; S36. Write the violation candidate actions and their corresponding safe candidate actions into the experience replay pool; S37. Update network parameters based on a two-layer weighted reward function and tower scheduling instruction vector.

[0057] In this embodiment, it should be noted that for S31, each unmanned device is encapsulated as an unmanned device agent. These agents include ground-based unmanned operational vehicle agents and aerial inspection drone agents. Each unmanned device agent is configured with a local Actor policy network. During the training phase, the global Critic evaluation network receives the airport's global state, the joint actions of all unmanned devices, and the tower scheduling command vector, and outputs action value evaluation results. During the execution phase, each unmanned device agent outputs original candidate actions based on its local observations. The improved MADDPG algorithm, based on the MADDPG centralized training and distributed execution framework, is a multi-agent collaborative decision-making method that adds airport safety constraint sets, tower dispatch command vectors, pre-emptive safety judgment, safety alternative action generation, and violation candidate action experience replay processing. Its input data includes a compliance environment state matrix, local observations, airport global state, joint actions, and tower dispatch command vectors. Its output includes original candidate actions, safety candidate actions, and network parameter update results from the training phase. Its application phase is S3 collaborative planning and decision-making. Its technical function is to subject the action generation of various types of unmanned equipment to airport safety rules and tower dispatch state constraints. This structure adopts the basic framework of MADDPG centralized training and distributed execution, and incorporates airport operation rules into the training and execution process through airport safety constraint sets, pre-emptive safety judgment, and tower dispatch command vectors.

[0058] For S32, local observations include the drone's position, speed, heading, remaining battery or fuel, mission status, relative positions of adjacent drones, relative positions of effective dynamic obstacles, local area type, local speed limit, and local authorization status. These local observations are derived from the airport global state constructed in S1 and the compliance environment state matrix generated in S2. Local observations do not require each drone to directly read all airport operational data; instead, they extract data relevant to the drone's decision-making from the airport edge server or the drone's local controller. This reduces the amount of data that a single drone needs to process and adapts to the distributed execution phase.

[0059] In S33, the local Actor policy network outputs raw candidate actions after receiving local observations. These raw candidate actions refer to action suggestions output by the local Actor policy network based on the local observations of the unmanned vehicle agent, which have not yet undergone pre-processing safety judgment based on airport safety constraints and have not yet been authorized and matched by the control tower. The data type varies depending on the type of unmanned vehicle: linear velocity and angular velocity for ground-based unmanned vehicles, and flight speed, heading angle adjustment, and target flight altitude for aerial inspection drones. For ground-based unmanned vehicles, raw candidate actions include linear velocity and angular velocity. For aerial inspection drones, raw candidate actions include flight speed, heading angle adjustment, and target flight altitude. The raw candidate actions output by the local Actor policy network only represent action suggestions generated by the algorithm based on current local observations and do not directly represent executable control commands. The local Actor policy network may output actions close to the safety boundary in the early stages of training or under unexpected conditions. If executed directly without pre-processing safety judgment, there is a risk of violating airport safety constraints. Therefore, the raw candidate actions must undergo safety judgment in S34.

[0060] Before the local Actor policy network outputs the original candidate actions, the system generates local observation vectors for the unmanned aerial vehicle (UAV) agent according to a fixed field order. These fixed fields include, in sequence, the UAV's position, speed, heading, remaining battery or fuel, mission status, relative position to adjacent UAVs, relative position to effective dynamic obstacles, local area type, local speed limit, local authorization status, and action boundary. For temporarily missing data that is not marked as anomalous, the system temporarily fills it with the previous valid value and simultaneously writes it into the missing data flag field. For data already marked as anomalous by S2, it is not included in the local observation vector but is replaced by a short-term predicted value or a safe alternative status field. After the local observation vectors are generated, the system normalizes the data according to the coordinate range, speed range, altitude range, battery range, and area level code range recorded in the airport safety rule configuration table, converting data from different units into numerical inputs readable by the local Actor policy network. The local Actor policy network outputs normalized action values, which the system then converts into physical action values ​​based on the type of unmanned equipment. For ground-based unmanned vehicles, these values ​​are converted into linear and angular velocities; for aerial inspection drones, they are converted into flight speed, heading angle adjustment, and target flight altitude. The converted actions are first truncated according to the equipment action boundary thresholds, and then used as original candidate actions in the preliminary safety assessment.

[0061] For S34, the pre-emptive safety assessment includes restricted area determination, collision avoidance distance determination, aircraft priority passage determination, tower authorization requirement pre-determination, and equipment action boundary determination. Restricted area determination is used to determine whether the predicted position or path of the unmanned equipment (UAV) enters a restricted area. Collision avoidance distance determination is used to determine whether the distance between the UAV and an aircraft, between UAVs, or between the UAV and a valid dynamic obstacle is less than the corresponding safe distance threshold. Aircraft priority passage determination is used to determine whether the predicted path of the UAV affects the aircraft's taxiing, pushback, takeoff preparation, or post-landing taxiing path. Tower authorization requirement pre-determination is used to determine whether the predicted path of the original candidate action enters the runway perimeter, taxiway perimeter, takeoff and landing airspace perimeter, or high-risk areas designated by the tower dispatch platform, and to determine whether a corresponding authorization status field exists in the compliance environment status matrix; precise matching of authorized equipment identifier, authorized area, authorized time window, authorized operation type, and authorized status is performed in S4. Equipment action boundary determination is used to determine whether the original candidate action exceeds the UAV's maximum speed, maximum angular velocity, flight altitude range, or braking capability boundary.

[0062] When performing preliminary safety assessments on original candidate actions, the system first predicts the position for the next cycle based on the current unmanned equipment position, current heading, current speed, original candidate actions, and decision cycle. For ground-based unmanned vehicles, the prediction process uses the vehicle's current position, linear velocity, angular velocity, and current heading to obtain the planar position and heading for the next cycle. For aerial inspection drones, the prediction process uses the current position, flight speed, heading angle adjustment, and target flight altitude to obtain the three-dimensional position for the next cycle. The system then compares the motion path from the current position to the next cycle position with the prohibited areas, buffer warning areas, effective dynamic obstacle states, and high-risk area authorization states in the compliance environment state matrix. If the same original candidate action triggers multiple violations simultaneously, the system determines the primary violation type in the following order: prohibited area violation, aircraft priority passage violation, collision avoidance distance violation, high-risk area authorization requirement violation, buffer area speed limit violation, and equipment action boundary violation. It then generates a safe alternative action based on the primary violation type and adds other violation types as additional violation tags to the lockout record and experience playback pool.

[0063] For S35, if the original candidate action meets all the pre-condition safety criteria, it is designated as a safe candidate action. If the original candidate action does not meet any of the pre-condition safety criteria, it is marked as a violation candidate action, and a safe alternative action is generated based on the type of unmanned equipment and the type of violation. The safe candidate action refers to the compliant action retained after the pre-condition safety criteria are met, or a safe alternative action generated by the system based on the type of unmanned equipment and the type of violation when the original candidate action is non-compliant. Its output still needs to enter S4 for tower dispatch status matching and authorization matching, and is not directly equivalent to the issued execution control command. Safe alternative actions for ground unmanned vehicles include deceleration, stopping and waiting, low-speed detour, or return-to-home actions. Safe alternative actions for aerial inspection drones include hovering, maintaining a safe altitude, ascending to a preset avoidance altitude, horizontal detour, or return-to-home actions. Low-speed detour actions are generated based on adjacent regular operating areas or low-safety-level areas in the compliant environment state matrix, avoiding restricted areas, effective dynamic obstacles, and priority aircraft paths. The horizontal avoidance maneuver is generated based on the UAV's current position, target direction, effective dynamic obstacle positions, and high-risk area boundaries, prioritizing horizontal avoidance paths that do not enter the prohibited altitude zone of the takeoff and landing airspace. The preset avoidance height can be set to 8m, which is derived from low-altitude operation tests of aerial inspection UAVs, sensor imaging altitude requirements, and airport takeoff and landing airspace avoidance rules. This preset avoidance height must not fall within the prohibited altitude zone of the takeoff and landing airspace.

[0064] The preset avoidance height is determined by considering the operating altitude range of the aerial inspection drone, sensor imaging distance, low-altitude safety control rules, and prohibited altitude zones in the take-off and landing airspace. The specific process involves first excluding prohibited altitude zones, then testing the drone's hovering stability, sensor imaging clarity, and horizontal maneuvering response time within the permissible operating altitude range, and finally selecting a height that balances inspection continuity and avoidance safety. For example, if the prohibited altitude zone is 0m to 4m, and the aerial inspection drone's operating altitude range is 5m to 18m, and the imaging distance and hovering stability meet the test requirements at an altitude of 8m, then 8m is selected as the preset avoidance height.

[0065] For S36, the system writes violation candidate actions, corresponding safe candidate actions, violation types, and reward / penalty data into the experience replay pool. The capacity of the experience replay pool can be set to 12,000 samples, which are derived from sample capacity tests and memory usage tests conducted on 1200 rounds of training on 12 unmanned devices in the airport digital twin simulation platform. The number of samples sampled in a single batch can be set to 256, which are derived from deep learning GPU training throughput tests and training stability calibration data. The experience replay pool retains both normal samples and violation candidate action samples, enabling the local Actor policy network to learn safe action boundaries, rather than relying solely on the device execution lockout layer as a fallback during the execution phase.

[0066] The capacity of the experience replay pool and the number of samples per batch were determined through tests of training sample coverage, memory usage, and training stability. Specifically, different experience replay pool capacities and batch sample numbers were set on a fixed hardware platform, and the retention rate of violation candidate actions, training convergence rounds, memory usage, and single parameter update time were statistically analyzed. For example, when the experience replay pool capacity was 6000 samples, early violation samples were covered quickly, but subsequent similar violation actions showed unstable decline. When the capacity was 12000 samples, normal and violation samples could be retained simultaneously, and memory usage remained within an acceptable range. A batch sample size of 128 samples resulted in significant gradient fluctuations, while 512 samples increased the single update time; therefore, 256 samples were selected as the single batch sample size.

[0067] Each sample written into the experience replay pool includes at least the current global airport state, current local observations, the joint actions of all unmanned aerial vehicles (UAVs), the original candidate actions of the current UAVs, safety candidate actions, violation types, tower dispatch command vectors, authorization matching results, reward values, the global airport state for the next period, and the local observations for the next period. During training, the system extracts batch samples from the experience replay pool according to a mixed ratio of normal samples and violation candidate action samples. If the number of violation candidate action samples is insufficient, the existing violation samples in the current experience replay pool are used, and no artificial forgery is performed. The global Critic evaluation network generates action value evaluation results based on the global airport state, joint actions, and tower dispatch command vectors. The local Actor policy network adjusts parameters based on the evaluation results of the global Critic evaluation network and action lockout penalty records to reduce the probability of outputting violation candidate actions again under the same state. After training is completed, the global Critic evaluation network is no longer called to output control actions during the execution phase. Instead, the local Actor policy networks of each UAV output original candidate actions based on their local observations, and release the execution control commands only after further pre-safety discrimination and authorization matching.

[0068] For S37, the system updates network parameters based on a two-layer weighted reward function and a tower dispatch instruction vector. The tower dispatch instruction vector refers to the numerical input formed after converting tower dispatch data into algorithm-readable fields. Its input data includes temporary control areas, temporary control time windows, aircraft priority passage instructions, unmanned equipment operation priority instructions, high-risk area operation authorization instructions, and unmanned equipment suspension operation instructions. It is generated by field-based processing according to area number, time window status, priority code, authorization status code, and suspension status code. Its output is used by the global Critic evaluation network and the pre-safety judgment process. Its technical role is to enable tower dispatch data to participate in action value evaluation and action locking judgment. The two-layer weighted reward function includes safety reward items and operational reward items. Safety reward items include positive rewards for avoiding aircraft and effective dynamic obstacles, positive rewards for maintaining compliant collision avoidance distances, penalties for violations in prohibited areas, penalties for violations of aircraft priority passage, and penalties for action locking. Operational reward items include task path advancement rewards, operation sequence rewards, task completion rewards, and cooperative avoidance rewards. The safety reward weight can be set to 0.85, where 0.85 comes from the weight calibration experiment in the airport digital twin simulation platform, which uses the number of intrusions into restricted areas, the minimum safe distance compliance rate, and the aircraft stand collaborative operation completion rate as evaluation indicators. The operation reward weight can be set to 0.15, where 0.15 comes from the same weight calibration experiment, and the sum of 0.15 and the safety reward weight is 1.

[0069] The weights for safety rewards and operational rewards were determined through multiple weight calibration experiments on a digital twin simulation platform. Specifically, under the same number of devices, the same task points, and the same dynamic interference injection rules, multiple weight combinations were tested. The number of intrusions into restricted areas, the minimum safe distance compliance rate, the aircraft stand collaborative operation completion rate, and the equipment operation resource utilization rate were statistically analyzed. Weight combinations that met airport compliance requirements for safety indicators and did not significantly decrease operational indicators were selected. For example, when α=0.75 and β=0.25, the task completion rate was high, but the safe distance compliance rate was insufficient; when α=0.90 and β=0.10, safety was high, but waiting time increased; when α=0.85 and β=0.15, both the safe distance compliance rate and the task completion rate were within a relatively stable range, therefore this weight combination was adopted.

[0070] The learning rate of the local Actor policy network for the ground-based unmanned vehicle can be set to 0.001, where 0.001 is derived from the low-speed motion characteristics of the ground-based unmanned vehicle and historical training convergence test data. The learning rate of the local Actor policy network for the aerial inspection drone can be set to 0.003, where 0.003 is derived from the state change frequency of the aerial inspection drone and historical training convergence test data. The learning rate of the global Critic evaluation network can be set to 0.002, where 0.002 is derived from the global value function training stability test data. The learning rates of the local Actor policy network and the global Critic evaluation network are determined through historical training convergence tests. Specifically, under the same simulation scenario, the same reward function, and the same experience replay pool settings, multiple sets of learning rates are tested for the ground-based unmanned vehicle, the aerial inspection drone, and the global Critic evaluation network, respectively, and the training oscillation amplitude, the descent speed of violation candidate actions, and the stability of the task completion rate are compared. For example, ground-based unmanned vehicles change their movements slowly, and the training curve is more stable when the learning rate is 0.001; aerial inspection drones change their states more frequently, and the movements are adjusted faster and without significant oscillations when the learning rate is 0.003; the global Critic evaluation network shows less fluctuation in value evaluation results in consecutive training rounds when the learning rate is 0.002, so the above learning rate combination is used.

[0071] In this embodiment, the two-layer weighted reward function can be expressed as:

[0072] in, Indicates the first Total reward for each decision-making cycle; Indicates the decision-making cycle number; Indicates the weight of security rewards; Indicates the first Security rewards for each decision-making cycle; Indicates the weight of the assignment reward; Indicates the first The task reward items for each decision cycle; in one embodiment, Pick , Pick ,and .

[0073] In some optional embodiments, regarding S3, assuming an unmanned baggage transfer vehicle is in the service lane, the initial candidate action output by the local Actor policy network is to continue moving forward and pass through the buffer warning area at a speed of 3 m / s. The system calls the compliance environment state matrix generated in S2 to identify the speed limit value of this area as 2 m / s, which comes from the low-speed passage test data of the buffer warning area and the airport slow passage rules. Since the initial candidate action exceeds the speed limit value of this area, the system marks it as a buffer area speed limit violation candidate action and generates a low-speed safety candidate action, allowing the unmanned baggage transfer vehicle to pass at a speed not exceeding 2 m / s. This violation candidate action and the corresponding safety candidate action are written into the experience replay pool to reduce the probability of the local Actor policy network outputting speeding actions in similar scenarios during subsequent training.

[0074] In some alternative embodiments, it is assumed that the initial candidate action of an aerial inspection UAV points to the 0m to 4m altitude range of the takeoff and landing airspace. The 0m to 4m range is derived from airport low-altitude safety control rules and tower control requirements for takeoff and landing related airspace. The system determines that the initial candidate action violates the restricted area constraint, marks it as a prohibited area violation candidate action, and generates a safe candidate action that maintains the current safe altitude or rises to a preset avoidance altitude of 8m. The 8m range is derived from the aerial inspection UAV's low-altitude operation test, sensor imaging altitude requirements, and airport takeoff and landing airspace avoidance rules. If there are aircraft taxiing paths in the vicinity, the system also writes the tower control instruction vector into the training samples, enabling the global Critic evaluation network to further reduce the value evaluation of such actions in subsequent training.

[0075] like Figure 5 As shown, in one specific embodiment, S4 includes: S41. Read safety candidate actions and tower dispatch data; S42. Perform scheduling status matching and authorization matching on security candidate actions; S43. Generate execution control instructions or safety alternative actions based on the matching results; S44. Perform collaborative operations with unmanned equipment; S45. Store decision logs, running trajectories, lockout records, and reward / punishment data; S46. Trigger emergency dynamic replanning under sudden working conditions.

[0076] In this embodiment, it should be noted that for S41, the safety candidate action originates from S3, and the tower dispatch data originates from S1 and is processed by S2 to form an authorization status flag. When reading a safety candidate action, the system simultaneously reads the corresponding unmanned equipment identifier, unmanned equipment type, predicted path, operation type, current time, and area safety level. The tower dispatch data includes high-risk area operation authorization instructions, temporary control areas, temporary control time windows, aircraft priority passage instructions, and unmanned equipment operation suspension instructions. If the safety candidate action does not involve a high-risk area, the system still needs to verify whether there are unmanned equipment operation suspension instructions, temporary control area conflicts, temporary control time window conflicts, or aircraft priority passage conflicts. If the safety candidate action involves a high-risk area, it must proceed to S42 for authorization matching.

[0077] For S42, scheduling status matching includes pause operation instruction matching, temporary control area matching, temporary control time window matching, and aircraft priority passage conflict matching. The authorization matching process includes equipment identifier matching, authorization area matching, authorization time window matching, authorized operation type matching, and authorization status matching. Equipment identifier matching confirms whether the unmanned equipment performing the safety candidate action is an authorized object. Authorization area matching confirms whether the predicted path of the safety candidate action is within the authorized area. Authorization time window matching confirms whether the current time is within the authorization effective time. Authorization operation type matching confirms that the current operation belongs to the authorized permitted operation type. Authorization status matching confirms that the authorization instruction is still valid. The authorization time window can be set to be consistent with the electronic authorization record issued by the tower dispatch platform; its source is the authorization instruction from the tower dispatch platform, and it is not generated by the algorithm itself.

[0078] For S43, if the scheduling status match and authorization match are successful, the device execution interlocking layer converts the safety candidate action into an execution control command and sends it to the corresponding unmanned equipment controller. If the scheduling status match or authorization match fails, the device execution interlocking layer does not send the corresponding execution control command, but instead generates a safety alternative action. Ground unmanned vehicles can enter a parking waiting state, low-speed detour, or return-to-home state. Aerial inspection drones can enter a hovering, maintaining a safe altitude, horizontal detour, or return-to-home state. If the tower dispatch platform issues a pause command for unmanned equipment, the corresponding unmanned equipment enters a waiting state, which continues until the tower dispatch platform lifts the pause command or the system generates a new compliant operation task.

[0079] For S44, unmanned equipment operates collaboratively according to executed control commands. Ground-based unmanned vehicles perform tasks such as guidance, cleaning, refueling, or baggage transfer within service lanes, aircraft stand operation areas, or peripheral pavements. Aerial inspection drones perform inspection, security, or airport status verification tasks in non-takeoff and landing airspace. During execution, unmanned equipment continuously reports its position, speed, task status, and fault status to the airport edge server. If the reported data is inconsistent with the executed control commands, the system writes the discrepancy to the log and, if necessary, triggers the equipment to execute a conservative safety action generated by the lockout layer. Conservative safety actions include ground-based unmanned vehicles slowing down and stopping, or aerial inspection drones hovering and returning to base.

[0080] For S45, the system stores decision logs, operational trajectories, interlock records, and reward / penalty data for each decision. Decision logs include the airport's global status, local observations, original candidate actions, safe candidate actions, tower authorization instructions, authorization matching results, and final execution control instructions. Operational trajectories include the unmanned equipment's position, speed, heading, altitude, and mission status. Interlock records include interlock time, interlocked unmanned equipment identifier, interlock reason, violation type, safe alternative actions, and tower authorization status. Reward / penalty data includes safety rewards, operational rewards, action interlock penalties, and mission completion rewards. This data is stored according to the unmanned equipment identifier and timestamp, and is used for airport unmanned equipment safety operation logs, model training review, and safety responsibility tracking.

[0081] For S46, contingency scenarios include temporary aircraft taxiing, personnel intrusion, changes to temporary fencing, addition of new construction closure areas, unmanned equipment malfunction and offline operation, issuance of temporary control instructions by the control tower, and effective dynamic obstacles entering the predicted path of unmanned equipment. Upon triggering a contingency scenario, the system writes new data into the airport's global state, regenerates or updates the airport safety constraint set and compliance environment state matrix, and uses the improved MADDPG algorithm to reuse trained network parameters to output new original candidate actions. For a scenario with 12 unmanned devices, the average time for dynamic replanning can be 47ms. This 47ms is derived from statistical results from the airport digital twin simulation platform under conditions of 12 unmanned devices, 1200 training rounds, and one dynamic disturbance injected every 20 rounds. This 47ms is an average time and does not mean that the time for every single replanning operation is less than 50ms.

[0082] In some optional embodiments, regarding S4, suppose an unmanned sweeper receives a safety candidate action that would cause it to enter a high-risk area around the taxiway. The system reads the tower dispatch data and finds that the tower dispatch platform has granted operation authorization for the area to another unmanned guided vehicle within the current time window, but has not granted authorization to this unmanned sweeper. When the system performs device identifier matching, it determines that the authorization does not match, therefore it does not issue an execution control command to the unmanned sweeper and generates a stop waiting action. The stop waiting action is written to the interlock record, and the interlock reason is recorded as unauthorized high-risk area. If the tower dispatch platform subsequently issues a valid authorization command to the unmanned sweeper, the system will re-perform authorization matching in the next decision cycle, and only issue the corresponding execution control command after the matching is successful.

[0083] In some alternative embodiments, assuming an aircraft temporarily changes its taxiing path, the tower dispatch platform issues a temporary control instruction to the system. In S1, the system updates the overall airport status; in S2, it marks the area surrounding the aircraft's predicted path as a higher safety level area; and in S3, it regenerates the original candidate actions for each unmanned vehicle and performs preliminary safety checks. If the original candidate action of an unmanned refueling truck conflicts with the aircraft's predicted path, the system marks it as a violation of the aircraft's priority passage rule and generates a stop-waiting or detour action. In S4, after matching the safety candidate actions with scheduling status and authorization, the waiting action is issued to the unmanned refueling truck controller, and the aircraft's temporary taxiing path, the unmanned refueling truck's interlock record, and the subsequent detour trajectory are written to the log.

[0084] like Figures 6-8 As shown, in one specific embodiment, this embodiment also relates to a safety constraint collaborative planning system for multiple types of unmanned equipment at airports. The system includes an airport area perception layer, a safety constraint preprocessing layer, a safety constraint improved MADDPG algorithm decision layer, an equipment execution interlocking layer, and a tower tracing interaction layer.

[0085] The airport area perception layer is used to execute S1. It connects to the airport apron millimeter-wave radar, vehicle-mounted lidar, BeiDou positioning terminals, aerial inspection UAV positioning modules, airport flight area electronic maps, the airport operations management system, and the tower dispatch platform. The output of the airport area perception layer is the overall airport status. To ensure that the output data can be used by the security constraint preprocessing layer, the airport area perception layer writes the data source, timestamp, coordinate system, effective time window, and data reliability status for each data item. The data reliability status can include valid, short-term prediction, abnormal, and pending confirmation. The data reliability status is marked based on sensor historical error statistics, equipment positioning drift tests, and multi-source data consistency verification results.

[0086] The safety constraint preprocessing layer is used to execute S2. It reads the airport's global status, generates a set of airport safety constraints, filters abnormal data, marks regional security levels and authorization statuses, and outputs a compliance environment status matrix. The preprocessing layer includes a built-in airport safety rule configuration table, which includes restricted area boundaries, buffer warning area boundaries, collision avoidance distance thresholds, speed limit thresholds, equipment action boundary thresholds, and tower authorization constraints. The source of each threshold is recorded. If a threshold comes from equipment manufacturer technical parameters, the equipment model and manufacturer parameter version are recorded. If a threshold comes from airport field testing, the test time, test scenario, and calibration results are recorded. If a threshold comes from the tower dispatch platform, the corresponding dispatch instruction number and effective time window are recorded.

[0087] The safety-constrained improved MADDPG algorithm's decision layer is used to execute S3. The decision layer comprises a local Actor policy network, a global Critic evaluation network, an experience replay pool, a two-layer weighted reward function processing unit, and a tower dispatch instruction vector encoding unit. The local Actor policy network, deployed in the unmanned equipment's native controller or airport edge server, outputs raw candidate actions. The global Critic evaluation network, deployed in the airport edge server, evaluates the value of actions corresponding to the airport's global state, joint actions, and tower dispatch instruction vectors during the training phase. The experience replay pool stores normal samples and violation candidate action samples. The tower dispatch instruction vector encoding unit transforms temporary control areas, passage priorities, authorization status, and delayed operation instructions into algorithm input fields. Through this structure, tower dispatch data is used not only for operation recording but also for algorithm training and action evaluation.

[0088] The equipment execution interlocking layer is used to perform scheduling status matching, authorization matching, and control command release in S4. It receives safety candidate actions, tower scheduling data, and airport safety constraint sets to determine whether to allow the issuance of execution control commands. The equipment execution interlocking layer has two functions: execution release verification and tower authorization interlocking. Execution release verification reads the safety candidate actions output from S3 and their pre-existing safety judgment results, and verifies whether new temporary control measures, aircraft path changes, or unmanned equipment suspension commands have occurred in the current airport global status before issuance. Tower authorization interlocking prevents high-risk area operations without valid authorization from entering the execution phase. The equipment execution interlocking layer performs machine matching based on the tower authorization command field and the safety candidate action field, without relying on manual confirmation step-by-step.

[0089] The tower traceability interaction layer is used to perform log storage and ledger generation in S4. This layer is a data interaction module for exchanging dispatch instructions, authorization results, and operational logs with the tower dispatch platform or airport operations management system. It receives temporary control instructions, high-risk area authorization instructions, and suspension of operations instructions from the tower dispatch platform, and transmits unmanned equipment decision logs, operational trajectories, interlock records, and reward / penalty data back to the tower dispatch platform or airport operations management system. The tower traceability interaction layer organizes data according to unmanned equipment identifiers and timestamps, enabling airport management personnel to trace the original candidate actions, safe candidate actions, tower authorization status, and final executed actions of a specific unmanned equipment within a certain time period. This data can also be used for subsequent model training and review; the review process does not change the stored execution action records. The equipment execution interlock layer and the tower traceability interaction layer together correspond to the matching, instruction release, and data storage processes in S4.

[0090] In one specific implementation, the training and simulation examples are described. The simulation examples are built on an airport flight area digital twin simulation platform. The simulation hardware platform includes an Intel i9-13900HX processor and an RTX 4090 deep learning graphics card. The above hardware configuration is derived from the actual test platform of the simulation examples. The simulation software platform includes Python 3.10, PyTorch 1.10, and Matlab. The above software versions are derived from the training environment configuration records of the simulation examples. The simulation field size is 420m × 360m, which is derived from the digital map modeling data of flight areas of small and medium-sized airports. The simulation field includes a runway restricted area, a taxiway 3m buffer warning area, an aircraft stand operation area, and an outer service lane operation area. The 3m is derived from airport on-site safety operation rules, taxiway boundary calibration data, and historical manual operation avoidance distance calibration data.

[0091] The simulation network comprises 12 unmanned intelligent agents, representing the scale of multi-equipment collaborative operations at a small to medium-sized airport as set in the simulation example. These 12 agents include 8 ground-based unmanned vehicles and 4 aerial inspection drones. The 8 ground-based unmanned vehicles consist of 2 unmanned guidance vehicles, 2 unmanned cleaning vehicles, 2 unmanned baggage transfer vehicles, and 2 unmanned refueling vehicles; this number is derived from the equipment configuration example for a single aircraft stand coordinated operation task. The 4 aerial inspection drones are configured based on the low-altitude inspection coverage area and the simulated field area. The maximum speed of the ground-based unmanned vehicles is 6 m / s, derived from the manufacturer's technical parameters and airport closed road test data. The maximum angular velocity of the ground-based unmanned vehicles is 0.3 rad / s, derived from the chassis controller parameters and airport low-speed turning test data. The maximum flight speed of the aerial inspection drones is 10 m / s, derived from the manufacturer's technical parameters and airport low-altitude inspection test data. The operating altitude range of the aerial inspection drone is 5m to 18m. The range of 5m to 18m is derived from the airport low-altitude inspection operation plan, sensor imaging distance test and take-off and landing airspace avoidance requirements.

[0092] The simulation implementation set up three sets of algorithm comparisons. The first set was the improved A-satellite path planning algorithm, the second set was the original MADDPG algorithm, and the third set was the safety-constrained improved MADDPG algorithm of this embodiment. All three algorithms were run under the same field map, the same number of unmanned devices, the same initial task points, and the same dynamic interference injection rules. The discount factor was set to 0.92, derived from the balance calibration experiment of short-term obstacle avoidance benefits and long-term task completion benefits in the training samples. The soft update coefficient was set to 0.005, derived from the target network training stability test data. The maximum number of training rounds was set to 1200, derived from the historical training round statistics required for the original MADDPG algorithm and the safety-constrained improved MADDPG algorithm to reach convergence or enter a stable training phase under the same simulation environment. The simulation duration per round was set to 50 seconds, derived from the single-round task window configuration of the aircraft stand linkage operation simulation task. Dynamic interference was injected once every 20 rounds, derived from the interference injection frequency calibration data used in the airport digital twin simulation platform to cover normal operations and emergency situations.

[0093] The discount factor, soft update coefficient, and maximum training epochs were determined through simulation training stability testing. Specifically, different discount factors, soft update coefficients, and training epochs were tested under fixed device counts and task scenarios, recording short-term obstacle avoidance response, long-term task completion rate, training curve fluctuations, and target network stability. For example, a discount factor below 0.90 indicates a bias towards short-term obstacle avoidance, resulting in insufficient progress on some tasks; a discount factor above 0.95 indicates slower training convergence, therefore 0.92 was selected. A soft update coefficient above 0.01 increases target network fluctuations, while a coefficient below 0.001 decreases convergence speed, therefore 0.005 was selected. After more than 1200 training epochs in the same scenario, the main metrics showed little change, therefore the maximum training epochs were set to 1200 epochs.

[0094] Simulation evaluation metrics include the number of intrusions into restricted areas, the minimum safe distance compliance rate, the number of training convergence iterations, the average time for dynamic replanning, the completion rate of coordinated operations between equipment and workstations, and the utilization rate of equipment resources. For the improved A* path planning algorithm, the number of training convergence iterations is not evaluated because it is not a neural network training algorithm and therefore does not have the same meaning as the original MADDPG algorithm and the safety-constrained improved MADDPG algorithm. The utilization rate of equipment resources is obtained by statistically analyzing the cumulative time the unmanned equipment is in an effective operating state or effective passage state, and then dividing this by the total runtime of the unmanned equipment. The determination of effective operating state and effective passage state is derived from the unmanned equipment's task status field, executed control command records, and running trajectory records.

[0095] In this embodiment, the equipment operation resource utilization rate can be expressed as:

[0096] in, Indicates the utilization rate of equipment operating resources; This indicates the cumulative duration, in seconds, of the unmanned equipment being in an effective operational or passageway state. This indicates the cumulative duration, in seconds, of the unmanned equipment being in a waiting, idle, or no-task state. This indicates the percentage conversion factor. and All data originates from the unmanned equipment's task status field, execution control command records, and operation trajectory records.

[0097] Under the above simulation conditions, the statistical results are as follows. The improved A-satellite path planning algorithm intruded into restricted areas 27 times in 1200 simulation rounds. The 1200 rounds were determined by the maximum training rounds or an equivalent number of simulation rounds, and the 27 instances were derived from simulation platform trajectory statistics. Its minimum safe distance compliance rate was 71.4%, derived from the simulation platform's periodic statistics of distances between unmanned equipment and aircraft, and between unmanned equipment. Its average dynamic replanning time was 2140ms, derived from the time taken to recalculate the global path for all unmanned equipment after dynamic interference was triggered. Its aircraft station collaborative operation completion rate was 76.5%, derived from simulation task completion records. Its equipment operation resource utilization rate was 62.4%, derived from the cumulative duration of unmanned equipment's effective operation status, effective passage status, and waiting status.

[0098] The native MADDPG algorithm intruded into restricted areas 19 times in 1200 simulation rounds, with these 19 instances derived from simulation platform trajectory statistics. Its minimum safe distance compliance rate was 78.2%, derived from cycle-by-cycle minimum distance statistics. Its training convergence iterations totaled 912 rounds, derived from iteration records of 50 consecutive rounds meeting the training stability criteria. These 50 consecutive rounds were derived from training curve stability calibration experiments. Its average dynamic replanning time was 320ms, derived from statistics on the time required to re-output actions after dynamic interference triggering. Its machine-station collaborative operation completion rate was 83.1%, derived from simulation task completion records. Its equipment operation resource utilization rate was 73.6%, derived from statistics of the cumulative duration of unmanned equipment's effective operating state, effective passage state, and waiting state.

[0099] In this embodiment, the improved MADDPG algorithm for safety constraints recorded zero intrusions into restricted areas across 1200 simulation rounds. This zero-intrusion count is derived from the operational trajectory statistics under the conditions of this simulation embodiment and does not imply that no intrusions into restricted areas occur at all airports, on all equipment, or under all operating conditions. Its minimum safe distance compliance rate is 98.6%, derived from cycle-by-cycle minimum distance statistics. The training convergence iteration rounds are 528, derived from 50 consecutive rounds of training records that satisfy the following conditions: zero intrusions into restricted areas, a collision avoidance compliance rate of no less than 98%, and a fluctuation in the aircraft stand collaborative operation completion rate of no more than 2%. The 98% is derived from the safety compliance threshold calibration experiment of the airport digital twin simulation platform, and the 2% is derived from the task completion rate stability judgment experiment. Its average dynamic replanning time is 47ms, derived from the time taken for the airport edge server to re-output the original candidate action and complete the pre-safety judgment after dynamic interference is triggered, excluding sensor sampling delay, wireless communication queuing delay, and unmanned equipment controller execution delay. Its machine-station collaborative operation completion rate was 97.2%, which was derived from the simulation task completion record. Its equipment operation resource utilization rate was 91.3%, which was derived from the cumulative duration statistics of the effective operation status, effective passage status, and waiting status of the unmanned equipment.

[0100] In one specific implementation, a multi-device scale-up example is described. Under the same hardware platform and algorithm configuration, the system was tested with 12, 20, and 35 unmanned intelligent agents. The 12, 20, and 35 agents represent simulation scale settings for single-aircraft station operations, multi-aircraft station parallel operations, and high-density unmanned equipment collaborative operations at small and medium-sized airports. In the 12-agent scenario, the average dynamic replanning time was 47ms, the station collaborative operation completion rate was 97.2%, and the equipment operation resource utilization rate was 91.3%. These figures are from statistics of the 12-agent simulation example. In the 20-agent scenario, the average dynamic replanning time was 68ms, the station collaborative operation completion rate was 95.8%, and the equipment operation resource utilization rate was 89.6%. These figures are from statistics of the 20-agent stress test. In the 35-agent scenario, the average dynamic replanning time was 118ms, the station collaborative operation completion rate was 92.4%, and the equipment operation resource utilization rate was 86.7%. These figures are from statistics of the 35-agent stress test. Therefore, under the aforementioned hardware platform and simulation parameters, a single airport edge server node can support collaborative planning computation for no more than 35 unmanned intelligent agents. It should be noted that in the scenario of 35 unmanned intelligent agents, an average dynamic replanning time of less than 50ms is not recommended; instead, a statistical result of 118ms obtained from stress testing is used for this scale.

[0101] In one specific implementation, the parameter acquisition method is described. The safety distance threshold, speed threshold, angular velocity threshold, flight altitude threshold, buffer zone width, decision cycle, training hyperparameters, and reward weight are not arbitrarily set values, but are obtained based on airport operation rules, equipment manufacturer parameters, airport on-site testing, digital twin simulation experiments, and a limited number of historical calibration data. Specifically, for the safety distance threshold, the system first reads the airport operation safety management rules and equipment operation safety requirements, and then corrects them by combining the unmanned equipment braking distance, positioning error, heading control error, and sensor recognition error. For the equipment action boundary threshold, the system first reads the equipment manufacturer's technical parameters, and then calibrates them through low-speed operation tests, braking tests, steering tests, hovering tests, and return-to-base tests in the airport closed area. For the training hyperparameters, the system compares training stability, convergence rounds, number of intrusions into restricted areas, and task completion rate using multiple sets of historical training data, selecting a parameter combination that balances safety constraints and operational efficiency. These parameters are written into the airport safety rule configuration table, and the source of acquisition, test scenario, and updated version are recorded for each parameter.

[0102] In one specific implementation, the correspondence between terms and data flows is explained. The airport global state is generated by S1 and serves as input to S2. The airport safety constraint set and compliance environment state matrix are generated by S2 and serve as input to S3. The original candidate actions are generated by the local Actor policy network in S3 and serve as input to the pre-safety judgment. The safety candidate actions are output from the pre-safety judgment and safety alternative action generation process in S3 and serve as input to the scheduling state matching and authorization matching in S4. The execution control command is generated by S4 after the scheduling state matching and authorization matching are passed and serves as input to the unmanned equipment controller. The decision log, operation trajectory, interlocking record, and reward and punishment data are stored by S4 and serve as the data source for subsequent model training review, airport unmanned equipment safety operation ledger, and safety responsibility tracking. Through the above data flow, the results of the previous step are all called by the next step, forming a closed loop relationship between the method steps.

[0103] In one specific implementation, the system operation flow supported by this method is described. The airport area perception layer first collects airport area environmental data, aircraft operation data, unmanned equipment status data, and tower scheduling data, performs time synchronization and coordinate unification, and constructs the airport's global state. The safety constraint preprocessing layer reads the airport's global state, generates an airport safety constraint set, and obtains a compliant environment state matrix through abnormal data filtering, regional security level marking, and authorization status marking. The safety constraint improved MADDPG algorithm decision layer reads the compliant environment state matrix, generates original candidate actions, performs pre-safety judgment on the original candidate actions, and outputs safe candidate actions. The equipment execution interlocking layer reads safe candidate actions and tower scheduling data, completes scheduling status matching and authorization matching, and issues execution control commands after successful matching. The tower traceability interaction layer stores decision logs, operation trajectories, interlocking records, and reward / penalty data. The above layers can be deployed between the airport edge server, the unmanned equipment's local controller, and the tower scheduling platform; the specific deployment location can be adjusted according to the airport's existing hardware conditions.

[0104] In one specific implementation, the emergency handling of this method is described. When the airport area perception layer detects a temporary aircraft taxiing in, personnel intrusion, changes to temporary fencing, addition of a construction closure area, unmanned equipment malfunction and offline status, or a temporary control command issued by the control tower, the system does not retrain the model from scratch. Instead, it updates the airport global state and compliance environment state matrix and reuses the trained local Actor policy network and global Critic evaluation network for action updates. If the malfunctioning unmanned equipment is located in a regular operation area, the system marks its location as a valid dynamic obstacle and replans the operation paths of the remaining unmanned equipment. If the malfunctioning unmanned equipment is located in a buffer warning area or a high-risk area, the system increases the safety level of that area and prioritizes generating deceleration, stopping, hovering, or detour actions for surrounding unmanned equipment. If a temporary aircraft taxiing in causes an area to temporarily become a high-priority passage area, the system writes the aircraft path into the aircraft priority passage constraint and blocks the original candidate actions of unmanned equipment that conflict with that path.

[0105] In one specific implementation, the technical effects of this method are objectively described. Under the simulation examples described in this specification, the safety constraint improved MADDPG algorithm, through airport safety constraint sets, pre-emptive safety discrimination, tower scheduling instruction vectors, a two-layer weighted reward function, and a device execution lockout layer, ensures that the original candidate actions of unmanned equipment are subject to the joint constraints of restricted areas, collision avoidance distances, aircraft priority passage, tower authorization, and device action boundaries before execution. Compared with the native MADDPG algorithm, in the simulation scenario of 12 unmanned devices, this method has lower training convergence iteration rounds and average dynamic replanning time, and higher minimum safety distance achievement rate, stand collaborative operation completion rate, and equipment operation resource utilization rate. The above effects are all based on specific simulation environments, specific equipment scales, and specific parameter conditions, and do not mean that completely identical numerical results can be obtained under any airport, any equipment scale, and any communication conditions. This description is consistent with the data in the examples and avoids extending the simulation results to conclusions that are not subject to any conditions.

[0106] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A collaborative planning method for safety constraints of multiple types of unmanned equipment at airports, characterized in that, include: Collect airport area environmental data, aircraft operation data, unmanned equipment status data and tower dispatch data, and after time synchronization and coordinate unification, construct the airport global status; Based on the airport's global state, an airport security constraint set is generated, and the airport's global state is subjected to abnormal data filtering, regional security level marking, and authorization status marking to obtain a compliance environment state matrix. The compliance environment state matrix is ​​input into the safety constraint improved MADDPG algorithm to generate the original candidate actions of each unmanned device. Based on the airport safety constraint set, the original candidate actions are pre-safety judged to obtain safe candidate actions. The safety candidate actions are matched with the tower scheduling data for scheduling status and authorization. After a successful match, the execution control command is issued and the corresponding decision log, running trajectory, interlocking record and reward / penalty data are stored.

2. The safety constraint collaborative planning method for multiple types of unmanned equipment at airports according to claim 1, characterized in that, The airport area environmental data includes the boundaries of the runway area, taxiway area, aircraft stand area, service lane area, temporary enclosure area, construction closure area, fixed structure location, and dynamic obstacle location. The aircraft operation data includes aircraft position, aircraft heading, aircraft speed, current aircraft operating phase, and aircraft passage priority; the unmanned equipment status data includes unmanned equipment identifier, unmanned equipment type, unmanned equipment position, unmanned equipment heading, unmanned equipment speed, unmanned equipment remaining battery power or remaining fuel, unmanned equipment mission status, unmanned equipment executable action boundaries, and unmanned equipment fault status. The tower dispatch data includes temporary control areas, temporary control time windows, aircraft priority passage instructions, unmanned equipment operation priority instructions, high-risk area operation authorization instructions, and unmanned equipment operation suspension instructions.

3. The safety constraint collaborative planning method for multiple types of unmanned equipment at an airport according to claim 1, characterized in that, The airport safety constraint set includes restricted area constraints, buffer zone speed limit constraints, collision avoidance distance constraints, aircraft priority passage constraints, tower authorization constraints, and equipment action boundary constraints. The restricted area constraint is used to restrict unmanned equipment from entering the core area of ​​the runway, the take-off and landing airspace, the core circle of aircraft parking positions, and the restricted areas temporarily designated by the tower dispatch platform; the collision avoidance distance constraint is used to limit the minimum safe distance between ground unmanned operation vehicles and aircraft, between ground unmanned operation vehicles, between aerial inspection drones and aircraft fuselages, and between aerial inspection drones. The tower authorization constraints are used to restrict unmanned equipment without valid authorization from entering the area around the runway, taxiway, takeoff and landing airspace, or high-risk areas designated by the tower dispatch platform.

4. The safety constraint collaborative planning method for multiple types of unmanned equipment at airports according to claim 1, characterized in that, Filtering abnormal data for the overall airport status includes: The location data of unmanned equipment is checked for speed constraints. When the estimated speed corresponding to the position change in adjacent decision cycles exceeds the sum of the maximum allowable speed and the speed anomaly judgment tolerance value for this type of unmanned equipment, the current positioning point is marked as an abnormal positioning point, and the previous valid positioning point is combined with the current speed estimate for short-term interpolation update. The dynamic obstacle data is continuously verified. When a dynamic obstacle appears only in a single decision cycle and is not confirmed by other data sources in two or more consecutive decision cycles, it is marked as a transient interference target. When a dynamic obstacle is detected in two or more consecutive decision cycles or confirmed by tower dispatch data, aircraft operation data or local perception data of unmanned equipment, it is marked as a valid dynamic obstacle.

5. The safety constraint collaborative planning method for multiple types of unmanned equipment at an airport according to claim 1, characterized in that, The compliance environment status matrix includes: area coordinates, area type, security level, speed limit, traffic direction, authorization status, temporary control status, and dynamic obstacle status; When the same area corresponds to multiple security levels, the compliance environment state matrix is ​​generated according to the higher security level. When the tower dispatch data sets a certain area as a temporary control area, the area is written into the compliance environment status matrix as a prohibited or restricted area within the corresponding temporary control time window.

6. The safety constraint collaborative planning method for multiple types of unmanned equipment at an airport according to claim 1, characterized in that, The safety constraint improved MADDPG algorithm includes multiple unmanned equipment agents. Each unmanned equipment agent is configured with a local Actor policy network. During the training phase, it receives the airport global state, the joint actions of all unmanned equipment agents and the tower scheduling instruction vector through the global Critic evaluation network. During the execution phase, each unmanned equipment agent outputs the original candidate actions based on its local observations. Among them, the original candidate actions for ground unmanned vehicles include linear velocity and angular velocity, while the original candidate actions for aerial inspection drones include flight speed, heading angle adjustment, and target flight altitude.

7. The safety constraint collaborative planning method for multiple types of unmanned equipment at an airport according to claim 1, characterized in that, Preliminary security assessment of the original candidate actions based on the airport security constraint set includes: Predict the next cycle position of the unmanned equipment after it executes the original candidate action, and determine whether the next cycle position or the movement path formed from the current position to the next cycle position enters a restricted area, is less than the corresponding collision avoidance distance threshold, violates the aircraft priority passage rules, involves the tower authorization requirements of high-risk areas, and exceeds the boundary of the equipment action. When any judgment result is non-compliant, the original candidate action is marked as a non-compliant candidate action, and a safe alternative action is generated as the safe candidate action.

8. The safety constraint collaborative planning method for multiple types of unmanned equipment at an airport according to claim 7, characterized in that, The safety alternative action is generated based on the type of unmanned equipment and the type of violation. When the unmanned equipment is a ground-based unmanned operating vehicle, the safety alternative actions include deceleration actions generated according to the maximum permissible braking deceleration, stopping and waiting actions, low-speed circling actions, or returning actions; when the unmanned equipment is an aerial inspection drone, the safety alternative actions include hovering actions, maintaining a safe altitude actions, ascending to a preset avoidance altitude actions, horizontal circling actions, or returning actions. The marked violation candidate actions, their corresponding safe alternative actions, and violation types are written into the experience replay pool to reduce the output probability of similar violation candidate actions in subsequent training.

9. The safety constraint collaborative planning method for multiple types of unmanned equipment at an airport according to claim 1, characterized in that, The safety constraint improved MADDPG algorithm is trained using a two-layer weighted reward function, which includes a safety reward item and a job reward item. The safety rewards include positive rewards for avoiding aircraft and effective dynamic obstacles, positive rewards for maintaining compliant collision avoidance distances, penalties for violations of restricted areas, penalties for violations of aircraft priority passage, and penalties for action lockouts. The task reward items include task path advancement rewards, task timing rewards, task completion rewards, and collaborative avoidance rewards; When ground-based unmanned vehicles and aerial inspection drones train their local Actor policy networks, they use different local Actor policy network learning rates but the same global Critic evaluation network learning rate.

Images