Design and deployment method for multi-agent crowd-sourcing decision-making system of intelligent airport apron

Abstract

The invention discloses a design and deployment method for a multi-agent crowd-sourcing decision-making system of an intelligent airport apron. The method comprises the following steps: constructing an edge sensing layer composed of a vehicle state sensing unit and a high-precision positioning unit; deploying a plurality of edge computing nodes in a key operation area on an airport apron in a distributed manner, and docking the edge computing nodes with the edge sensing layer and an airport collaborative decision-making system; a materialized crowd-sourcing decision-making device is constructed in the edge computing node; and constructing a visual man-machine interface connected with the interior of the edge computing node, and the like. The method has the advantages that a valuable reference direction is provided for solving the risk of a flight delay system, improving the satisfaction degree of passengers, building a smart airport and the like, and the method is a flow and management innovation; application innovation is provided for actual production and operation of civil aviation, and a new landing instance is established for fusion of AI and digital driving decision technologies; the achievement can be popularized in the national airport ground service vehicle group resource allocation demand occasion, and possibility is provided for reducing resource consumption and personnel cost under normal operation.

Description

Technical Field

[0001] This invention belongs to the field of smart airport construction technology, specifically relating to a design and deployment method for a smart apron multi-agent swarm intelligence decision-making system based on edge computing architecture and reinforcement learning algorithm, which is particularly suitable for the intelligent scheduling and management of ground service vehicles in large airports. Background Technology

[0002] Currently, most domestic airports still rely on static scheduling or even manual scheduling, which is labor-intensive and cannot achieve real-time configuration and scheduling. During flight scheduling, there is an imbalance in the allocation of aircraft, crew members, and ground service resources. Some flight resources are idle, while others face resource shortages, leading to a decline in overall operational efficiency. Airport ground services involve various vehicles, such as aircraft ground power vehicles, aircraft ground air supply units, aircraft towing equipment, and shuttle buses. These service vehicles have different functions, making scheduling difficult. Furthermore, the information system development of some airports is lagging behind, preventing real-time data sharing and analysis. This results in scheduling decisions lacking a scientific basis, impacting overall operational efficiency. Summary of the Invention

[0003] To address the aforementioned problems, the present invention aims to provide a design and deployment method for a smart apron multi-agent swarm intelligence decision-making system.

[0004] To achieve the above objectives, the design and deployment method of the intelligent apron multi-agent swarm intelligence decision-making system provided by the present invention includes the following steps performed in sequence:

[0005] 1) Construct an edge perception layer consisting of vehicle state perception units and high-precision positioning units to collect real-time location and key operational status data of service vehicles;

[0006] 2) Distribute multiple edge computing nodes in key operational areas on the apron and interface with the aforementioned edge perception layer and airport collaborative decision-making system to receive location and key operational status data from the edge perception layer and structured flight support data from the airport collaborative decision-making system, so as to achieve multi-source data fusion and local scheduling decisions.

[0007] 3) Construct a physical crowd intelligence decision-making device within the edge computing node, and use the crowd intelligence decision-making device to fuse multi-source data received from the aforementioned edge computing node, generate the final scheduling instructions in real time, and issue them to the corresponding service vehicles.

[0008] 4) Construct a visual human-machine interface connected to the edge computing node to display real-time data pushed by the edge computing node, so as to realize real-time monitoring of the status of the entire smart apron multi-agent swarm decision-making system.

[0009] In step 1), the vehicle status perception unit adopts an industrial-grade on-board automatic diagnostic system terminal, which is inserted into the standard OBD-II diagnostic interface on each service vehicle. Through this interface, it communicates with the electronic control device of the service vehicle via a CAN bus. The automatic diagnostic system terminal includes a GPS / BeiDou dual-mode positioning module and a 5G / 4G communication module.

[0010] The high-precision positioning unit includes an ultra-wideband positioning base station, an ultra-wideband positioning tag, and a positioning engine server; multiple ultra-wideband positioning base stations are deployed in a honeycomb grid pattern in key operating areas and main roads on the apron; at the same time, an ultra-wideband positioning tag is fixedly installed on the top of each service vehicle; the positioning engine server is integrated into the following edge computing nodes or deployed independently.

[0011] The service vehicles include baggage carts, shuttle buses, refueling trucks, clean water trucks, sewage trucks, and aircraft towing vehicles.

[0012] The vehicle status perception unit is the XYZ-2000 4G / 5G OBD collector; the ultra-wideband positioning base station uses the Decawave DW3000 series chip.

[0013] In step 2), the key operating areas include the baggage sorting area, the flight boarding gate jet bridge area, the aircraft maintenance hangar, the apron main road and taxiway, the fuel supply station, and the cargo terminal.

[0014] The edge computing node uses industrial-grade embedded AI computing equipment, with a built-in 5G communication module and 5G URLLC network interface. It conducts highly reliable data transmission with the service vehicle and cloud server where the edge perception layer is located through the 5G URLLC slicing network. The edge computing node also interfaces with the airport collaborative decision-making system. The interface method is to develop a multi-protocol adapter on the edge computing node.

[0015] The edge computing node uses the NVIDIA Jetson AGX Orin module.

[0016] In step 3), the collective intelligence decision-making device includes an environment state perception interface, an intelligent agent policy executor cluster system, and an instruction distribution and issuing system that are logically interconnected; wherein the intelligent agent policy executor cluster system is physically manifested as multiple parallel processes or threads running on a multi-core CPU or GPU on an edge computing node, each process / thread serving as an intelligent agent policy executor; each intelligent agent policy executor internally encapsulates and runs a pre-trained policy neural network model;

[0017] The input end of the environmental state perception interface is connected to the network interface or memory buffer of the edge computing node to acquire fused global data and flight support structured data in real time. Then, using the internally pre-built state vector assembly logic, according to predefined rules and templates, it extracts, filters, and assembles the corresponding local observation state vector for each agent policy executor from the above data. The local observation state vector includes the real-time location, energy status, current task queue, and related flight support structured data of the service vehicle itself. The output end is connected to each agent policy executor in the agent policy executor cluster system through an internal inter-process communication mechanism to distribute the assembled local observation state vector to these agent policy executors in parallel.

[0018] The intelligent agent policy executor continuously receives its own, assembled local observation state vector from the environmental state perception interface through an internal inter-process communication mechanism. Then, it calls its internal policy neural network model to perform forward propagation calculation on the input local observation state vector to obtain an instruction data packet representing the optimal action. Finally, it sends the instruction data packet to the instruction distribution and delivery system through the internal inter-process communication mechanism.

[0019] The input end of the instruction distribution and issuing system is connected to each intelligent agent policy executor in the intelligent agent policy executor cluster system through an internal inter-process communication mechanism, which is used to collect instruction data packets output by all intelligent agent policy executors; the output end is directly connected to the 5G communication module of the edge computing node, which is used to issue the formatted specific scheduling instructions to the corresponding service vehicles through the 5G URLLC network.

[0020] The policy neural network model is generated through an independent, offline centralized training system, which includes:

[0021] Simulation environment server: running a high-fidelity apron digital twin system;

[0022] Parameter server: Stores and manages the latest central policy network parameters;

[0023] Multiple experience collectors: interact with the simulation environment to collect a large amount of interactive data;

[0024] Model trainer: includes gradient calculation module and parameter update module.

[0025] In step 4), the visual human-machine interface is deployed in the airport's dispatch and command center, duty rooms in key operational areas, and mobile dispatch terminals. It is used to provide airport operators with a centralized monitoring and visual interactive interface for the entire smart apron dispatch system. The interface includes an apron digital map component, a vehicle status table component, and a flight timeline component. These components constitute the core visualization components.

[0026] The design and deployment method of the intelligent apron multi-agent swarm intelligence decision-making system provided by this invention has the following beneficial effects: it can provide valuable reference directions for solving flight delay system risks, improving passenger satisfaction, and building smart airports, representing a process and management innovation; it is an application innovation for actual civil aviation production and operation, and establishes a new practical example for the integration of AI and digital-driven decision-making technologies; the results can be extended to situations requiring resource allocation for airport ground service vehicle fleets nationwide, providing the possibility of reducing resource consumption and personnel costs under normal operation. The visual human-machine interface can provide operators with a unified monitoring view that is real-time, smoothly displayed, and conveniently interactive, greatly improving the transparency of apron scheduling and the efficiency of manual intervention. Attached Figure Description

[0027] Figure 1 Flowchart of the design and deployment method of the intelligent apron multi-agent swarm intelligence decision-making system provided by the present invention;

[0028] Figure 2 This invention provides an information transmission flowchart based on an edge computing architecture.

[0029] Figure 3 This is the iterative curve of vehicle mileage in the simulated airport environment operation of this invention; Detailed Implementation

[0030] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0031] like Figure 1 As shown, the design and deployment method of the intelligent apron multi-agent swarm intelligence decision-making system provided by the present invention includes the following steps executed in sequence:

[0032] 1) Construct an edge perception layer consisting of vehicle state perception units and high-precision positioning units to collect real-time location and key operational status data of service vehicles;

[0033] The edge perception layer is used to acquire basic data for the smart apron multi-agent swarm intelligence decision-making system. Its core task is to achieve comprehensive, high-precision, and high-reliability perception of the status and location of service vehicles on the apron. The service vehicles include baggage carts, shuttle buses, refueling trucks, clean water trucks, sewage trucks, and aircraft towing vehicles.

[0034] The vehicle status perception unit employs an industrial-grade on-board diagnostic (OBD) terminal, specifically the XYZ-2000 4G / 5G OBD collector in this invention. This collector is inserted into the standard OBD-II diagnostic interface on each service vehicle, typically located below the steering wheel in the driver's cab. Through this interface, the system communicates with the service vehicle's Electronic Control Unit (ECU) via CAN (Controller Area Network). The system communicates via a bus. The automatic diagnostic system terminal includes a GPS / BeiDou dual-mode positioning module and a 5G / 4G communication module. By parsing the CAN bus protocol, it periodically reads and encapsulates the following key operating status data of the service vehicles: Location information: The latitude and longitude coordinates of the service vehicles are obtained through the GPS / BeiDou dual-mode positioning module, with a positioning accuracy of approximately 10 meters, used for auxiliary positioning and trajectory backtracking; Power system status information: Including engine speed of fuel vehicles, motor operating status data, instantaneous vehicle speed, and total mileage of electric vehicles; Energy status information: Including fuel quantity (unit: liters) of refueling vehicles, internal water storage / tank capacity of clean water vehicles / sewage vehicles, remaining power (SOC) of electric service vehicles' power batteries, and battery health status data; Then, the automatic diagnostic system terminal uploads the above-mentioned key operating status data in real time to the nearest edge computing node via the 5G / 4G communication module through a 5G URLLC network or airport private network, using MQTT (Message Queuing Telemetry Transport Protocol) or HTTP (Hypertext Transfer Protocol) protocols.

[0035] The high-precision positioning unit includes an ultra-wideband (UWB) positioning base station, an ultra-wideband positioning tag, and a positioning engine server, used to solve the problems of insufficient accuracy and susceptibility to obstruction of GPS / BeiDou in the tarmac environment. Multiple ultra-wideband positioning base stations are deployed in a honeycomb grid pattern in key operating areas and main passages on the tarmac. The deployment spacing should be planned according to the coverage area and required accuracy (usually 50-100 meters) to ensure that there are no signal blind spots in the target area. In this invention, the ultra-wideband positioning base station uses Decawave DW3000 series chips. At the same time, an ultra-wideband positioning tag is fixedly installed on the top of each service vehicle. This system adopts the TDOA (Time Difference of Arrival) positioning algorithm. The ultra-wideband tag on the service vehicle periodically emits pulse signals, which are received by at least three ultra-wideband positioning base stations. Then, the positioning engine server, which is integrated into the edge computing node or deployed independently, calculates the time difference between the pulse signals arriving at different ultra-wideband positioning base stations, and combines this with the precise coordinates of the ultra-wideband positioning base stations to calculate the real-time three-dimensional position coordinates of the service vehicle where the ultra-wideband positioning tag is located with centimeter-level accuracy, and uploads them to the nearest edge computing node.

[0036] 2) Distribute multiple edge computing nodes in key operational areas on the apron and interface with the aforementioned edge perception layer and airport collaborative decision-making system to receive location and key operational status data from the edge perception layer and structured flight support data from the airport collaborative decision-making system, so as to achieve multi-source data fusion and local scheduling decisions.

[0037] Edge computing nodes are the core of this system to achieve low-latency, high-reliability intelligent scheduling. Their core advantage lies in moving data processing and decision-making capabilities from the cloud to the network edge, closer to the data source, thereby significantly reducing communication latency, achieving real-time response, and ensuring the continuity of critical business when network connectivity is unstable. Figure 2 This is a flowchart illustrating the information transmission process of the edge computing architecture in the context of this invention.

[0038] The key operational areas include baggage sorting areas, flight boarding gate bridge areas, aircraft maintenance hangars, apron main roads and taxiways, fuel supply stations, and cargo terminals. In addition, in order to achieve seamless coverage of apron operations, the deployment location of edge computing nodes aims to ensure that their wireless communication range can cover all service vehicles in the area, forming multiple collaborative edge computing units.

[0039] To meet the stable operation requirements of the apron under harsh conditions including wide temperature range, vibration, and electromagnetic interference, the edge computing node uses industrial-grade embedded AI computing equipment with built-in 5G communication modules and 5G URLLC network interfaces. In this invention, the NVIDIA Jetson AGX Orin module is selected, whose powerful built-in GPU computing power is sufficient to run complex multi-agent reinforcement learning models in real time at the edge. The equipment supports stable operation in a wide temperature range of -40℃ to 70℃ and has certain dust and water resistance capabilities (such as reaching IP67 protection level). The 5G communication module supports SA (Standalone) mode with an uplink bandwidth of no less than 100Mbps. It conducts highly reliable data transmission with the service vehicle and cloud server where the edge perception layer is located through the 5G URLLC (Ultra-Reliable Low-Latency Communication) slice network. It is also equipped with redundant gigabit Ethernet interfaces as wired backup or the main means of communication with fixed facilities. The edge computing node receives real-time three-dimensional position coordinates and key operating status data from the edge perception layer through its 5G URLLC network interface, performs timestamp alignment and data fusion, thereby forming the fused global data.

[0040] To achieve linkage between scheduling decisions and flight status, the edge computing nodes also interface with the Airport Collaborative Decision Making (ACDM) system. The interface method involves developing a multi-protocol adapter on the edge computing node. According to the interface specifications provided by the Airport Collaborative Decision Making system, it selects any protocol, including HTTP / REST API, WebSocket, or AMQP (Advanced Message Queuing Protocol), for communication. Through the interface, it periodically receives structured flight support data from the Airport Collaborative Decision Making system, including flight number, aircraft type, scheduled arrival time, scheduled departure time, actual arrival time, allocated gate, and baggage carousel number. This data provides core business input for scheduling decisions.

[0041] 3) Construct a physical crowd intelligence decision-making device within the edge computing node, and use the crowd intelligence decision-making device to fuse multi-source data received from the aforementioned edge computing node to generate the final scheduling instruction in real time;

[0042] Logically, the collective decision-making device is a fixed data processing pipeline dedicated to handling scheduling decisions. It receives multi-source data from edge computing nodes, fuses the multi-source data through a series of internal serial and parallel processing units, and outputs the final scheduling instructions.

[0043] The collective intelligence decision-making device includes an environment state perception interface, an intelligent agent policy executor cluster system, and an instruction distribution and issuing system that are logically interconnected; wherein the intelligent agent policy executor cluster system is physically manifested as multiple parallel processes or threads running on a multi-core CPU or GPU on an edge computing node, each process / thread serving as an intelligent agent policy executor; each intelligent agent policy executor internally encapsulates and runs a pre-trained policy neural network model.

[0044] The input end of the environmental state perception interface is connected to the network interface or memory buffer of the edge computing node to acquire fused global data and flight support structured data in real time. Then, using the internally pre-built state vector assembly logic, according to predefined rules and templates, it extracts, filters, and assembles the corresponding local observation state vector for each agent policy executor from the above data. The local observation state vector includes the real-time location, energy status, current task queue, and related flight support structured data of the service vehicle itself. The output end is connected to each agent policy executor in the agent policy executor cluster system through an internal inter-process communication mechanism to distribute the assembled local observation state vector to these agent policy executors in parallel.

[0045] The intelligent agent policy executor continuously receives its own, assembled local observation state vector from the environmental state perception interface through an internal inter-process communication mechanism. Then, it calls its internal policy neural network model to perform forward propagation calculation on the input local observation state vector. This is a deterministic and high-speed mathematical operation process to obtain an instruction data packet representing the optimal action. Finally, through the internal inter-process communication mechanism, the above instruction data packet is sent to the instruction distribution and issuing system.

[0046] The input end of the instruction distribution and issuing system is connected to each intelligent agent policy executor in the intelligent agent policy executor cluster system through an internal inter-process communication mechanism, used to collect instruction data packets output by all intelligent agent policy executors. The instruction distribution and issuing system internally stores airport map data, traffic rule base, and instruction encoding table, used to perform the following deterministic operations: conflict detection and verification, which is based on the instructions and real-time locations of all service vehicles, and performs rapid logical judgment by querying airport map data and traffic rule base to detect whether there are path or resource conflicts; instruction encoding, which is based on abstract action instructions (such as "path selection 3", "reject task"), and converts them into specific scheduling instructions (such as specific navigation point sequences) that the service vehicle's control system or driver can understand by querying the instruction encoding table; the output end is directly connected to the 5G communication module of the edge computing node, used to issue the formatted specific scheduling instructions to the corresponding service vehicles through the 5G URLLC network.

[0047] The policy neural network model is generated through an independent, offline centralized training system, which includes:

[0048] Simulation environment server: running a high-fidelity apron digital twin system;

[0049] Parameter server: Stores and manages the latest central policy network parameters;

[0050] Multiple experience collectors: interact with the simulation environment to collect a large amount of interactive data;

[0051] Model trainer: includes gradient calculation module and parameter update module;

[0052] a. Training Data Flow and Algorithm: The experience collector feeds data into the model trainer; the gradient calculation module in the model trainer uses the MAPPO algorithm framework, which includes a central policy network to integrate the experience of all agent policy executors during training and output the global policy gradient. The central policy network is a deep neural network whose input is a global state vector composed of the concatenation of the local observation state vectors of all agent policy executors, and whose output is the action distribution parameters for each agent policy executor. During training, the gradient calculation module performs calculations including generalized advantage estimation (GAE) and proximal policy optimization (PPO) pruning of the objective function to obtain the gradient; then the parameter update module updates the parameters of the central policy network based on the above gradient.

[0053] b. Reward Function Design: Throughout the training process, the reward function is the key to guiding the convergence of the policy neural network model. This function comprehensively considers task completion efficiency and resource utilization, and specifically introduces a soft time window penalty mechanism and a game theory equilibrium model as system-level optimization constraints to train an agent policy executor that can achieve global resource optimization and conflict resolution.

[0054] c. Model solidification and deployment: After training is completed, the parameters of the finally converged central policy network are compiled, optimized and solidified into model files that can be loaded on edge computing nodes, and then deployed to the agent policy executor cluster system of each edge computing node.

[0055] 4) Construct a visual human-machine interface (VHMI) connected to the edge computing node to display real-time data pushed by the edge computing node, so as to realize real-time monitoring of the status of the entire smart apron multi-agent swarm decision-making system;

[0056] The visual human-machine interface is deployed in the airport's dispatch and command center, duty rooms in key operational areas, and mobile dispatch terminals. It is used to provide airport operators with a centralized monitoring and visual interactive interface for the entire smart apron dispatch system. Its core objective is to achieve high-frequency data updates, smooth screen rendering, and intuitive information presentation.

[0057] The visual human-machine interface includes an apron digital map component, a vehicle status table component, and a flight timeline component, which together constitute the core visualization components; among them, the apron digital map component is based on Qt Location and Qt Graphics. The View framework is built by importing 2D vector map data of the apron. On this layer, based on real-time data received from edge computing nodes, it dynamically draws and updates the icons, locations, driving trajectories, and task statuses of all service vehicles, such as using different colors to distinguish "in operation," "idle," and "faulty" statuses. This component supports interactive operations such as zooming and panning. The vehicle status table component is implemented using QTableView control and QStandardItemModel data model, clearly displaying detailed status information of all service vehicles in rows and columns, including service vehicle ID, real-time latitude and longitude coordinates, instantaneous speed, energy status (electricity / fuel / water), and current task description, and supports sorting and filtering by field. The flight timeline component is drawn using custom QGraphicsView or a timeline-specific control (such as a self-developed one based on QWidget). It visualizes the planned arrival time, planned departure time, and actual support status of all flights in the future in the form of a Gantt chart, and displays them in association with the gate positions and service vehicle tasks on the 2D vector map.

[0058] To achieve low-latency, high-throughput data synchronization between edge computing nodes and the visual HMI (human-machine interface), ZeroMQ is used as the communication middleware, employing a PUB-SUB (publish-subscribe) model. The edge computing nodes act as publishers, and the visual HMIs as subscribers. This model supports efficient data broadcasting from one publisher to multiple subscribers, making it ideal for monitoring applications. The edge computing nodes encapsulate the merged full-service vehicle status data into Google Protocol Buffers (an efficient serialization format) or JSON format and push it to all connected visual HMIs via ZeroMQ sockets at a fixed frequency of 10Hz. The pushed data includes all the information needed for map rendering and table updates, such as service vehicle IDs, location coordinates, speed, and task information.

[0059] Simulation experiment:

[0060] The simulation experiment of the design and deployment method of the smart apron multi-agent swarm intelligence decision-making system based on edge computing architecture provided by the present invention is as follows: Windows 11 operating system, CPU processor is AMD Ryzen 7 7840H, and display adapter is NVIDIA GeForce RTX 4060 Laptop GPU.

[0061] To verify the effectiveness of the method of the present invention in an airport tarmac environment, big data resources and data from a domestic airport's information system were selected. An tarmac scene map was imported using Python as the platform to complete the experimental simulation system platform. The system simulated flight information for a random day, and the experiment was conducted with the goal of minimizing vehicle mileage, to verify the effectiveness of the present invention.

[0062] Figure 3 The iterative curve of service vehicle mileage in the simulated airport environment of this invention is shown. The horizontal axis data represents the number of decision transformations (every hundred iterations), and the vertical axis represents the mileage of all service vehicles on the apron (taken as the reciprocal). It can be seen that the method of this invention can achieve dynamic decision optimization of the system for multiple objects, multiple tasks, and multiple objectives by establishing and improving the dynamic configuration scheduling decision algorithm based on data such as actual maps, actual working condition process parameters, optimization objectives, and constraints.

Claims

1. A design and deployment method for a smart apron multi-agent swarm intelligence decision-making system, characterized in that: The method includes the following steps performed sequentially: 1) Construct an edge perception layer consisting of vehicle state perception units and high-precision positioning units to collect real-time location and key operational status data of service vehicles; 2) Distribute multiple edge computing nodes in key operational areas on the apron and interface with the aforementioned edge perception layer and airport collaborative decision-making system to receive location and key operational status data from the edge perception layer and structured flight support data from the airport collaborative decision-making system, so as to achieve multi-source data fusion and local scheduling decisions. 3) Construct a physical crowd intelligence decision-making device within the edge computing node, and use the crowd intelligence decision-making device to fuse multi-source data received from the aforementioned edge computing node, generate the final scheduling instructions in real time, and issue them to the corresponding service vehicles. 4) Construct a visual human-machine interface connected to the edge computing node to display real-time data pushed by the edge computing node, so as to realize real-time monitoring of the status of the entire smart apron multi-agent swarm decision-making system.

2. The design and deployment method of the intelligent apron multi-agent swarm intelligence decision-making system according to claim 1, characterized in that: In step 1), the vehicle status perception unit adopts an industrial-grade on-board automatic diagnostic system terminal, which is inserted into the standard OBD-II diagnostic interface on each service vehicle. Through this interface, it communicates with the electronic control device of the service vehicle via a CAN bus. The automatic diagnostic system terminal includes a GPS / BeiDou dual-mode positioning module and a 5G / 4G communication module. The high-precision positioning unit includes an ultra-wideband positioning base station, an ultra-wideband positioning tag, and a positioning engine server; multiple ultra-wideband positioning base stations are deployed in a honeycomb grid pattern in key operating areas and main roads on the apron; at the same time, an ultra-wideband positioning tag is fixedly installed on the top of each service vehicle; the positioning engine server is integrated into the following edge computing nodes or deployed independently. The service vehicles include baggage carts, shuttle buses, refueling trucks, clean water trucks, sewage trucks, and aircraft towing vehicles.

3. The design and deployment method of the intelligent apron multi-agent swarm intelligence decision-making system according to claim 2, characterized in that: The vehicle status perception unit is the XYZ-2000 4G / 5G OBD collector; the ultra-wideband positioning base station uses the Decawave DW3000 series chip.

4. The design and deployment method of the intelligent apron multi-agent swarm intelligence decision-making system according to claim 1, characterized in that: In step 2), the key operating areas include the baggage sorting area, the flight boarding gate jet bridge area, the aircraft maintenance hangar, the apron main road and taxiway, the fuel supply station, and the cargo terminal. The edge computing node uses industrial-grade embedded AI computing equipment, with a built-in 5G communication module and 5G URLLC network interface. It conducts highly reliable data transmission with the service vehicle and cloud server where the edge perception layer is located through the 5G URLLC slicing network. The edge computing node also interfaces with the airport collaborative decision-making system. The interface method is to develop a multi-protocol adapter on the edge computing node.

5. The design and deployment method of the intelligent apron multi-agent swarm intelligence decision-making system according to claim 4, characterized in that: The edge computing node uses the NVIDIA Jetson AGX Orin module.

6. The design and deployment method of the intelligent apron multi-agent swarm intelligence decision-making system according to claim 1, characterized in that: In step 3), the collective intelligence decision-making device includes an environment state perception interface, an intelligent agent policy executor cluster system, and an instruction distribution and issuing system that are logically interconnected; wherein the intelligent agent policy executor cluster system is physically manifested as multiple parallel processes or threads running on a multi-core CPU or GPU on an edge computing node, each process / thread serving as an intelligent agent policy executor; each intelligent agent policy executor internally encapsulates and runs a pre-trained policy neural network model; The input end of the environmental state perception interface is connected to the network interface or memory buffer of the edge computing node to acquire fused global data and flight support structured data in real time. Then, using the internally pre-built state vector assembly logic, according to predefined rules and templates, it extracts, filters, and assembles the corresponding local observation state vector for each agent policy executor from the above data. The local observation state vector includes the real-time location, energy status, current task queue, and related flight support structured data of the service vehicle itself. The output end is connected to each agent policy executor in the agent policy executor cluster system through an internal inter-process communication mechanism to distribute the assembled local observation state vector to these agent policy executors in parallel. The intelligent agent policy executor continuously receives its own, assembled local observation state vector from the environmental state perception interface through an internal inter-process communication mechanism. Then, it calls its internal policy neural network model to perform forward propagation calculation on the input local observation state vector to obtain an instruction data packet representing the optimal action. Finally, it sends the instruction data packet to the instruction distribution and delivery system through the internal inter-process communication mechanism. The input end of the instruction distribution and issuing system is connected to each intelligent agent policy executor in the intelligent agent policy executor cluster system through an internal inter-process communication mechanism, which is used to collect instruction data packets output by all intelligent agent policy executors; the output end is directly connected to the 5G communication module of the edge computing node, which is used to issue the formatted specific scheduling instructions to the corresponding service vehicles through the 5G URLLC network.

7. The design and deployment method of the intelligent apron multi-agent swarm intelligence decision-making system according to claim 6, characterized in that: The policy neural network model is generated through an independent, offline centralized training system, which includes: Simulation environment server: running a high-fidelity apron digital twin system; Parameter server: Stores and manages the latest central policy network parameters; Multiple experience collectors: interact with the simulation environment to collect a large amount of interactive data; Model trainer: includes gradient calculation module and parameter update module.

8. The design and deployment method of the intelligent apron multi-agent swarm intelligence decision-making system according to claim 1, characterized in that: In step 4), the visual human-machine interface is deployed in the airport's dispatch and command center, duty rooms in key operational areas, and mobile dispatch terminals. It is used to provide airport operators with a centralized monitoring and visual interactive interface for the entire smart apron dispatch system. The interface includes an apron digital map component, a vehicle status table component, and a flight timeline component. These components constitute the core visualization components.

Images