Efficient cooperative control method for airport unmanned equipment

Abstract

The invention relates to the technical field of airport ground traffic automatic control, in particular to an efficient cooperative control method for airport unmanned equipment, which comprises the following steps of: acquiring and fusing airport scene multi-source heterogeneous data to construct unified comprehensive sensing information, and further establishing an intelligent agent model containing state, action and reward and punishment mechanisms for each unmanned equipment; on the basis, when an emergency task is responded, task assignment and conflict autonomous resolution are performed by adopting a dynamic game process fusing local bidding, reputation feedback and alliance game, and a decision strategy and path planning of equipment are continuously optimized by utilizing a deep reinforcement learning algorithm based on digital twin environment training; and finally, through distributed instruction execution, rapid, accurate and adaptive collaborative response of multiple types of unmanned equipment taking the real-time state of the aircraft as a core is realized.

Description

Technical Field

[0001] This invention relates to the field of airport ground traffic automation control technology, and in particular to a highly efficient collaborative control method for unmanned airport equipment. Background Technology

[0002] As an indispensable part of the modern transportation system, the civil aviation industry has extremely stringent safety requirements—any minor oversight can trigger systemic risks, threatening passenger lives and aircraft safety. Airports, as crucial nodes in this system, require their special ground equipment to handle not only highly time-sensitive and complex tasks, but also, due to serving aircraft—a unique entity—exhibiting operational logic drastically different from that of mines, ports, or public roads. From cargo transfer to emergency rescue, from perimeter patrols to passenger shuttles, the operation of each type of unmanned equipment (such as unmanned logistics vehicles, unmanned fire trucks, unmanned patrol vehicles, and unmanned shuttle buses) must prioritize the real-time status of the aircraft, ensuring precise coordination for its safe and efficient operation.

[0003] However, current multi-device collaborative control technologies still have significant limitations. Existing solutions mainly fall into two categories: one focuses on the collaborative mechanisms within or between unmanned systems, achieving basic functions such as formation and collision avoidance through communication protocols and control laws, but neglects the core characteristic of airport operations being "aircraft-centric," leading to a disconnect between equipment scheduling and flight dynamics and operational needs; the other, while constructing scheduling models for special vehicle groups and introducing technologies such as genetic algorithms and digital twins to optimize mileage and waiting time, is still limited to resource allocation within the vehicle group, failing to use the real-time status of the aircraft (such as the location of fuselage damage after an emergency landing and the situation of personnel being trapped) as the core basis for scheduling decisions, and also lacks accurate modeling of the operational connection logic between aircraft and vehicles.

[0004] This technological disconnect is particularly pronounced in extreme scenarios such as emergency aircraft landings. These events are characterized by dense task nodes, short decision-making windows, and highly complex resource coordination, requiring various types of unmanned equipment to be solely guided by aircraft needs, achieving rapid, precise, and adaptive collaborative responses. However, existing methods, lacking deep integration with the core role of aircraft, often lead to rigid scheduling schemes, delayed responses, and even resource misallocation, seriously threatening emergency response efficiency and aviation safety. Summary of the Invention

[0005] The purpose of this invention is to provide an efficient collaborative control method for unmanned airport equipment, which solves the problems of scheduling rigidity, response delay and resource mismatch caused by the inability of existing airport scheduling methods to conduct in-depth collaboration based on the real-time status of aircraft when dealing with emergencies.

[0006] To achieve the above objectives, the present invention provides an efficient collaborative control method for unmanned airport equipment, comprising the following steps: Collect multi-source heterogeneous data within the airport's operating area, and fuse the multi-source heterogeneous data to generate unified comprehensive sensing information; Based on the comprehensive perception information, an intelligent agent model is constructed for each unmanned device. The intelligent agent model includes state information for describing the device status, action information for describing the actions that the device can perform, and a reward and punishment mechanism for evaluating the effect of the actions. In response to the received emergency task, based on the state information of the intelligent agent model, the task is assigned to each unmanned device and a cooperative alliance is formed through a dynamic game process that integrates local bidding, reputation feedback and alliance game. At the same time, the task and path conflicts between devices are detected and resolved in real time during the game process. The decision-making strategy of the agent model is trained and optimized using a deep reinforcement learning algorithm to generate optimized task execution instructions and path planning instructions. The task execution instructions and path planning instructions are distributed to the corresponding unmanned devices to control them to perform collaborative operations.

[0007] This includes collecting multi-source heterogeneous data within the airport's operating area and fusing this multi-source heterogeneous data to generate unified comprehensive sensing information, specifically including: Collect multi-source heterogeneous data from roadside non-cooperative sensors, vehicle-mounted cooperative sensors, and airport surface surveillance systems; the multi-source heterogeneous data includes the position, speed, acceleration, and remaining resource data of each unmanned device, the three-dimensional pose, speed, and acceleration data of aircraft, as well as traffic signs, road conditions, and weather conditions data.

[0008] This includes collecting multi-source heterogeneous data within the airport's operating area, fusing the multi-source heterogeneous data, and generating unified comprehensive sensing information. Specifically, it also includes: The multi-source heterogeneous data is cleaned to remove noise and outliers, and then standardized or normalized to eliminate dimensional differences. The processed data is spatiotemporally registered to achieve time synchronization and spatial alignment. By using feature extraction and decision-level fusion methods, unified comprehensive perception information is generated.

[0009] Based on the comprehensive perception information, an intelligent agent model is constructed for each unmanned device. The intelligent agent model includes state information describing the device's status, action information describing the actions the device can perform, and a reward / punishment mechanism for evaluating the effectiveness of the actions. Specifically, it includes: The integrated sensing information is mapped to the status information, action information, and reward / punishment mechanism corresponding to each unmanned device; The status information includes the device's real-time task status, spatial location information based on the airport's high-precision map, remaining resources, and operational capability parameters. The action information includes a set of actions that the device can execute, including response tasks, collaborative operations, path transfers, and fault reporting, and the triggering conditions of the actions are dynamically associated with the status information. The reward and punishment mechanism uses task completion rate, response latency, and resource consumption rate as core indicators. It assigns positive rewards to successful task completion and conflict avoidance, and assigns negative penalties to task overtime and resource waste.

[0010] In response to received emergency tasks, based on the state information of the intelligent agent model, a dynamic game process integrating local bidding, reputation feedback, and alliance game is used to assign tasks to each unmanned device and form a cooperative alliance. Simultaneously, during the game process, task and path conflicts between devices are detected and resolved in real time. Specifically, this includes: An emergency task containing task attributes is issued by the task requester; Each unmanned device, based on its own status information, engages in local bidding with the revenue parameters of task value, resource adaptability, and response cost as the revenue parameters. Based on the bidding results and the reputation feedback corresponding to the equipment's historical task completion quality, the initial task assignment is determined; For tasks requiring multi-device collaboration, a cooperative game mechanism is used to guide the unmanned devices to form a collaborative alliance in order to achieve shared benefits. In the dynamic game process, an embedded conflict detection module is used to identify path overlap and resource contention between devices in real time, and to impose additional cost penalties on the conflicting parties by adjusting the payout parameters, thereby guiding the devices to autonomously avoid conflicts.

[0011] Specifically, deep reinforcement learning algorithms are used to train and optimize the decision-making strategy of the agent model, generating optimized task execution instructions and path planning instructions, including: In light of the heterogeneity, high dynamism, and distributed collaboration requirements faced by unmanned airport equipment in emergency scenarios, a deep reinforcement learning training framework is constructed that integrates Proximal Policy Optimization (PPO) and Multi-Agent Deep Deterministic Policy Gradient (MADDPG). The PPO algorithm is used to enable rapid policy iteration and path adjustment for a single unmanned device in a complex airport environment, while the MADDPG algorithm is used to achieve collaborative decision-making and task priority allocation among multiple unmanned devices through a centralized critic mechanism. Design a reward function that integrates multiple key factors. The reward function includes four core elements: priority matching degree, path feasibility, agent conflict degree, and response timeliness. The weight of each element is dynamically adjusted through a dynamic weighting mechanism. A digital twin simulation environment is constructed based on the real airport topology. The policy network is trained in the environment through the interaction between the agent and the environment. An experience playback and target network delay update mechanism are introduced to improve the training stability. The agent learns optimized strategies through training, and generates task execution instructions and path planning instructions for unmanned equipment based on these optimized strategies.

[0012] The training process enables the agent to learn an optimization strategy, and based on this optimization strategy, task execution instructions and path planning instructions for the unmanned device are generated, specifically including: The PPO algorithm is applied to a single unmanned device to help it accurately prioritize tasks and effectively learn path selection strategies. The MADDPG algorithm is applicable to scenarios where multiple unmanned devices coexist. When these devices have resource coupling or need to cooperate to complete specific tasks, the algorithm trains the Critic network in a centralized manner, prompting each agent to make the optimal action choice based on a full consideration of the behavior of other agents.

[0013] This invention discloses an efficient collaborative control method for airport unmanned equipment. It constructs unified comprehensive perception information by collecting and fusing multi-source heterogeneous data from the airport surface, and then establishes an intelligent agent model for each unmanned device, including its state, actions, and reward / punishment mechanisms. Based on this, when responding to emergency tasks, a dynamic game process integrating local bidding, reputation feedback, and alliance game is used for task assignment and autonomous conflict resolution. Furthermore, a deep reinforcement learning algorithm trained in a digital twin environment is used to continuously optimize the device's decision-making strategy and path planning. Finally, through distributed command execution, it achieves rapid, accurate, and adaptive collaborative response for multiple types of unmanned equipment, with the real-time status of the aircraft as the core. By employing multi-agent modeling and deep reinforcement learning, this system enables unmanned equipment to autonomously respond to emergencies, freeing it from reliance on fixed rules and human experience, and adapting to diverse tasks and heterogeneous resources. A dynamic game theory mechanism resolves task competition and path conflicts between equipment, while distributed scheduling optimization reduces global computational latency, improving collaborative response speed by over 30% (based on digital twin scenario test data). An online disturbance adjustment mechanism can complete disturbance response and strategy correction within 10 seconds, preventing scheduling interruptions due to emergencies and ensuring the continuous progress of multiple service tasks related to aircraft support. The multi-agent modeling framework supports adding new unmanned equipment types, requiring only the expansion of the state and action space without reconstructing the overall algorithm system, thus adapting to the future intelligent upgrade needs of airports. This system addresses the problems of scheduling rigidity, response delays, and resource mismatches caused by the inability of existing airport scheduling methods to achieve deep collaboration based on real-time aircraft status when responding to emergencies. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0015] Figure 1 This is a flowchart of the steps of the efficient collaborative control method for unmanned airport equipment of the present invention. Detailed Implementation

[0016] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0017] Please see Figure 1 ,in, Figure 1 This is a flowchart of the steps of the efficient collaborative control method for unmanned airport equipment of the present invention.

[0018] This invention provides a highly efficient collaborative control method for unmanned airport equipment, comprising the following steps: S101: Collect multi-source heterogeneous data within the airport operating area, and perform fusion processing on the multi-source heterogeneous data to generate unified comprehensive perception information; S102: Based on the comprehensive perception information, construct an intelligent agent model for each unmanned device. The intelligent agent model includes state information for describing the device status, action information for describing the actions that the device can perform, and a reward and punishment mechanism for evaluating the effect of the actions. S103: In response to the received emergency task, based on the state information of the intelligent agent model, through a dynamic game process that integrates local bidding, reputation feedback and alliance game, tasks are assigned to each unmanned device and a cooperative alliance is formed. At the same time, during the game process, task and path conflicts between devices are detected and resolved in real time. S104: Using a deep reinforcement learning algorithm, the decision-making strategy of the agent model is trained and optimized to generate optimized task execution instructions and path planning instructions; S105: Distribute the task execution instructions and path planning instructions to the corresponding unmanned equipment to control them to perform collaborative operations.

[0019] Specifically, data acquisition relies on various sensors deployed on the airport surface. Among these, environmental perception sensors utilize external sensors to perceive traffic signs, road conditions, weather conditions, personnel status, and the status (position, speed, acceleration, etc.) of obstacles, including other equipment, as well as their future status. Environmental perception sensors determine the ability of unmanned equipment to interact with the external environment and form the foundation of the hardware architecture for autonomous driving equipment. Environmental perception sensors are divided into two main categories: autonomous and cooperative. Autonomous sensors typically emit energy in the form of electromagnetic waves and measure the return time to determine parameters such as distance. Existing autonomous environmental perception sensors include ultrasonic radar, millimeter-wave radar, and lidar. Cooperative environmental sensing, on the other hand, receives perception messages from vehicle-to-everything (V2X) or vehicle-road cooperative networks, such as cameras based on light and infrared. Comprehensive three-dimensional perception is achieved by establishing a vehicle-machine-airport integrated three-dimensional perception, transmission, and fusion framework. Based on a lidar target detection model, target feature information is extracted at multiple temporal and spatial levels. Building upon the fusion of non-cooperative and cooperative sensor information from roadside equipment deployed on-site, information from vehicle-mounted lidar, vehicle-mounted GPS positioning, BeiDou positioning, ADS-B IN, field surveillance radar, panoramic video, and multi-point positioning systems is incorporated. Chain synchronization and feature matching methods are used to perform multi-dimensional decision-level data fusion of high-dynamic data from on-site, vehicle-mounted, and airborne sensors to obtain aircraft 3D pose data, comprehensively perceiving vehicle-aircraft-field information. Specific data includes: the position, speed, and acceleration of unmanned equipment; traffic signs; road conditions; weather conditions; driver status; aircraft position, speed, and acceleration; obstacle status (position, speed, acceleration, etc.) including other equipment; and the future status of the unmanned equipment. Other data that airport systems can obtain include flight information, real-time task status (idle / in operation / fault), spatial location (coordinate information based on high-precision airport maps), remaining resources (such as water volume of unmanned fire trucks, battery power of rescue robots, number and location of support vehicles, etc.), and operational capability parameters (such as rescue coverage, operational efficiency, and support scope).

[0020] Subsequently, the multi-source heterogeneous data is processed and transmitted. Data processing adopts a multi-source fusion method, and data transmission is carried out through 5G and 5GAeroMACS. The specific data processing method is as follows: First, data preprocessing is performed, (1) Data cleaning: remove "dirty data", the goal is to eliminate noise, correct errors, and process missing values. The methods used are: noise removal: smoothing is used for continuous data (such as temperature, speed) (moving average method, wavelet transform); anomaly detection is used for discrete data (such as text labels) (isolated forest, DBSCAN clustering). (2) Data standardization / normalization: eliminate "differences in units". The units (units) and numerical ranges of multi-source heterogeneous data may differ greatly (such as "vehicle speed (km / h)" and "distance (m)"). Direct fusion will lead to "features with large values ​​dominating the fusion result", so a unified scale is required: standardization (Z-Score): suitable for scenarios where the data is approximately normally distributed. The formula is: ( The mean, The standard deviation is 0, and the output data has a mean of 0 and a standard deviation of 1. Normalization (Min-Max): suitable for scenarios with irregular data distribution, the formula is: The output data range is compressed to [0,1] or [-1,1]. (3) Data registration: To achieve "spatiotemporal alignment", multi-source heterogeneous data may have time asynchrony (such as differences in sensor sampling frequency) or spatial non-overlap (such as different camera viewpoints). Registration is required to ensure that "information corresponds at the same time and in the same space". (4) Data conversion: To solve "type heterogeneity", different types of data (such as text, images, and numerical values) are converted into a unified fusion input format. Then, data fusion is performed. First, features (such as image edges and textures, and sensor data statistics) are extracted from each data source, and then the features are integrated. The specific method is to use the CNN method.

[0021] Subsequently, based on the comprehensive perception information, an agent model is constructed for each unmanned device, i.e., multi-agent modeling is performed for heterogeneous multi-service resources. This modeling framework provides a formal description of the devices, specifically including: State space definition: It covers the real-time task status of the equipment (idle / in execution / fault), spatial location (coordinate information based on high-precision airport map), remaining resources (such as water volume of unmanned fire trucks and power of rescue robots), and operational capability parameters (such as rescue coverage and operational efficiency).

[0022] Action space design: Define the set of actions that the device can perform, including "responding to tasks", "collaborative operation", "path transfer" and "fault reporting", and the action triggering conditions are dynamically associated with the device state space (e.g., when the remaining power is below the threshold, the "long distance transfer" action is prohibited).

[0023] Reward and penalty function construction: With "task completion rate", "response latency" and "resource consumption rate" as core indicators, positive rewards are given to the device for completing tasks and avoiding conflicts, and negative penalties are given for tasks exceeding time limits and wasting resources.

[0024] Distributed interface adaptation: Considering the complex communication environment of airports (such as weak signal areas caused by building obstruction), a local sensing interface is designed to support distributed collaboration between devices based on information interaction with neighboring devices, without relying on global information synchronization, thus reducing communication latency.

[0025] Then, in response to the received emergency tasks, tasks are assigned through dynamic game modeling and coordination mechanisms for scheduling tasks.

[0026] This process is abstracted as a multi-agent dynamic game process, with the specific mechanism as follows: Dynamic game model design: Each unmanned device is defined as an independent player in the game. The core payoff parameters are "task value" (e.g., the priority of aircraft personnel rescue is significantly higher than site cleanup), "resource adaptability" (e.g., the capability matching coefficient between fire fighting tasks and unmanned fire trucks), and "response cost" (e.g., the energy consumption and time conversion value of equipment traveling to and from the aircraft operation point). A non-zero-sum game model is constructed. By quantifying the differences in parameters, high-value tasks are given priority in response, while avoiding excessive resource consumption by a single device.

[0027] Multi-mechanism collaborative guidance strategy: Integrating three layers of mechanisms to achieve efficient collaboration: "Local competitive auction" mechanism: The airport emergency command virtual node publishes task requirements, and equipment bids in real time based on its own capabilities (such as fire extinguishing coverage and rescue speed). The system selects the optimal equipment to undertake the task based on a comprehensive score of "value-suitability-cost"; "Reputation feedback" mechanism: The quality of task completion (such as rescue timeliness and resource utilization) is converted into equipment reputation value, dynamically increasing the bidding weight of high-reputation equipment and incentivizing equipment to continuously optimize its operational performance; "Alliance game" mechanism: For collaborative tasks such as large-area fire extinguishing and multi-stage aircraft support, equipment is guided to spontaneously form temporary alliances based on complementary capabilities. Through preset revenue distribution rules (such as splitting the total alliance revenue according to contribution ratio), the system promotes proactive collaboration rather than disorderly competition among equipment.

[0028] Dynamic optimization of conflict resolution: An "airport scenario conflict detection module" is embedded in the game process to capture scenario-based conflicts in real time, such as the overlap of equipment paths with aircraft operating areas and the competition for emergency resources (such as fire water sources). By dynamically adjusting the game payoff parameters (such as imposing high cost penalties on equipment that intrudes into the runway and deducting payoff value from the resource grabbing party), the equipment is guided to autonomously avoid conflicts in decision-making, and finally achieve a balance of "ensuring efficiency in competition and preventing internal friction in cooperation".

[0029] Furthermore, in combination with the dual requirements of autonomous decision-making and group collaboration in airport emergency scenarios, an innovative distributed scheduling and path optimization scheme driven by deep reinforcement learning is adopted to achieve dynamic iteration and continuous evolution of decision-making strategies.

[0030] A hierarchical training framework is specifically constructed that integrates PPO (Proximal Policy Optimization) and MADDPG (Multi-Agent Deep Deterministic Policy Gradient). The PPO algorithm focuses on intelligent optimization of single devices, adapting to the individual decision-making needs of airport unmanned equipment (such as emergency rescue robots, unmanned fire trucks, unmanned guidance vehicles, and unmanned towing vehicles). By introducing a policy update cutoff threshold, it suppresses drastic policy fluctuations caused by sudden aircraft conditions (such as changes in emergency landing locations or the addition of new rescue points), ensuring that single devices are trained stably during task priority ranking (such as prioritizing responses to trapped personnel signals) and path dynamic planning (such as bypassing temporarily blocked runway areas), and quickly converge to the local optimal policy for the adapted scenario. The MADDPG algorithm is designed for multi-device collaboration scenarios. It is designed with a "centralized training and distributed execution" mechanism for airport resource coupling (such as shared fire water sources) and multi-stage collaborative tasks (such as "firefighting-patient transfer-equipment replenishment" linkage). It uses a centralized Critic network to integrate global information (including the status of all equipment, real-time aircraft requirements, and dynamic restricted areas of the airport) for joint training, so that each agent can predict the behavior of other equipment when making decisions (such as fire trucks can avoid the transfer path of ambulances in advance), and finally achieve conflict-free collaboration in distributed execution.

[0031] By combining the two algorithms, we can ensure the rapid response capability of a single device to sudden airport scenarios, and improve the overall collaborative efficiency of multiple devices, thus forming an intelligent decision-making system that is adapted to the complex airport environment.

[0032] To address the unique characteristics of airport emergency dispatch scenarios, an innovative multi-dimensional reward function system is designed. This system includes not only basic rewards based on task urgency (e.g., aircraft emergency landing rescue has a higher weight than site clearing) and completion efficiency, but also positive incentives based on route efficiency (e.g., avoiding detours and congested sections), energy consumption costs (e.g., electricity / water consumption conversion), and multi-agent collaboration quality (e.g., task seamlessness and information exchange effectiveness). Furthermore, a tiered penalty mechanism is implemented for task response delays (exceeding the preset time window), duplicate responses (multiple devices competing for the same task), and resource conflicts (e.g., competition for fire hydrants). The weights of each factor are dynamically optimized through airport dispatch priority strategies (e.g., increasing the response timeliness weight to above 0.3 in emergency scenarios) to achieve precise matching between reward signals and actual operational needs.

[0033] The training environment utilizes real airport topology data (including static elements such as runways, terminals, and fire lanes, as well as dynamic information such as flight dynamics and equipment deployment) to construct a digital twin simulation system. This system can reproduce over 10 emergency scenarios, including single-point incidents (such as localized aircraft fires), concurrent multi-task events (such as simultaneous fuselage fire suppression and personnel transfer), and road disruptions (such as rescue routes being blocked by obstacles). The agent accumulates experience data through continuous interaction with the simulation environment. Combined with a priority experience replay mechanism (prioritizing the learning of high-value emergency cases), the system optimizes the policy network. This improves the model's convergence stability in complex scenarios and enhances its generalization ability to unseen emergencies, ensuring that training results can be directly transferred to real airport environments.

[0034] To address the core characteristics of airport unmanned equipment scheduling—namely, "dynamic state changes and decision-making dependent on real-time information"—the scheduling environment is formalized as a Markov decision process tailored to the specific needs of the scenario. ,in, For state space, For the action space, It is the state transition probability function. For the reward function, Let [the discount factor] be [the discount factor]. The policy function is defined as [the policy function]. State transition function and award letter number.

[0035] state space Around the core elements of the airport dispatch system, a three-dimensional information architecture of "resource-task-environment" is constructed to comprehensively cover the key data required for dispatch decisions: Resource status focuses on dispatch units such as fire trucks and ambulances, including not only real-time geographical location and travel speed, but also detailed to the remaining executable resource quantity (such as fire truck water volume and ambulance material reserves) and historical task execution records, accurately reflecting the dynamics of equipment capabilities; Task status is centered on "space-attribute-timeliness," clearly presenting the spatial distribution, type (such as emergency rescue and flight support), urgency level, and remaining time limit of tasks, providing a basis for prioritization; Environmental status captures scene variables such as airport road accessibility, taxiway obstacles and visibility, and weather conditions in real time, ensuring that status perception closely matches the actual operating environment. These three status dimensions are interconnected and complementary, jointly forming the complete information foundation of the dispatch environment.

[0036] Action space To address the decision-making needs of intelligent agents, a two-layer action system of "scheduling decision-making + path control" is designed to adapt to the complexity of multi-service support scenarios at airports. In multi-task parallel scenarios, the intelligent agent needs to make real-time decisions on whether to accept new tasks, select specific tasks, and plan the path and driving strategy to reach the task point. Task selection requires determining the target from a discrete task list, which is a discrete action; path control requires optimizing the driving trajectory and speed in a continuous space, which is a continuous action. Together, they form a hybrid action space of "discrete + continuous." After the intelligent agent executes an action, the airport scheduling environment status is updated synchronously with equipment movement and task progress. The system then provides immediate rewards, and the intelligent agent enters the next decision-making cycle, constructing a complete interactive closed loop of "action-state transition-reward."

[0037] To address the decision-making needs of different devices in airport dispatching scenarios, two reinforcement learning algorithms are adapted to specific scenarios: the Proximal Policy Optimization (PPO) algorithm is more suitable for individual dispatching units with independent decision-making capabilities, such as fire trucks, ambulances, tow trucks, and guidance vehicles. It ensures training stability through a policy update truncation mechanism and improves sample utilization by leveraging the multiple reuse characteristics of sampled data, efficiently supporting single devices in making independent decisions such as task priority ranking and dynamic path adjustment. The Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm focuses on scenarios where multiple devices have resource coupling (such as shared fire water sources and shared rescue channels) or collaborative task requirements (such as "firefighting-injury transfer" joint operations). With its "centralized training and distributed execution" framework, it can accurately resolve the problems of task contention and repeated responses among multiple devices. At the same time, it ensures rapid convergence of multi-agent collaborative behavior policies through global information-assisted training, adapting to the complex group collaborative dispatching needs of airports.

[0038] In the PPO algorithm, the core objective of policy optimization is to maximize the objective function that combines the constraints of the policy improvement direction and the update magnitude. .

[0039] in, It is the ratio of the old and new strategies. For the estimation of the advantage function, A truncation threshold is used. In the policy optimization stage of the PPO algorithm, a clip function is introduced to dynamically constrain the policy update magnitude—that is, to limit the probability ratio of the new policy to the old policy to a preset range. This effectively suppresses drastic policy fluctuations caused by sudden disturbances in the airport scenario (such as temporary task insertion or sudden path blockage), fundamentally preventing policy divergence and significantly improving the stability and robustness against interference in the training process. Based on this advantage, the PPO algorithm can be widely adapted to various scheduling units in airport scenarios, efficiently supporting task priority ranking (such as prioritizing response to aircraft emergency needs) and dynamic path selection policy learning for equipment such as fire trucks, ambulances, and shuttle buses, ensuring that single-equipment decision-making can stably adapt to the complex and ever-changing airport operating environment.

[0040] For airport support tasks requiring multi-agent collaboration, such as the joint transport of passengers by multiple shuttle buses, the system employs the MADDPG algorithm to construct a dedicated multi-agent centralized training framework. Based on the core logic of "centralized training and distributed execution," the framework incorporates global airport scheduling information (such as real-time passenger numbers at passenger distribution points, flight departure countdowns, and the status of other support vehicles) during the training phase. This provides comprehensive data support to the Critic network of each agent, ensuring the network can accurately evaluate the overall benefits of multi-agent collaborative actions. This lays the training foundation for efficient collaboration during subsequent distributed execution, as defined below: .

[0041] in, It is the policy of the i-th agent. Based on the overall state and actions The method employs a centralized value function. It allows each agent to make optimal action choices based on the behavior of others, making it suitable for complex scenarios such as multi-vehicle collaboration, task allocation, and conflict avoidance. Based on this design, each agent can perceive the global state of the system during the learning phase, thereby forming strategic coordination during the execution phase, avoiding conflicts or repetitive responses, and improving collaborative efficiency.

[0042] The reward function is designed as follows: .

[0043] in, This represents the basic reward for successfully completing a task, taking into account the urgency of the task and the efficiency of completion. Measuring route efficiency and energy consumption; Reflects the quality of multi-agent collaboration; The penalty for delayed task response; Used to limit duplicate task responses or resource conflicts. Weight parameters The system will be optimized based on the airport's scheduling priority strategy to ensure that the strategy learning direction aligns with emergency response objectives.

[0044] Ultimately, the task execution instructions (such as task acceptance sequences and operation timing planning) and path planning instructions (such as optimal driving trajectories and obstacle avoidance routes) generated through training and optimization will be precisely distributed to the corresponding unmanned equipment (such as fire trucks and shuttle buses) in a distributed manner, driving each device to complete the final operation according to the collaborative strategy. The entire system architecture adopts a distributed decision-making mode, allowing devices to autonomously achieve collaborative responses based on local perception information (such as the status of neighboring devices and the dynamics of the surrounding environment), without relying on real-time synchronization of global information. This design significantly improves the robustness of operation in the complex communication environment of airports (such as weak signal areas caused by building obstruction and interference from concurrent communication of multiple devices), ensuring the continuity and reliability of collaborative operations.

[0045] Example 1: Using a 4E-level airport as the test subject, a scenario simulating a Boeing 737 aircraft catching fire and making an emergency landing was simulated. Ten unmanned devices (including three unmanned fire trucks, four rescue robots, and three unmanned ambulance transport vehicles) were deployed, and the dispatch response was achieved through the method of this invention. Multi-agent modeling stage: Complete the initialization of the equipment state space, and clarify the parameters of the unmanned fire truck such as "water volume 1000L, fire extinguishing coverage radius 5m", and the parameters of the rescue robot such as "maximum load 50kg, climbing angle 30°". Dynamic game phase: The emergency command center issues three tasks: "fuselage fire extinguishing", "personnel transfer" and "runway clearing". The unmanned fire truck wins the "fuselage fire extinguishing" task through bidding, and the rescue robot and the unmanned ambulance transfer vehicle form an alliance to complete the "personnel transfer" task. Deep reinforcement learning optimization phase: Based on the MADDPG algorithm, the equipment autonomously plans its path, avoids restricted areas on the runway, and reduces the path overlap rate between unmanned fire trucks and rescue robots to below 5%.

[0046] By employing multi-agent modeling and deep reinforcement learning, this system enables unmanned equipment to autonomously respond to emergencies, freeing it from reliance on fixed rules and human experience, and adapting to diverse tasks and heterogeneous resources. A dynamic game theory mechanism resolves task competition and path conflicts between equipment, while distributed scheduling optimization reduces global computational latency, improving collaborative response speed by over 30% (based on digital twin scenario test data). An online disturbance adjustment mechanism can complete disturbance response and strategy correction within 10 seconds, preventing scheduling interruptions due to emergencies and ensuring the continuous progress of multiple service tasks related to aircraft support. The multi-agent modeling framework supports adding new unmanned equipment types, requiring only the expansion of the state and action space without reconstructing the overall algorithm system, thus adapting to the future intelligent upgrade needs of airports. This system addresses the problems of scheduling rigidity, response delays, and resource mismatches caused by the inability of existing airport scheduling methods to achieve deep collaboration based on real-time aircraft status when responding to emergencies.

[0047] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A highly efficient collaborative control method for unmanned airport equipment, characterized in that, Includes the following steps: Collect multi-source heterogeneous data within the airport's operating area, and fuse the multi-source heterogeneous data to generate unified comprehensive sensing information; Based on the comprehensive perception information, an intelligent agent model is constructed for each unmanned device. The intelligent agent model includes state information for describing the device status, action information for describing the actions that the device can perform, and a reward and punishment mechanism for evaluating the effect of the actions. In response to the received emergency task, based on the state information of the intelligent agent model, the task is assigned to each unmanned device and a cooperative alliance is formed through a dynamic game process that integrates local bidding, reputation feedback and alliance game. At the same time, the task and path conflicts between devices are detected and resolved in real time during the game process. The decision-making strategy of the agent model is trained and optimized using a deep reinforcement learning algorithm to generate optimized task execution instructions and path planning instructions. The task execution instructions and path planning instructions are distributed to the corresponding unmanned devices to control them to perform collaborative operations.

2. The efficient collaborative control method for unmanned airport equipment as described in claim 1, characterized in that, Collect multi-source heterogeneous data within the airport's operating area, and fuse this multi-source heterogeneous data to generate unified comprehensive sensing information, specifically including: Collect multi-source heterogeneous data from roadside non-cooperative sensors, vehicle-mounted cooperative sensors, and airport surface surveillance systems; the multi-source heterogeneous data includes the position, speed, acceleration, and remaining resource data of each unmanned device, the three-dimensional pose, speed, and acceleration data of aircraft, as well as traffic signs, road conditions, and weather conditions data.

3. The efficient collaborative control method for unmanned airport equipment as described in claim 2, characterized in that, Collecting multi-source heterogeneous data within the airport's operating area and fusing the multi-source heterogeneous data to generate unified comprehensive sensing information, specifically also includes: The multi-source heterogeneous data is cleaned to remove noise and outliers, and then standardized or normalized to eliminate dimensional differences. The processed data is spatiotemporally registered to achieve time synchronization and spatial alignment. By using feature extraction and decision-level fusion methods, unified comprehensive perception information is generated.

4. The efficient collaborative control method for unmanned airport equipment as described in claim 3, characterized in that, Based on the comprehensive perception information, an intelligent agent model is constructed for each unmanned device. The intelligent agent model includes state information describing the device's condition, action information describing the actions the device can perform, and a reward / penalty mechanism for evaluating the effectiveness of the actions. Specifically, it includes: The integrated sensing information is mapped to the status information, action information, and reward / punishment mechanism corresponding to each unmanned device; The status information includes the device's real-time task status, spatial location information based on the airport's high-precision map, remaining resources, and operational capability parameters. The action information includes a set of actions that the device can execute, including response tasks, collaborative operations, path transfers, and fault reporting, and the triggering conditions of the actions are dynamically associated with the status information. The reward and punishment mechanism uses task completion rate, response latency, and resource consumption rate as core indicators. It assigns positive rewards to successful task completion and conflict avoidance, and assigns negative penalties to task overtime and resource waste.

5. The efficient collaborative control method for unmanned airport equipment as described in claim 4, characterized in that, In response to received emergency tasks, based on the state information of the intelligent agent model, a dynamic game process integrating local bidding, reputation feedback, and alliance game is used to assign tasks to each unmanned device and form a cooperative alliance. Simultaneously, during the game process, task and path conflicts between devices are detected and resolved in real time, specifically including: An emergency task containing task attributes is issued by the task requester; Each unmanned device, based on its own status information, engages in local bidding with the revenue parameters of task value, resource adaptability, and response cost as the revenue parameters. Based on the bidding results and the reputation feedback corresponding to the equipment's historical task completion quality, the initial task assignment is determined; For tasks requiring multi-device collaboration, a cooperative game mechanism is used to guide the unmanned devices to form a collaborative alliance in order to achieve shared benefits. In the dynamic game process, an embedded conflict detection module is used to identify path overlap and resource contention between devices in real time, and to impose additional cost penalties on the conflicting parties by adjusting the payout parameters, thereby guiding the devices to autonomously avoid conflicts.

6. The efficient collaborative control method for unmanned airport equipment as described in claim 5, characterized in that, Using deep reinforcement learning algorithms, the decision-making strategy of the agent model is trained and optimized to generate optimized task execution instructions and path planning instructions, specifically including: In light of the heterogeneity, high dynamism, and distributed collaboration requirements faced by unmanned airport equipment in emergency scenarios, a deep reinforcement learning training framework is constructed that integrates Proximal Policy Optimization (PPO) and Multi-Agent Deep Deterministic Policy Gradient (MADDPG). The PPO algorithm is used to enable rapid policy iteration and path adjustment for a single unmanned device in a complex airport environment, while the MADDPG algorithm is used to achieve collaborative decision-making and task priority allocation among multiple unmanned devices through a centralized critic mechanism. Design a reward function that integrates multiple key factors. The reward function includes four core elements: priority matching degree, path feasibility, agent conflict degree, and response timeliness. The weight of each element is dynamically adjusted through a dynamic weighting mechanism. A digital twin simulation environment is constructed based on the real airport topology. The policy network is trained in the environment through the interaction between the agent and the environment. An experience playback and target network delay update mechanism are introduced to improve the training stability. The agent learns optimized strategies through training, and generates task execution instructions and path planning instructions for unmanned equipment based on these optimized strategies.

7. The efficient collaborative control method for unmanned airport equipment as described in claim 6, characterized in that, The agent learns an optimized strategy through the training process, and generates task execution instructions and path planning instructions for the unmanned device based on the optimized strategy. Specifically, this includes: The PPO algorithm is applied to a single unmanned device to help it accurately prioritize tasks and effectively learn path selection strategies. The MADDPG algorithm is applicable to scenarios where multiple unmanned devices coexist. When these devices have resource coupling or need to cooperate to complete specific tasks, the algorithm trains the Critic network in a centralized manner, prompting each agent to make the optimal action choice based on a full consideration of the behavior of other agents.

Images