Airport ground service task scheduling method based on collaborative multi-agent reinforcement learning

Abstract

The invention discloses an airport ground service task scheduling method based on collaborative multi-agent reinforcement learning, and relates to the technical field of airport operation management and intelligent scheduling, and the method comprises the steps: S1, obtaining the real-time position of a flight, the occupation state of a machine position, the operation progress of a special vehicle, and the sign-in data of personnel in a rigid business system, s2, inputting the first running state snapshot into a trained multi-agent reinforcement learning model, outputting a first decision instruction set for resource scheduling by the model, and enabling each instruction to carry an identifier for sending an agent; according to the airport ground service task scheduling method based on collaborative multi-agent reinforcement learning, the real-time performance, the self-adaptability and the global optimization capability of ground service resource scheduling are improved, the manual intervention frequency is reduced, and the airport operation efficiency and the overall operation stability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of airport operation management and intelligent scheduling, and particularly relates to an airport ground service task scheduling method based on collaborative multi-agent reinforcement learning. BACKGROUND

[0002] In the field of airport operation management, ground service scheduling is a core link for ensuring flight punctuality and operation efficiency, and the decision quality directly affects passenger experience and airport operation cost, so it is necessary to realize intelligent scheduling and adaptive optimization. At present, airports generally rely on scheduling systems based on fixed rules or manual experience. When facing dynamic and complex scenes such as large-scale flight delays and sudden resource conflicts, the strategy adjustment of this kind of method often relies on manual intervention and is slow to respond, and it is difficult to realize the optimal real-time matching of global resources, which exposes the fundamental deficiency of the existing system in dealing with high dynamic uncertainty environment. This deficiency leads to the urgent need to introduce multi-agent reinforcement learning models with autonomous decision-making ability into actual operation systems, but there are serious technical challenges in moving from laboratory training environment to real business systems. A core difficulty is the "safe and controllable fusion" between the decision-making logic of the model and the existing rigid business system. The strategy learned by the model through interaction with the environment is dynamic and implicit, while the existing airport operation management system and resource management system carry stable and reliable business processes, and there are essential differences between the decision-making mechanisms of the two. This difference directly leads to the second key technical difficulty: "gradual safe import" of the learned strategy. If the model strategy suddenly takes over completely, its unpredictable decision-making behavior may conflict with the safety boundaries of the existing system, causing operation risks. Specifically, due to the dynamic and implicit nature of the model strategy, it cannot be clearly traced and pre-verified like traditional software modules, which makes it extremely difficult to directly embed it into the flight information database and resource management system that require high reliability. SUMMARY

[0003] The present application relates to the technical field of airport operation management and intelligent scheduling, and particularly relates to an airport ground service task scheduling method based on collaborative multi-agent reinforcement learning.

[0004] To achieve the above object, the application provides the following technical scheme: an airport ground service task scheduling method based on cooperative multi-agent reinforcement learning is executed in an airport operation system configured with progressive import control parameters and provided with an instruction coordination channel, comprising S1, obtaining flight real-time position, stand occupation state, special vehicle operation progress and personnel check-in data in a rigid business system, and fusing to form a first operation state snapshot containing a timestamp; S2, inputting the first operation state snapshot into a trained multi-agent reinforcement learning model, the model outputting a first decision instruction set for resource scheduling, each instruction having an identifier of the issuing agent; S3, judging the compliance of the first decision instruction set according to a preset rule, generating a first verification pass signal if all the instructions comply with the preset rule, and replacing the instruction that violates the preset rule with an instruction generated by a rigid business system rule engine to generate a first corrected instruction set, and recording the rule conflict type and the source of the corrected instruction; S4, using a monitorable mechanism to compare the first decision instruction set and the first corrected instruction set for instruction difference, marking the offset reason of the instructions with differences, and storing the difference data, the rule conflict type, the associated resource state and the first operation state snapshot for snapshot association to form a first strategy offset record containing the agent behavior trajectory and the rule violation context.

[0005] Preferably, S1 comprises obtaining a standardized business data stream, the standardized business data stream being obtained by cleaning and converting heterogeneous source data of the rigid business system; generating a space state data set based on the standardized business data stream, the space state data set being generated based on spatial mapping of flight real-time position coordinates and determination of stand occupation state codes; constructing a job chain data based on the space state data set, the job chain data being established by matching job node identifiers in the special vehicle operation progress message and the personnel check-in log; and performing time correction and multi-dimensional fusion on the job chain data to generate the first operation state snapshot.

[0006] Preferably, S2 comprises obtaining the first operation state snapshot, inputting the first operation state snapshot into the multi-agent reinforcement learning model to obtain candidate action probability distribution for the shuttle bus, the cleaning team and the luggage tractor; performing joint action deduction according to the candidate action probability distribution, if the deduction result has a space-time conflict, updating the policy gradient by exchanging local observation information until a conflict-free joint action sequence is generated; and analyzing the conflict-free joint action sequence, converting the actions in the conflict-free joint action sequence into scheduling operation codes and associating the agent identifiers to generate the first decision instruction set for resource scheduling.

[0007] Preferably, the S3 comprises obtaining a first decision instruction set containing scheduling information, mapping the first decision instruction set to preset flight shortest transit time rules and slot conflict safety boundaries to construct a compliance verification matrix; extracting unique identifiers of rule violation instructions from the compliance verification matrix to generate a rule violation instruction index list; inputting the rule violation instruction index list into a rigid business system rule engine to generate rigid alternative instructions that meet safety specifications; and replacing corresponding instructions in the first decision instruction set with the rigid alternative instructions to generate a first modified instruction set.

[0008] Preferably, the S4 comprises obtaining the first decision instruction set and the first modified instruction set, extracting an instruction difference data list and retrieving a rule conflict type according to the instruction difference data list to generate a difference data set with a reason mark; calling a first running state snapshot according to the difference data set with the reason mark, extracting associated resource states and structurally associating with the first running state snapshot; analyzing the structurally associated data set to construct an agent behavior trajectory, mapping the rule conflict type and the associated resource state to the agent behavior trajectory to form a first strategy deviation record containing the agent behavior trajectory and the rule violation context.

[0009] Preferably, the S5 further comprises determining a proportion of the shuttle vehicle scheduling function taken over by the model decision in the current operation cycle according to the historical deviation sequence and the deviation frequency statistical result in the gradual introduction control parameter, specifically comprising obtaining the historical deviation sequence and the deviation frequency statistical result in the gradual introduction control parameter, extracting shuttle vehicle scheduling task feature data in the current operation cycle; generating a scheduling instruction stability index according to the historical deviation sequence, and converting the deviation frequency statistical result into a business risk level value; inputting the scheduling instruction stability index and the business risk level value into a gradual introduction control model to obtain an initial takeover proportion coefficient; correcting the initial takeover proportion coefficient with a model decision confidence to generate a dynamic weight factor, and determining the proportion of the shuttle vehicle scheduling function taken over by the model decision in the current operation cycle according to the dynamic weight factor.

[0010] Preferably, the method further includes S6: receiving a first verification pass signal or a first modified instruction set via the instruction coordination channel; selectively sending model instructions or hybrid instructions to the vehicle terminal of the rigid business system based on the current shuttle bus dispatching function takeover ratio; the system executing the instructions and updating the vehicle position and task status; specifically, receiving the first modified instruction set and the shuttle bus dispatching function takeover ratio value transmitted via the instruction coordination channel; extracting model instructions or recombining them to generate hybrid instructions based on the comparison result of the shuttle bus dispatching function takeover ratio value and a preset threshold; sending the model instructions or hybrid instructions to the vehicle terminal interface to drive the vehicle to perform actions, and collecting the vehicle position coordinate data and task status change data generated after execution; and updating the real-time vehicle mapping status based on the vehicle position coordinate data and task status change data.

[0011] Preferably, the process also includes S7: In the next decision cycle, obtain updated vehicle location and task status data, and combine them with new flight arrival events to form a second operational status snapshot. Specifically, this includes receiving real-time vehicle location coordinates and task status change data, and analyzing the vehicle road network distribution status using the airport road network topology model; calculating the driving distance value based on the vehicle road network distribution status and flight arrival events; if the driving distance value meets the preset shuttle service range, generating a task queue to be matched; obtaining the vehicle driving trajectory data and remaining battery information associated with the task queue to be matched, and constructing a multi-dimensional data matrix by combining the vehicle driving trajectory data and the remaining battery information; performing spatiotemporal dimension standardization processing on the multi-dimensional data matrix to generate a second operational status snapshot containing complete vehicle location and flight event information.

[0012] Preferably, it also includes S8, continuously accumulating strategy offset records, analyzing the offset trend within the time series index, and if the analysis results show that the conflict rate of the model decision with the shortest flight turnaround time rule is lower than a preset threshold in multiple consecutive periods, then the takeover ratio of the shuttle bus scheduling function during the gradual import process is automatically increased. Specifically, this includes obtaining strategy offset records, constructing an offset status dataset based on the strategy offset records, performing time series feature analysis on the offset status dataset, and generating an offset trend feature vector.

[0013] Preferably, step S8 further includes calculating the rule conflict rate value based on the offset trend feature vector; if the rule conflict rate value is lower than a preset threshold in multiple consecutive periods, a takeover ratio increment instruction is generated to automatically increase the takeover ratio of the shuttle bus scheduling function according to the takeover ratio increment instruction.

[0014] As can be seen from the above technical solution, the present invention has the following beneficial effects: This airport ground handling task scheduling method, based on collaborative multi-agent reinforcement learning, introduces a multi-agent reinforcement learning model to make intelligent decisions on ground handling resources such as shuttle buses. It also incorporates rule verification, monitorable offset recording, and a gradual import control mechanism into the rigid business system, achieving a safe and controllable integration of the learning-based scheduling strategy with the existing high-reliability airport operation system. The invention utilizes operational state snapshots to uniformly model the multi-dimensional states of flights, aircraft stands, vehicles, and personnel, enabling the model to continuously output scheduling decisions in dynamic and complex operating environments. Pre-set rules verify the compliance of model instructions and make necessary corrections, effectively preventing unpredictable model decisions from impacting operational safety. Simultaneously, by continuously recording and statistically analyzing the differences between model decisions and rule correction results, traceable agent behavior trajectories and strategy offset information are formed. Based on this, the model's takeover ratio of shuttle bus scheduling functions is dynamically adjusted, thus achieving a gradual and safe import of the reinforcement learning strategy. This method improves the real-time performance, adaptability, and global optimization capabilities of ground handling resource scheduling while ensuring the shortest flight turnaround time, aircraft stand safety boundaries, and vehicle operation constraints. It also reduces the frequency of manual intervention, improving airport operational efficiency and overall operational stability. Attached Figure Description

[0015] Figure 1 This is a flowchart of the airport ground service task scheduling method based on cooperative multi-agent reinforcement learning according to the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] like Figure 1As shown, this invention provides a technical solution: an airport ground service task scheduling method based on collaborative multi-agent reinforcement learning, executed in an airport operation system configured with progressive import control parameters and an instruction coordination channel, including: S1, acquiring real-time flight locations, gate occupancy status, special vehicle operation progress, and personnel check-in data from the rigid business system, and fusing them to form a first operational state snapshot containing timestamps; S2, inputting the first operational state snapshot into a trained multi-agent reinforcement learning model, the model outputting a first set of decision instructions for resource scheduling, each instruction bearing the identifier of the issuing agent; S3, judging the compliance of the first set of decision instructions according to preset rules, generating a first verification pass signal if all instructions comply with the preset rules, and replacing any instruction that violates the preset rules with an instruction generated by the rule engine of the rigid business system, generating a first set of corrected instructions, and recording the rule conflict type and the source of the corrected instruction; S4, using a monitoring mechanism to compare the instruction differences between the first set of decision instructions and the first set of corrected instructions, and marking the offset reasons for instructions with differences. The system performs the following steps: S5. It stores the difference data, rule conflict types, associated resource status, and first operating state snapshot together to form a first policy offset record containing the agent's behavior trajectory and rule violation context; S6. Based on the historical offset sequence and offset frequency statistics in the progressive import control parameters, it determines the proportion of shuttle bus scheduling functions taken over by the model decision in the current operating cycle; S7. The instruction coordination channel receives the first verification pass signal or the first corrected instruction set, and selectively sends model instructions or mixed instructions to the vehicle terminal of the rigid business system according to the current shuttle bus scheduling function takeover proportion. The system executes the instructions and updates the vehicle location and task status; S8. In the next decision cycle, it obtains the updated vehicle location and task status data, and combines it with new flight arrival events to form a second operating state snapshot; S9. It continuously accumulates policy offset records and analyzes the offset trend in the time series index. If the analysis results show that the conflict rate of the model decision for the shortest transit time rule of the flight is lower than the preset threshold in several consecutive cycles, the takeover proportion of the shuttle bus scheduling function during the progressive import process is automatically increased.

[0018] In the above implementation, the method uses the airport's existing rigid business systems as the basis for security and compliance constraints, and introduces a collaborative multi-agent reinforcement learning model to generate efficient ground support resource scheduling decisions. By collecting and aligning key operational elements such as flight positions, gate status, special vehicle operation progress, and personnel check-in in real time, a complete operational status snapshot is constructed, providing the model with highly reliable environmental state input.

[0019] In the multi-agent reinforcement learning model, each agent corresponds to a different scheduling decision-making body or resource unit, and they jointly output a set of scheduling instructions through a collaborative mechanism. The model output is not executed directly, but first undergoes compliance judgment by a rule verification module to ensure that it does not violate the safety, efficiency, and business constraints that must be followed in airport operations. When a model instruction conflicts with a rigid rule, the system automatically replaces it with an instruction generated by the rule engine and records all relevant conflict information completely.

[0020] By analyzing the differences between the original and revised instructions of the model, the system can construct a mapping relationship between the agent's behavioral trajectory and rule deviations, thus providing a traceable data foundation for subsequent model evaluation and policy evolution. Based on this, a progressive import control parameter is introduced to dynamically adjust the takeover ratio of the model during actual operation, enabling the model's decision-making ability to gradually replace manual or rigid rule scheduling under controllable and monitorable conditions.

[0021] By employing the above implementation method, a scheduling approach combining collaborative multi-agent reinforcement learning and a progressive introduction mechanism is introduced into the airport operation system, achieving continuous optimization of ground service resource scheduling efficiency while ensuring operational safety and rule compliance. On one hand, the model can make global collaborative decisions based on real-time operational status, effectively reducing shuttle bus empty runs and waiting times, thereby shortening flight turnaround times. On the other hand, through rule verification and policy offset recording mechanisms, the risks to actual operation caused by the instability of the model's initial decisions are avoided.

[0022] Furthermore, the gradual introduction of control parameters makes the takeover ratio of model decisions adjustable and adaptive, dynamically increasing model participation based on historical operational performance and reducing the cost of manual intervention. Long-term analysis of policy deviation trends also provides data support for model retraining and rule optimization, thereby improving the overall intelligence level and operational reliability of the airport operation system.

[0023] S1 includes acquiring standardized business data streams, which are obtained by cleaning and transforming heterogeneous source data from rigid business systems; generating spatial status datasets based on standardized business data streams, which are generated based on spatial mapping of real-time flight location coordinates and determination of gate occupancy status codes; constructing operation chain data based on the spatial status datasets, which are established by matching special vehicle operation progress messages with operation node identifiers in personnel check-in logs; and performing time correction and multi-dimensional fusion on the operation chain data to generate a first operational status snapshot.

[0024] In this implementation, the formation of standardized business data flows is centered on "unified field definitions" and "data quality convergence." The heterogeneous source data of rigid business systems typically comes from multiple subsystems such as flight operations, gate management, vehicle operations, and personnel check-in. The system pre-defines a unified data dictionary for each type of data, specifying the field name, type, unit, legal value range, and enumerated state set, and configures field mapping relationships to map the original field names of each subsystem to unified field names.

[0025] The field mapping relationship is determined as follows: extract the field set from the interface definition and historical message examples of each subsystem, merge them according to business semantics and determine a unified field name; when the same semantic field is reported by multiple sources, the source priority is set based on the principle of "authoritative source first". For example, the status of the aircraft position is reported by the aircraft position management system first, the vehicle location is reported by the vehicle terminal or vehicle dispatch system first, and the personnel attendance status is reported by the personnel management system first, so as to reduce the conflict caused by the inconsistency of cross-system standards.

[0026] After completing the field mapping, the system performs cleaning and transformation on each original record. Duplicate record processing employs a "sliding deduplication window" mechanism: based on the business primary key and event time, the system identifies duplicate deliveries within a time range matching the reporting cycle, retaining only the earliest arriving record or the one with the most recent timestamp. The deduplication window is determined by: calculating the reporting cycles of each data source, taking the longest reporting cycle, and then multiplying it by a safety factor to cover network jitter. This safety factor is typically set to 2 to cover the time range where most duplicate deliveries occur.

[0027] Missing fields are filled using a "last valid value backfill" mechanism, but the backfill is not unlimited: the system configures a maximum backfill span for different fields. For example, vehicle location is allowed to be backfilled for a shorter duration to avoid location distortion, while personnel check-in is allowed to be backfilled for a longer duration within the same shift to avoid gaps in attendance status due to delays in check-in reporting. The maximum backfill span is determined by statistically analyzing historical data, taking the typical duration for which the field remains continuous and stable under normal circumstances, and setting an upper limit based on the business's real-time requirements to ensure that backfill can both suppress missing reports and not mask real changes.

[0028] Outlier handling employs an "adaptive threshold based on historical distribution." Continuous numerical fields (such as speed, location jump distance, and operation duration) are grouped by field type and business scenario. The distribution of similar historical data within the same time period, region, or resource type is statistically analyzed, and robust statistics are used to determine outlier boundaries. Specifically, historical samples are first sorted by value from smallest to largest, and the sample in the middle is taken as the typical level. Then, the deviation of the sample from the typical level is used to characterize the fluctuation range. When a new value significantly exceeds this fluctuation range, and this deviation is a rare tail event in the historical data, it is judged as an outlier and removed or marked for review. The historical sample window used to calculate the distribution typically covers at least 7 days, and data from the same hourly period is preferred to reduce the impact of circadian rhythms on the threshold.

[0029] For discrete state fields, the system uses the enumeration set of the data dictionary as the standard. Any state not included in the enumeration set is classified as an "unknown state" and enters the subsequent conflict resolution process instead of being directly used for spatial or job chain determination. After the above cleaning is completed, the system unifies the time field to a millisecond-level timestamp, unifies fields with inconsistent units to the same unit according to unit conversion rules, and standardizes the identifier field, such as unifying the character set, capitalization, and delimiters, to ensure that different ways of writing the same entity in different systems converge to the same standard identifier, thereby forming a standardized business data stream with continuous output.

[0030] Secondly, the generation of the spatial state dataset focuses on "spatial attribution of flight positions" and "reliable determination of gate occupancy." The system pre-configures the spatial boundaries of the airport operating area to determine whether a flight position is within the airport's effective area. If the flight coordinates fall outside the effective area, the system does not discard them directly but corrects them using a "nearest boundary point pullback" method to avoid misjudging a flight as departing due to short-term satellite positioning drift. The maximum allowable pullback distance is obtained from historical positioning drift statistics. Specifically, it calculates the high quantile level of drift distance from a large number of historical trajectories and takes the distance corresponding to that high quantile as the upper limit to cover the vast majority of normal drifts while suppressing significant erroneous positioning.

[0031] After flight coordinates are validated, the system performs spatial mapping: when the airport has precise polygonal boundaries for aircraft stands, point assignment is used to assign flights to specific stand areas; when the airport only has gridded areas, a grid index is used to assign flights to a specific spatial unit. The principle for determining the grid size is: not less than the minimum distinguishable displacement of the vehicle in a scheduling sense, while considering the upper bound of positioning error to ensure that adjacent grids are identifiable in actual scheduling; this value is usually determined by combining the vehicle's minimum turning radius, minimum safety distance, and the statistical upper bound of positioning error. The determination of stand occupancy status adopts a state mapping rule: the original status reported by the stand management system is mapped to a unified occupancy status code, and a "stable confirmation window" mechanism is used to suppress jitter when inconsistencies occur from multiple sources. The stable confirmation window is determined by statistically analyzing the distribution of the shortest stable duration during which the stand status remains unchanged under normal operation, and taking the smaller high-frequency level as the confirmation window, so that the system can respond quickly to state changes without being falsely triggered by momentary jitter. To support the selection of credibility during subsequent fusion, the system also generates source consistency markers: when the same state is consistently supported by multiple sources, the credibility of that state is higher; when there are fewer or contradictory supporting sources, the credibility of that state is lower, and subsequent fusion prioritizes records with higher credibility. Through the above processing, the spatial state dataset can stably represent the key spatial facts of "where the flight is and whether the gate is available" under a unified time axis.

[0032] Secondly, the construction of the operational chain data revolves around "consistent node identification" and "reliable link connection," aiming to organize the progress of special vehicle operations and personnel check-in from scattered events into a traceable and verifiable support process chain. The system pre-configures an operational node dictionary, including node identifiers, node semantics, allowed sets of preceding nodes, and allowed sets of succeeding nodes. This dictionary is derived from airport support process specifications and is fixed during deployment. During the matching process, the system uses the operational node identifier as the primary key to associate vehicle progress messages with personnel check-in logs. However, considering the discrepancies between system clocks, network latency, and personnel check-in delays, the system sets a time tolerance window for matching. This time tolerance window is not determined arbitrarily but is synthesized from three historical upper bound statistics: the maximum deviation level of different systems relative to the reference clock, the maximum jitter level of message transmission latency, and the typical lag level of personnel check-in relative to the start of the operation. These three are obtained through historical reconciliation data, message link monitoring data, and check-in behavior statistics, respectively, and high-quantile statistical values ​​are used to cover abnormal fluctuations.

[0033] When there is more than one candidate matching record, the system selects the most likely one based on the time proximity. When there is no matching record, the system does not disconnect the job chain, but marks the personnel field of that node as unknown and retains the vehicle progress information to avoid the link breaking due to missing personnel data. When connecting links, the system simultaneously refers to the preceding and following constraints of the node dictionary and the actual time order: when a subsequent node meets the allowed successor relationship and its start time is not earlier than a reasonable range of the preceding node, the system connects it to the same job chain. The reasonable range is obtained by statistically analyzing the parallel or overlapping situations of historical jobs. Specifically, the system calculates the overlap duration between nodes of the same type in historical data and takes the high quantile level covering most normal parallel situations as the tolerance threshold, thus allowing parallel jobs that conform to the process to be correctly linked, without mistakenly linking obviously unreasonable time inversions together.

[0034] Subsequently, the workflow data undergoes time correction and multi-dimensional fusion to generate a first operational snapshot that can be directly consumed by the model. Time correction first identifies a unified benchmark time source, typically selecting a core system synchronized with a unified time synchronization system to ensure continuous time consistency. For each non-benchmark data source, the system estimates its fixed deviation relative to the benchmark and its drift trend over time through reconciliation events: a reconciliation event refers to the moment when the same business entity can be simultaneously observed in two systems, such as the record of the same vehicle location or the status switch of the same machine in both systems. The system continuously collects these event pairs, estimates the typical level of deviation using robust statistical methods, and smoothly updates the deviation trend over time to avoid abrupt changes in the correction results due to single anomalies.

[0035] The window length for smooth updates is typically set to cover at least five reporting cycles, ensuring that corrections follow long-term changes without being affected by short-term fluctuations. After time correction, the system aligns the spatial state with the job chain state along a unified timeline and encapsulates them according to preset snapshot time slices. The width of the snapshot time slice needs to meet two conditions simultaneously: first, it must cover the fluctuations in the data source reporting cycles, ensuring that valid data is available in most time slices; second, it must not exceed the minimum response time required by scheduling to prevent slow state updates. This width is determined by first statistically analyzing the reporting cycle distribution of key data sources, taking the high percentile level that covers most reporting rhythms as the lower limit, and then using the requirements of operation control for scheduling response as the upper limit, selecting a stable and usable value between the upper and lower limits. During encapsulation, the system selects the latest spatial state record within each time slice that meets the reliability requirements; the reliability requirement is typically set to the level of "consistent support from at least half of reliable sources" to maintain robustness even when multiple sources are inconsistent. If no record that meets the credibility requirements is found within the time slice, the most recent valid record is allowed to be used according to the maximum backfill span, and a backfill mark is attached for subsequent tracing.

[0036] The projection of the job chain status within a time slice is determined using interval relationships: when the execution time period of a node overlaps with the time slice's coverage interval, the node is considered to be in progress; when a node has finished before the start of the time slice, it is considered to be completed; when a node starts after the end of the time slice, it is considered to have not started. For conflicting fields, the system handles them according to field type: discrete status fields preferentially use a combination of source priority and stable confirmation window rules; continuous numerical fields use a robust "weighted median level" fusion method, with weights given by source consistency flags to avoid unnecessary offsets in the fusion results caused by extreme values. Finally, the system encapsulates key fields such as flight space ownership, gate occupancy status, job node progress, vehicle availability and occupancy status, and personnel arrival and participation status within the same time slice into a first operational status snapshot, and assigns a unique identifier and timestamp index to this snapshot.

[0037] S2 includes acquiring a first snapshot of the operating state, inputting the first snapshot of the operating state into a multi-agent reinforcement learning model to obtain the probability distribution of candidate actions for the shuttle bus, cleaning team, and baggage towing vehicle; performing joint action deduction based on the probability distribution of candidate actions, and updating the policy gradient by exchanging local observation information if there is a spatiotemporal conflict in the deduction result, until a conflict-free joint action sequence is generated; parsing the conflict-free joint action sequence, converting the actions in the conflict-free joint action sequence into scheduling operation codes and associating them with the agent identifiers to generate a first set of decision instructions for resource scheduling.

[0038] In this implementation, firstly, after obtaining the first snapshot of the operating state from step S1, it is converted into an input representation that can be processed by the multi-agent reinforcement learning model. Specifically, the system encodes the fields in the snapshot according to a preset feature template: discrete fields (e.g., gate occupancy status, job node status, resource availability status) are mapped with fixed numbers and expanded into sparse features; continuous fields (e.g., vehicle location, estimated arrival time, remaining job duration) are standardized in terms of dimensions and pruned in terms of range, with the upper and lower bounds of pruning determined by the high quantile statistics of historical extreme values ​​to avoid amplifying model output fluctuations due to abnormal noise; aggregate fields (e.g., multiple flights, multiple gates, multiple vehicles, and multiple personnel statuses within the same time slice) are aligned in length according to a fixed sorting rule, with the sorting rule prioritizing those with higher relevance to the current conflict risk, such as prioritizing those closer to the current resource or with more urgent transit time, and the portion exceeding the length limit is processed in the next decision cycle.

[0039] The upper limit of the set length is determined by two factors: the upper bound of the airport's operational scale and the budget for computational latency in a single inference iteration. Typically, the lower limit is determined by statistically analyzing the high quantile of the number of similar objects during peak periods, and the upper limit is determined by a safe proportion where the inference latency does not exceed the decision-making cycle. This ensures coverage of most real-world operating conditions while maintaining real-time operation. After encoding, the first operational state snapshot becomes the unified state input within a single decision-making cycle.

[0040] After receiving a unified state input, the multi-agent reinforcement learning model outputs candidate action probability distributions for shuttle buses, cleaning teams, and baggage tractors, respectively. In practice, the system defines a set of candidate actions for each resource type and enumerates and encodes these sets to ensure that the meaning of each action can be strictly interpreted as a business operation. Shuttle bus candidate actions typically cover types such as "go to a designated waiting area, go to a designated boarding gate or aircraft stand area, undertake a designated flight shuttle task, exit the current area and transfer, and remain waiting"; cleaning team candidate actions typically cover types such as "enter a designated aircraft stand for operation, queue for entry, transfer to the next task point, and remain on standby"; and baggage tractor candidate actions typically cover types such as "tow to a designated sorting or loading / unloading location, undertake a designated flight baggage transfer node, queue for loading / unloading windows, and transfer to the next work point".

[0041] The size of the candidate action set is not arbitrarily expanded, but controlled by the action granularity parameter: the action granularity is determined by the smallest executable scheduling unit on-site. Too fine a granularity will lead to an excessively large action space and high inference overhead, while too coarse a granularity will make it unworkable on-site. Therefore, the action granularity is usually determined by combining the opcode granularity of the existing rigid business system, so that the model actions correspond one-to-one with the business system instructions or can be transformed through finite mapping. In one inference, the model first calculates the relative preference score for each candidate action, and then normalizes the score into a probability distribution so that the sum of the probabilities of each candidate action is 1.

[0042] The normalization process employs an exponential re-normalization method to ensure that score differences are reasonably converted into probability differences. The smoothness involved is controlled by a "probability mildness parameter," which is used to avoid an overly sharp probability distribution leading to insufficient exploration or an overly flat distribution leading to unfocused decision-making. This parameter is determined as follows: during offline training, historical replay data is used to evaluate the "concentration of correct actions in the same state" and the "allowable pluralism in the environment." When multiple feasible solutions are allowed in the environment, the mildness is appropriately increased to retain alternatives; when the environment constraints are strict, the mildness is decreased to enhance determinism. After going live, the model can be fine-tuned based on its performance without changing the model structure, but the adjustment range is limited by a safety strategy to avoid introducing sudden changes.

[0043] After obtaining the probability distribution of candidate actions for each resource type, the joint action simulation phase begins to ensure that action combinations of multiple resources can be simultaneously valid in time and space within the same decision-making cycle. The implementation of joint action simulation is not simply a matter of piecing together the highest-probability actions of each resource; rather, it involves first generating a candidate set of joint actions and then verifying their feasibility one by one. The generation of the candidate set of joint actions is typically constrained by two key parameters: the upper limit on the number of candidate actions for each resource type participating in the combination, and the maximum number of combinations in the joint candidate set. The former limits each resource to only include a few actions with relatively high probabilities in the combination, avoiding an explosion of combinations due to low-probability actions; its value is determined statistically by the long-tailedness of the probability distribution, typically selecting the minimum number of actions that can cover a cumulative probability reaching a preset percentage. This preset percentage is determined based on the business risk appetite; the more conservative the risk appetite, the higher the percentage. The latter limits the overall combination scale, and its value is determined by the simulation calculation budget, ensuring that the simulation is completed and instructions are output within a single decision-making cycle.

[0044] After the combined actions are generated, the system performs a spatiotemporal simulation for each candidate joint action. The simulation calculation process includes: calculating the travel path and travel time based on the current location of the resource and the target location. The path calculation can use the shortest feasible path from the airport road network map or the output of the existing vehicle navigation module; the travel time is determined by the path length, the speed limit of the resource type, and the speed limit zone rules. The speed limit comes from vehicle specifications or airport control rules, and the speed limit rules come from the airport operation rule configuration; after arrival, the duration of the operation is estimated based on the standard working hours of the operation node. The standard working hours are obtained from historical operation data statistics, usually stratified by the same gate type, same flight type, and same shift configuration, and a robust center value is taken to reduce the impact of occasional delays on the estimation. During the simulation, the system also incorporates gate occupancy status, the sequential dependency of operation nodes, and resource mutual exclusion rules into the verification. For example, if the cleaning team has not completed its work, the baggage tractor cannot enter a specific operation window, and the shuttle bus and the tractor cannot occupy the mutual exclusion zone of the same road segment at the same time. The above mutual exclusion zones and dependencies are fixed by business rules or generated by the airport road network topology configuration and are one-time configuration content before going live.

[0045] When spatiotemporal conflicts are detected during joint action simulation, the system needs to localize these conflicts to drive subsequent strategy updates. Conflict identification is based on deterministic judgment using "time slice discretization" and "spatial unit affiliation": The system first divides the simulation timeline into continuous time slices with a fixed granularity. The granularity of the time slice is determined by two constraints: the minimum reliable frequency of vehicle position updates and the minimum reaction step size allowed by on-site dispatch. Typically, the larger of the "typical cycle for stable and updatable vehicle position updates" and the "minimum step size for real-time closed-loop dispatching" is chosen to avoid instability caused by excessively fine time slices. Subsequently, the system projects resource trajectories onto the time slices to obtain the spatial unit and occupied interval of the resource within each time slice, and determines the conflict type accordingly: when two resources occupy the same mutually exclusive spatial unit in the same time slice, it is determined to be a spatial conflict; when the same resource is scheduled to execute two mutually exclusive tasks in the same time slice, it is determined to be a resource conflict; when a subsequent task node is scheduled to execute before the preceding task node is completed, it is determined to be a process conflict.

[0046] To avoid misjudgments caused by boundary jitter, a safety buffer parameter is introduced. Spatial conflicts are handled using "safety interval expansion," and temporal conflicts are handled using "minimum interval buffer." The value of the safety buffer is not arbitrarily set, but is determined jointly by safety specifications and historical operational statistics: the spatial buffer is usually taken as the high quantile level of the sum of the minimum safe distance of the vehicle and the upper bound of the positioning error; the temporal buffer is usually taken as the typical shortest interval of the handover of similar operations, taking into account the upper bound of network and execution delays, thus ensuring that there is still a margin when the simulation is conflict-free and the field execution is carried out.

[0047] After a conflict is located, a closed loop of local observation information exchange and policy update is initiated. The specific implementation of local observation information exchange is as follows: the system extracts the minimum necessary information set from the conflict-related agents and shares it through instruction coordination channels or internal model messaging mechanisms, enabling conflict-related agents to obtain each other's key constraints. This minimum necessary information set typically includes: the current location or state of the conflicting resource, its candidate action list and probability ranking of each action, its estimated arrival and start times, its estimated operation duration, and the context of related flights, gates, and operation nodes. The scope of the minimum necessary information set is determined by both information exchange bandwidth and privacy isolation requirements: too little information cannot resolve the conflict, while too much information increases communication overhead and introduces irrelevant noise. Therefore, the relationship between "information set size and conflict resolution success rate" is usually evaluated through offline playback to select the minimum set size that allows the success rate to reach a preset threshold. The preset threshold is determined by setting a target based on the airport's requirements for scheduling stability, such as requiring a conflict-free solution to be obtained within a limited number of iterations in most decision cycles during peak periods; this target is also constrained by the computational budget, and the threshold is ultimately solidified based on the principle of "achieving the target success rate within a specified maximum number of iterations."

[0048] The strategy update is implemented by "penalizing conflicting actions and relatively enhancing alternative feasible actions" to correct the action selection tendency in the current decision-making cycle, making the next round of simulation more likely to generate feasible joint actions. The specific calculation process is as follows: the system transforms the conflict identification results into quantifiable feedback signals, including conflict severity, conflict duration, the amount of resources involved in the conflict, and the impact on key business indicators (e.g., the potential disruption to the shortest transit time). The quantification standard for conflict severity is determined by the business risk level configuration: for example, conflicts entering security-sensitive areas have a higher weight than conflicts on general road sections, and conflicts affecting critical flight nodes have a higher weight than conflicts affecting non-critical nodes; these weights are derived from the operational rules and security priority configuration.

[0049] Upon receiving feedback signals, the system performs one or more policy corrections on the agents involved in the conflict, reducing the probability of actions causing the conflict and increasing the probability of feasible alternative actions. The correction step size is controlled by the "online correction amplitude parameter," which follows the principle of "not introducing drastic distribution changes within a single cycle": its upper limit is determined by historical online stability assessment, ensuring the continuity of action distribution after correction and avoiding instruction oscillations; its lower limit is determined by conflict resolution efficiency requirements, ensuring sufficient amplitude to escape conflict combinations within the maximum allowed number of iterations. The maximum allowed number of iterations is also a configurable parameter, determined by the decision cycle duration and the computation time of the deduction: the system ensures that the "exchange-correction-re-deduction" closed loop is completed within this number of iterations without exceeding the time budget of the decision cycle, and adopts a degradation strategy when convergence within the budget, prioritizing joint action candidates with the fewest conflicts and the least impact on key constraints to ensure continuous output of instructions without blocking operation.

[0050] Once the closed-loop iteration generates a conflict-free sequence of joint actions, the system enters the instruction parsing and scheduling operation code generation stage. This stage is based on a "mapping table from action to business operation": the system predefines the correspondence between action semantics and rigid business system operation codes for each type of resource, and defines necessary parameter fields for each operation code, such as resource identifier, target task identifier, target location or machine area, expected arrival time or execution time window, job node identifier, and necessary safety constraint flags.

[0051] During parsing, the system reads each action in the joint action sequence, looks up the corresponding opcode in a table, and extracts the time window, target location, and associated tasks of the action from the deduction results, filling them into the parameters of the issueable instruction. To ensure traceability, the agent identifier that issued the action is written into each instruction. The generation rules for the agent identifier are fixed during model deployment and typically correspond one-to-one with resource type and resource grouping, facilitating subsequent compliance verification, instruction difference comparison, and policy offset attribution. At this point, the system forms the first set of decision instructions for resource scheduling.

[0052] S3 includes obtaining a first set of decision instructions containing scheduling information, mapping the first set of decision instructions to preset shortest flight turnaround time rules and gate conflict safety boundaries to construct a compliance verification matrix; extracting unique identifiers of rule-violating instructions based on the compliance verification matrix to generate a list of violation instruction indexes; inputting the list of violation instruction indexes into the rigid business system rule engine to generate rigid replacement instructions that comply with safety specifications; and replacing the corresponding instructions in the first set of decision instructions with rigid replacement instructions to generate a first set of corrected instructions.

[0053] In this implementation, a first set of decision instructions containing scheduling information is first obtained. Each instruction is then broken down and standardized to ensure it can be uniformly mapped to the rule validation dimension. Instruction elements include at least a unique instruction identifier, resource type and resource identifier, associated flight identifier, associated work node identifier, target gate or target area, planned start time, planned end time or expected duration, and necessary priority markers. These elements are obtained as follows: if the model output already contains all fields, they are read directly; if the model output only contains action semantics, the target gate, target area, work node type, and standard working hours are supplemented using a mapping table from action to operational code. The standard working hours are derived from stratified statistical results of historical support data, with the stratification dimension including at least gate type, flight type, work category, and shift configuration to avoid bias caused by using a single average value.

[0054] To ensure comparability of time fields, the system uniformly converts all times into timestamps of the same precision, and calculates the planned end time for instructions with missing end times based on "planned start time plus estimated duration". The principle for determining the estimated duration is to prioritize the remaining man-hour estimate reported in real time by the field system, followed by the historical robust center value, and allows marking the source category of the duration on the instruction for subsequent offset analysis.

[0055] After standardizing the instruction elements, the system maps the first set of decision instructions to preset shortest turnaround time rules for flights and gate conflict safety boundaries to construct a compliance verification matrix. Mapping transforms "instructions" into "occupancy and time window expressions that can be calculated by rules." The structure of the compliance verification matrix is ​​determined by preset dimensions, including at least flight, gate, time, and resource dimensions. Each matrix cell records "the occupancy or constraint impact of a certain instruction on a certain flight, gate, or resource within a certain time range." The flight dimension carries the associated information of the shortest turnaround time rule; the gate dimension carries gate occupancy and safety boundary information; the time dimension carries the execution time window of the instruction; and the resource dimension carries information on mutually exclusive resource execution and resource capacity limitations. The dimension set is determined as follows: based on the hard constraint list of airport operations, rules are first classified according to constraint objects, and then the constraint objects are abstracted into dimensions. Flights and gates are the core constraint objects of the scheduling system, time windows are the basic objects that all rules depend on, and the resource dimension covers basic restrictions such as the inability to execute the same resource in parallel, thus it is fixed as a basic dimension.

[0056] The time granularity of the compliance verification matrix is ​​not arbitrarily set, but is determined by two factors: the minimum perceptible change step size of instruction execution and the reliable frequency of field system location and status updates. The time granularity is typically the larger of the "minimum period that the update frequency can stably cover" and the "minimum step size allowed by the scheduling loop," to avoid overly fine granularity leading to status jitter or overly coarse granularity leading to missed detection boundary conflicts. To prevent misjudgments due to boundary errors, the system introduces a safety buffer for the time window of each instruction during mapping. The value of the safety buffer is jointly determined by the upper bound of execution delay, the upper bound of network transmission jitter, and the arrival time uncertainty caused by device positioning errors. These three are obtained through historical link monitoring, terminal receipt statistics, and positioning error statistics, respectively, and high-quantile statistical values ​​are used to cover abnormal fluctuations, thereby ensuring that instructions judged to be compliant still have a safety margin when executed on-site.

[0057] After the compliance verification matrix is ​​constructed, the system iterates through the matrix according to preset rules to extract instructions that violate the rules. Regarding the shortest turnaround time rule for flights, the system does not only compare individual instructions, but organizes all operation node instructions related to the same flight into a support chain in chronological order, and calculates whether the chain can be completed within the time frame of the flight's arrival and planned departure. Specifically, the system first reads the planned arrival time and planned departure time from the flight's basic data and calculates the available turnaround window from these two values; then, it retrieves the set of instructions related to the flight from the matrix, determines the priority relationship of nodes according to the operation node type (the priority relationship comes from a preset operation node dictionary); finally, it calculates the node occupancy interval based on the planned start time and planned end time of each instruction, and determines whether there is a situation where "the start time of a subsequent node is earlier than the end time of a preceding node minus the allowed overlap threshold".

[0058] The overlap threshold is used to tolerate parallel operations permitted by regulations. For example, partial cleaning and partial loading / unloading can be done in parallel under specific conditions. The threshold value is determined jointly based on airport support process specifications and historical parallel sample statistics: for node pairs that are explicitly allowed to run in parallel according to the specifications, the minimum interval requirement given in the specifications is adopted; for node pairs that are not explicitly allowed to run in parallel according to historical data, the threshold is set at the high percentile level of historical parallel overlap time to cover common parallel modes while avoiding misjudgment of abnormal overlap. After completing the node-level consistency check, the system further determines whether the total completion time of the chain exceeds the fixed buffer reserved by subtracting the minimum turnaround time requirement from the planned departure time. The value of the minimum turnaround time requirement is determined based on the airport's support standard configuration for different aircraft types and flight classes. It is usually based on the minimum support time published by the operations department, and additional buffers are set according to historical delay tolerance to ensure that the minimum standard can still be met during peak periods. If any judgment condition is not met, the system will determine that there is a conflict in the shortest transit time rule and mark the instruction that caused the conflict as a violation candidate. In order to avoid misjudging a single early or late non-critical node as a violation, the system will also filter based on the set of critical nodes. The set of critical nodes consists of nodes that have a decisive impact on departure in the guarantee process and is configured in a fixed manner according to business specifications.

[0059] To address security boundary conflicts at workstations, the system uses workstation occupancy and entry conditions as its core for verification. The system first extracts the occupancy status of each workstation in each time slice from the matrix, and then projects all instructions pointing to that workstation onto its occupancy trajectory according to time windows. Conflict determination includes at least three scenarios: The first scenario is when two or more instructions within the same time slice cause different resources to enter the same mutually exclusive area, and this area is marked as non-mergeable in the security configuration; the second scenario is when an instruction requests a resource to enter the workstation area, but the workstation is occupied or locked, and the occupant or the reason for locking is not allowed to be shared; the third scenario is when the time window for an instruction to enter overlaps with the workstation's security boundary time window, for example, prohibiting vehicles from entering within a specified security check window.

[0060] The mutual exclusion zone is determined by the airport road network and aircraft stand geometry, typically through offline modeling and solidification based on stand adjacency, taxiway intersections, and vehicle access restriction zones. The safety boundary time window is determined by operational safety regulations and on-site control strategies, and different boundary strategies can be activated for different time periods, weather conditions, or operational levels. To reduce misjudgments caused by positioning errors and execution delays, the system employs spatial safety boundary expansion for the aircraft stand area. The expansion range is jointly determined by the minimum safe distance for vehicles, the upper bound of the positioning error statistics, and the upper bound of the driver's execution deviation. The upper bound of the execution deviation is obtained through historical statistics of the difference between the terminal's acknowledgment time and the actual arrival time, and similarly uses a high quantile value to cover abnormal fluctuations.

[0061] After completing the above verification, the system extracts the unique identifiers of the violating rules based on the compliance verification matrix and generates a list of violating instructions. This list is not a simple compilation, but rather each violating instruction is accompanied by a violation type label and a trigger location label, enabling the subsequent rule engine to generate more suitable alternatives based on the cause. The rules for generating unique identifiers are fixed during the instruction generation stage, and are typically derived from a combination of information such as resource identifier, target object identifier, planned start time, and action type, ensuring that they are not repeated within the same decision cycle. When there is a need for cross-cycle traceability, the system also embeds a decision cycle marker into the unique identifier to avoid identifier conflicts for the same action in different cycles.

[0062] The classification criteria for violation type labels are consistent with the aforementioned verification rules, including at least shortest transit time conflicts, station occupancy conflicts, safety boundary conflicts, and resource mutual exclusion conflicts; the trigger location label is used to indicate the station number, area number, or job node number where the conflict occurs, so that the rule engine can directly locate the constraint object.

[0063] The list of violation instructions is input into the rigid business system's rule engine to generate rigid alternative instructions that comply with safety specifications. The rule engine's input includes at least the original elements of the violation instruction, violation type tags, trigger location tags, and context such as current gate occupancy status, flight schedule, and resource availability. The rule engine's process of generating alternative instructions follows the principle of "maintaining consistency with objectives and minimizing conflict factors," meaning that the flights and operational nodes served by the original instruction are kept unchanged as much as possible, with only necessary adjustments made to the target gate, execution time window, resource allocation, or route selection. In practice, the rule engine first selects an alternative strategy based on the type of violation: For shortest turnaround time conflicts, the priority strategy is to move the time window forward or adjust it in parallel for non-critical nodes, and compress the waiting interval without breaking process dependencies; if compression is not possible, alternative resources or adjacent available gates are selected to reduce turnaround time; for gate occupancy or security boundary conflicts, the priority strategy is to adjust the entry time window to avoid gate locking windows or security check windows, while ensuring that the adjusted time window still falls within the flight guarantee allowable range; if the gate itself is unavailable, a gate that meets the requirements of similar aircraft type compatibility, adjacency restrictions, and passage restrictions is selected from the candidate gate set as an alternative.

[0064] The candidate aircraft stand set is generated based on the stand matching table, stand adjacency constraint table, and area access rule table. These tables are fixed by the airport's basic data and operational rule configuration. When multiple feasible solutions exist in the candidate set, the rule engine selects the optimal alternative based on a preset priority ranking. The priority is determined by combining operational goal configuration, such as prioritizing critical flights, minimizing cross-regional movement, and reducing the cascading impact on other flights. To avoid introducing new conflicts with the alternative instructions, the rule engine performs a quick rule check with the same caliber again before outputting the alternative instructions. If it still does not meet the requirements, it iterative selection is performed within the candidate set. The upper limit of the number of iterations is determined by the time budget of the current decision cycle and is set to ensure continuous system output, avoiding blocking instruction issuance in extreme cases.

[0065] Finally, rigid substitution instructions are used to replace the corresponding instructions in the first decision instruction set to generate the first revised instruction set. The substitution process locates each original instruction entry in the violation instruction index list and replaces its entire instruction element with the elements of the substitution instruction, while preserving traceable relationships. These traceable relationships include at least the original instruction's unique identifier, the substitution instruction's unique identifier, the violation type label, and the substitution source marker. The substitution source marker indicates that the substitution was generated by the rule engine rather than the model and is used for subsequent difference comparisons and policy offset records. To avoid new resource mutual exclusion issues within the instruction set caused by the substitution, a set-level consistency check is performed on the first revised instruction set after all substitutions are completed. The check includes at least whether the same resource is assigned to multiple tasks within the same time window, whether the same machine position is occupied by multiple resources within the same time window, and whether the shortest transit time chain conflict is triggered again. If new conflicts caused by the substitution are found, a second correction by the rule engine is triggered based on the principle of "substitution instructions prioritizing safety and hard constraints."

[0066] S4 includes obtaining a first set of decision instructions and a first set of corrected instructions; extracting a list of instruction difference data and retrieving rule conflict types based on the list of instruction difference data to generate a difference dataset with cause labels; retrieving a first running state snapshot based on the difference dataset with cause labels; extracting associated resource states and performing a structured association with the first running state snapshot; parsing the structured association dataset to construct an agent behavior trajectory; mapping rule conflict types and associated resource states to the agent behavior trajectory to form a first policy offset record containing the agent behavior trajectory and the rule violation context.

[0067] In this implementation, firstly, a first set of decision instructions and a first set of corrected instructions are simultaneously acquired and aligned using the instruction's unique identifier as the primary key. The instruction's unique identifier is permanently written during the instruction generation stage and typically uniquely identifies "a scheduling action of a resource on a target object" within the same decision cycle, thus avoiding ambiguity caused by relying solely on resource identifiers or target positions. During alignment, the system establishes two types of index structures: one is a fast retrieval index using the unique identifier as the key, used to complete a one-to-one match between the two sets; the other is an auxiliary index using the resource identifier and time window as the key, used to identify special cases such as "cancel, split, merge," which cannot be directly matched by the unique identifier. The time window tolerance used in the auxiliary index is not arbitrarily set, but is jointly determined by the minimum precision of the instruction time field, the upper bound of the terminal receipt delay, and the upper bound of the system's internal queue forwarding delay. The value is determined by statistically analyzing historical receipt data and selecting a high quantile level that covers the vast majority of normal delays, ensuring that the same action can still be identified as the same type of action in different sets despite slight timestamp offsets.

[0068] After alignment, the system performs field-level comparisons on each matched instruction, generating a list of instruction difference data. The comparison field set is predefined by a "key field set affecting compliance and execution effectiveness," and includes at least resource type, resource identifier, target flight or operation node, target gate or target area, operation type, planned start time, planned end time or duration, priority, and constraint flags. The principle for determining the field set is to include only fields that will change resource usage, gate usage, time window usage, or process dependencies, avoiding the inclusion of purely display fields or log fields that could lead to noise inflation. During difference identification, discrete fields are judged based on "inequality," while time fields are judged based on "exceeding the time error tolerance threshold." This threshold is determined by a comprehensive upper bound of time accuracy, link jitter, and terminal execution deviation, aiming to filter out minor jitter that does not affect actual execution. For unmatched instructions, the system marks them as cancellation-type differences if "the model set exists but the correction set is missing", and as addition-substitution-type differences if "the correction set exists but the model set is missing". It also records the adjacency relationship between the instruction and instructions with the same resource, the same target object, and adjacent time windows, so as to facilitate the restoration of decision intent when connecting trajectories in the future.

[0069] After the discrepancy data list is formed, the cause of each discrepancy record is attributed, generating a discrepancy dataset with cause tags. Cause attribution is not based on manual interpretation, but rather on retrieving rule conflict types. Specifically, in step S3, the system has already generated a list of violation instruction indexes and written a rule conflict type and trigger location tag for each violation instruction. Therefore, the system uses the instruction's unique identifier as the search key to match the discrepancy record with the violation instruction index list. If a match is found, the discrepancy cause is marked as the corresponding conflict type, such as a flight minimum turnaround time conflict, gate occupancy conflict, safety boundary conflict, or resource mutual exclusion conflict, and the trigger location tag is also written into the discrepancy record to indicate the gate, area, or operational node where the conflict occurred.

[0070] If a discrepancy does occur but is not found in the violation index list, the system further categorizes the cause based on the source of the correction instruction: when the source is rule engine reordering, the cause is marked as rule reordering; when the source is a system degradation strategy, the cause is marked as degradation adjustment. The triggering conditions for rule reordering and degradation adjustment are pre-defined in the system strategy. For example, degradation may be enabled when there is insufficient computing budget, some states are unreliable, or input is missing, or suboptimal reordering may be performed on non-critical tasks under hard constraints. To avoid fragmentation of causes due to multiple discrepancy records for the same instruction, the system aggregates discrepancies for the same instruction within the same decision-making cycle. The aggregation rule is constrained by "cause priority," with security-related causes taking precedence over business efficiency causes, and hard constraint conflict causes taking precedence over reordering causes. This ensures that the final cause label reflects the fundamental triggering factor of the discrepancy.

[0071] After completing the cause labeling, the system backtracks to the first operational state snapshot based on the discrepancy dataset with the cause labels, and extracts the associated resource states and performs a structured association with the snapshot. Snapshot location uses a time index: the system first locates the corresponding first operational state snapshot based on the decision cycle timestamp to which the discrepancy record belongs; when the instruction execution time window spans multiple time slices, the system prioritizes the time slice that best explains the discrepancy trigger. This selection is not subjective judgment, but determined based on the trigger location label and conflict type: if it is a machine position occupancy or safety boundary conflict, the time slice where the machine position status conflict occurs is prioritized; if it is a shortest transit time conflict, the time slice near the planned start time of the critical operation node is prioritized; if it is a resource mutual exclusion conflict, the time slice where the same resource is occupied in parallel is prioritized. To prevent a single time slice from failing to reflect the cause and effect, the system simultaneously takes adjacent time slices as context supplements according to a preset extended window. The value of this extended window is determined by the state update cycle and the conflict propagation speed, typically covering at least two state updates to ensure that state changes before the conflict trigger and correction results after the conflict trigger are captured.

[0072] Subsequently, based on the resource identifier, target aircraft position or target area, and target flight or operation node identifier in the difference record, the associated resource status is extracted from the snapshot and its supplementary time slice. The associated resource status includes at least the resource's current location or area affiliation, resource availability status, current task occupancy and remaining duration, target aircraft position occupancy status and locking reason, target flight support node status, adjacent resource occupancy information with mutual exclusion relationships, and necessary access restriction markers. The resource mutual exclusion relationships and access restriction markers come from the preset mutual exclusion area configuration and road network topology configuration, which are pre-deployed data used to ensure that "what hard constraints existed at that time" can be restored in the offset record. If a field is missing in the snapshot, the system allows it to be filled in by the "most recent valid value" in the extended window, but the filling must be marked with a filling mark and the filling source to ensure that the actual observation and inferred filling can be distinguished during subsequent statistics, avoiding misleading the judgment of offset trends.

[0073] After completing the snapshot and resource status extraction, the difference records, cause markers, associated resource statuses, and the first running state snapshot are structurally associated and written to the associated storage. The structured association uses a unified set of association keys, which includes at least the decision cycle identifier, instruction unique identifier, snapshot identifier, resource identifier, flight identifier, and gate identifier. The principle for determining the association key set is that it should be able to achieve unique location during subsequent lookups in any dimension and support multi-dimensional aggregation, such as aggregation by agent, aggregation by gate scenario, aggregation by flight type, or aggregation by time series.

[0074] To avoid the same difference being counted repeatedly in multiple time-slice contexts, a primary time-slice tag and an auxiliary time-slice tag are written for each difference record upon data entry. The statistical definition explicitly states that the primary time-slice is used as the counting basis, while the auxiliary time-slice is only used for contextual interpretation. Simultaneously, the system standardizes the encoding of difference fields. For example, time window differences are categorized as shifting forward, shifting backward, shortening, or extending; target gate differences are categorized as same-zone replacement or cross-zone replacement; and task differences are categorized as node replacement, flight replacement, or cancellation. This facilitates subsequent trend analysis by category. The category classification rules are fixed and driven by business meaning, avoiding caliber drift caused by subjective adjustments by operations personnel.

[0075] After the structured association dataset is formed, it is parsed to construct agent behavior trajectories, and rule conflict types and associated resource states are mapped to the trajectories, ultimately forming the first policy offset record. Trajectory construction is based on agent identifiers. Agent identifiers are written when the model outputs instructions and have a stable correspondence with resource types or resource groups, enabling the system to aggregate instruction outputs to specific agents. The system first groups all difference records by agent identifier, and then sorts the records of the same agent according to the decision cycle time order, forming a time series trajectory. To avoid only recording "what was corrected" without understanding "what was originally intended to be done," the trajectory simultaneously writes the original and corrected instruction elements of the model, along with the difference category, cause label, trigger location label, and snapshot context summary.

[0076] The context summary is not arbitrarily extracted, but rather selects the most relevant fields based on the conflict type: for example, for gate conflicts, the system prioritizes retaining the reasons for gate occupancy and locking, adjacent gate occupancy, and mutual exclusion zone markers; for shortest transit time conflicts, the system prioritizes retaining the status of critical operation nodes, planned time windows, and remaining margin information; for resource mutual exclusion conflicts, the system prioritizes retaining the current task occupancy of resources, queue length, and availability of adjacent resources. To ensure that the trajectory can be used for quantitative statistics, the system calculates the offset intensity level for each offset event. The offset intensity level is pre-configured based on the degree of business impact of the difference fields: for example, changes in target gate positions are generally higher than time window fine-tuning, cross-area replacements are generally higher than intra-area replacements, and delays in critical nodes are generally higher than delays in non-critical nodes; this level configuration comes from operational objectives and risk level requirements and is a configurable parameter that requires approval. Finally, the system encapsulates the agent trajectory and offset event context into a first-strategy offset record and writes it to storage with a time-series index. The index includes at least the agent identifier, decision cycle, conflict type, and resource type, facilitating subsequent statistics on offset frequency by time window, proportion by conflict type, and sensitive scenarios by resource type, and providing data basis for the gradual import of control parameters.

[0077] S5 includes obtaining the historical offset sequence and offset frequency statistics from the progressive import control parameters, extracting shuttle bus scheduling task feature data within the current operating cycle; generating a scheduling instruction stability index based on the historical offset sequence, and converting the offset frequency statistics into a business risk level value; inputting the scheduling instruction stability index and the business risk level value into the progressive import control model to obtain the initial takeover ratio coefficient; using the model decision confidence to correct the initial takeover ratio coefficient to generate a dynamic weight factor, and determining the proportion of shuttle bus scheduling functions taken over by the model decision within the current operating cycle based on the dynamic weight factor.

[0078] In this embodiment, the historical offset sequence and offset frequency statistics of the progressive import control parameters are first obtained, and the shuttle bus scheduling task feature data in the current operating cycle is extracted. The historical offset sequence is derived from the summary result of the policy offset record formed in step S4 in the time series index, and its content includes at least the offset event list of shuttle bus related instructions in each decision cycle, the offset reason category, the offset intensity level, and the context label of the offset trigger time.

[0079] The offset frequency statistics are a table summarizing offset events periodically within a preset statistical window. It includes at least the number of offset occurrences, the percentage of instructions involved in the offset, the percentage of causes of the offset, the offset intensity distribution, and the length of consecutive offset cycles. The length of the statistical window is not arbitrary but determined by two constraints: firstly, the volatility of airport operations within a short period, where a window that is too short is easily affected by occasional events, leading to drastic fluctuations in the takeover ratio; secondly, the need for timely reflection of model learning and rule adaptation effects during gradual implementation, where a window that is too long will cause the takeover ratio to react slowly to improvements or degradation. A common practice is to use "covering at least 7 days of similar operating periods" as the lower limit and "not exceeding the response lag period allowed by the operating strategy" as the upper limit, selecting a fixed window between these limits, and using different window configurations for peak and off-peak periods to balance stability and sensitivity. The characteristic data of shuttle bus scheduling tasks in the current operating cycle are extracted from the first operating state snapshot, vehicle task queue and flight event stream. The extraction process uses the current cycle time index as the anchor point and reads elements such as the number of flights to be guaranteed, the number of gates involved, the number of available shuttle buses, the average travel time of shuttle buses, the length of the shuttle bus task queue, the proportion of critical flights, the marking of high-incidence areas of gate conflicts, and the marking of road control or traffic restriction.

[0080] To ensure comparability of features across different periods, these elements are normalized according to predefined criteria: count-type elements are scaled using the upper limit of the airport's historical peak statistics as the scale; duration-type elements are truncated using typical values ​​and fluctuation ranges from historical stratified statistics and then mapped to a fixed range; and labeled elements are encoded using an enumeration method. The aforementioned scales, typical values, and fluctuation ranges are all fixed through historical data statistics before the system goes live and are allowed to be updated in versions according to seasonal or operational strategy changes to maintain consistency.

[0081] After obtaining historical offsets and current task characteristics, the system generates a scheduling instruction stability index based on the historical offset sequence and converts the offset frequency statistics into a business risk level value. The scheduling instruction stability index characterizes the degree to which model shuttle bus instructions can remain uncorrected under rule constraints. Its calculation follows the principle of "fewer offsets mean greater stability, lighter offsets mean greater stability, and safety-related offsets have higher weights." Specifically, the system first filters shuttle bus-related offset events within the statistics window and calculates the proportion of offset events to the total number of shuttle bus instructions. Simultaneously, it weights and summarizes the offset intensity levels. The intensity level classification has been solidified according to business impact in step S4; for example, cross-regional replacement and critical node delays are considered higher intensity, while time window fine-tuning is considered lower intensity. To prioritize safety, the system assigns different weights to different conflict causes. The weights are determined based on the operational safety level and hard constraint priority: safety boundary conflicts have the highest weight, followed by machine slot occupancy conflicts, then resource mutual exclusion conflicts, and rule rearrangement or downgrade adjustments have the lowest weight.

[0082] The weight values ​​are not subjectively set, but rather converted into numerical proportions based on the risk level stratification requirements of the operations management department. These weights are then calibrated using the severity of the impact of different conflicts on operations as analyzed in historical event reviews to ensure they reflect the true risks. Subsequently, the system merges the "offset ratio," "weighted intensity summary," "cause weighted summary," and "continuous offset cycle penalty item" to form the stability evaluation result. The continuous offset cycle penalty item is used to suppress repeated triggering of similar corrections within a short period. Its value is determined by adding a penalty when the offset ratio exceeds a preset threshold for several consecutive cycles. The threshold is determined by statistically analyzing the distribution of offset ratios in historical data and selecting a dividing point that distinguishes between "normal fluctuations" and "significant anomalies." Typically, the high quantile level of the offset ratio is used as the threshold to ensure that the penalty is triggered only when the offset is significantly abnormal, avoiding oversensitivity to normal small fluctuations. Finally, the stability index is mapped to a fixed range with directional constraints, so that a larger value indicates greater stability, facilitating its use in subsequent control models.

[0083] The business risk level value is used to characterize the business risk background within the current operating cycle where "even if the model is stable, it is not advisable to take over significantly." Its generation is based on the statistical results of offset frequency and is stratified in conjunction with the task characteristics of the current cycle. In specific implementation, the system first forms a set of risk assessment elements, including the proportion of critical flights, the degree of gate congestion, the shortage of shuttle buses, the activity level of high-conflict areas, and whether the offset of neighboring cycles has increased. The degree of gate congestion is calculated based on the gate occupancy status and the expected release time. The system statistically analyzes the proportion of gates that are occupied or locked in the current cycle and compares it with the typical level of similar historical periods. The shortage of shuttle buses is determined based on the ratio of available vehicles to the number of tasks to be executed, taking into account the reduction in the effective available number due to vehicles being occupied en route.

[0084] The aforementioned ratios and proportion thresholds are derived from quantile statistics of historical operational data and are combined with operational control strategies to set stratified boundaries. For example, risks are divided into several levels, each corresponding to a monotonically increasing value. The principle for determining the stratified boundaries is: in historical samples, high-risk levels should cover known peak congestion and abnormal operational scenarios, low-risk levels should cover stable normal operational scenarios, and sufficient differences should be maintained between adjacent levels to avoid frequent level jumps. Finally, the system maps the risk assessment elements of the current period to the corresponding risk level and outputs the numerical representation of that level. The larger the value, the higher the risk, which serves as the negative input to the control model.

[0085] After obtaining the stability index and business risk level of the dispatch instructions, the system inputs both into the incremental control model to obtain the initial takeover ratio coefficient. The incremental control model can be implemented in engineering using rule-based mapping or a trained control model, but its outputs must satisfy monotonicity and amplitude constraints: the takeover tendency increases with stability and decreases with risk, while the takeover ratio must not change beyond a preset maximum range within a single cycle to prevent sudden changes in on-site operations. The maximum range is determined by the operational safety strategy, and its value is determined by combining historical manual handover experience with the system's tolerable dispatch strategy drift range, ensuring that the rate of change in the takeover ratio does not exceed the pace that on-site dispatchers and vehicle terminals can adapt to.

[0086] To avoid outputting overly aggressive takeover ratios when risk is high but stability is also high, the control model introduces a risk suppression term, which hardens the takeover ratio due to high risk levels. The strength of this suppression term is determined by the operational risk preference configuration; the more conservative the risk preference, the stronger the suppression. After the initial takeover ratio coefficient is output, the system will also perform boundary pruning to ensure it is neither lower than the minimum takeover ratio nor higher than the maximum takeover ratio. The minimum takeover ratio is used to ensure that the model still has participation in low-risk situations to continuously accumulate feedback data; its value is determined by the model's online learning and evaluation needs. The maximum takeover ratio is used to ensure that even if the model performs well, a certain proportion of the rule engine or hybrid decision-making channel is reserved as a safety margin; its value is determined by operational safety requirements and fixed in the strategy.

[0087] Subsequently, the initial takeover ratio coefficient is modified using the model decision confidence to generate dynamic weight factors, and the final takeover ratio is determined based on these dynamic weight factors. Model decision confidence reflects the model's self-consistency and feasibility regarding its output in the current state. Its generation is not dependent on external annotations but is intrinsically derived from the model output and the deduction process. Specifically, the system extracts concentration features from the probability distribution of candidate actions related to the shuttle bus. When the probability is highly concentrated on a few actions, and these actions are more likely to yield conflict-free solutions in joint action deduction, the confidence increases; when the probability distribution is dispersed and the deduction requires multiple conflict resolution iterations to converge, the confidence decreases. To avoid misjudging "overconfident but incorrect" situations as high confidence, the system performs consistency checks between the confidence and recent offset performance: when the recent offset rate increases but the current output still shows extreme concentration, the system lowers the confidence to suppress possible overfitting or state misreading.

[0088] The confidence level is standardized within the system and converted into a correction intensity using pre-defined mapping rules. The principle for determining the correction intensity parameter is that the impact of confidence level on the takeover ratio should be less than its fundamental determinant of stability and risk; that is, confidence level is used for fine-tuning rather than radicalization. Therefore, the correction magnitude is limited to a small range. The upper bound of this range is determined by the sensitivity assessment of stability to fluctuations in the takeover ratio during historical operation, while the lower bound ensures that changes in confidence level do indeed produce observable adjustment effects. The system combines the initial takeover ratio coefficient with the confidence level correction intensity to generate a dynamic weighting factor. This dynamic weighting factor is adjusted upwards as confidence level increases and downwards as confidence level decreases, and after each adjustment, amplitude limiting and smoothing are performed again. The smoothing process aims to prevent fluctuations in the takeover ratio due to single-cycle confidence level fluctuations. The smoothing window length is determined to cover at least three decision cycles to filter out instantaneous fluctuations while ensuring timely response even when the operational situation changes significantly. Finally, the system determines the proportion of shuttle bus scheduling functions to be taken over by the model decision within the current operational cycle based on the dynamic weighting factor.

[0089] S6 includes receiving the first modified instruction set transmitted through the instruction coordination channel and the shuttle bus dispatching function takeover ratio value; based on the comparison result of the shuttle bus dispatching function takeover ratio value and the preset threshold, extracting model instructions from the first modified instruction set or recombining them to generate hybrid instructions; sending the model instructions or hybrid instructions to the vehicle terminal interface to drive the vehicle to perform actions, and collecting the vehicle position coordinate data and task status change data generated after execution; updating the vehicle real-time mapping status based on the vehicle position coordinate data and task status change data.

[0090] In this embodiment, the system first receives a first revised instruction set and a shuttle bus dispatch function takeover ratio value through the instruction coordination channel. The first revised instruction set is formed after compliance verification and rule substitution, and the instructions are already executable before entering the instruction coordination channel. Upon receiving the instruction set, the system performs structural and integrity verification. The structural verification is used to confirm that the instruction fields are complete and comply with the vehicle terminal interface protocol requirements. The fields include at least a unique instruction identifier, vehicle identifier, action type, target location or target machine area, planned execution time window, necessary safety constraint markers, and instruction source markers.

[0091] Integrity checks are used to confirm whether there are obvious gaps or mutual contradictions in the instruction sequence of the same vehicle within the current operating cycle. For example, the same vehicle may be assigned two overlapping and non-parallel actions in time windows. If structural deficiencies or mutual contradictions are found, the system does not issue a direct order but triggers an instruction set consistency review process. Instructions marked as originating from the rule engine in the first corrected instruction set are used as a fallback to avoid terminal execution failures due to instruction format or sequence issues. The shuttle bus scheduling function takeover ratio is written into the current cycle control context after being received and bound to the current cycle identifier, ensuring that subsequent threshold determination and sampling selection are performed within the same cycle caliber.

[0092] After obtaining the takeover ratio value, the system determines the issuance strategy based on the comparison between the takeover ratio value and the preset threshold, and extracts model instructions from the first revised instruction set or reassembles them to generate hybrid instructions. The preset threshold is used to map the takeover ratio to discrete execution modes, and its configuration follows the principles of "multi-segment controllability, risk interpretability, and smooth change". The threshold is not arbitrarily specified, but determined by combining historical deviation performance and operational risk tolerance: the system statistically analyzes the rule conflict rate, critical flight impact probability, and on-site rework rate under different takeover ratios in historical playback, selects the takeover ratio boundary point that can keep the above risk indicators within the allowable range of the operational strategy as the threshold, and allows the use of different threshold versions during peak and off-peak periods to adapt to differences in operational intensity.

[0093] The number of thresholds can be one or more. When using one threshold, the system enters the model-instruction-first mode when the takeover ratio is above or equal to the threshold, and enters the hybrid instruction mode when it is below the threshold. When using multiple thresholds, the system can be divided into low takeover intervals, medium takeover intervals, and high takeover intervals, corresponding to rule-first, hybrid, and model-first modes, respectively, thereby achieving a more granular gradual implementation. To avoid frequent switching of execution modes due to fluctuations in the takeover ratio around the threshold, the system introduces a threshold hysteresis band, which sets a small buffer range above and below the threshold. The mode is switched only when the takeover ratio crosses this buffer range and remains there for several consecutive cycles. The width of the hysteresis band is determined by the statistical results of the historical fluctuation amplitude of the takeover ratio, and is usually selected to cover the high-frequency fluctuation range of common jitter, to ensure that the mode switching reflects the real trend rather than short-term noise.

[0094] Once the execution mode is determined, the first revised instruction set is classified and filtered based on instruction source markers and risk tags. Instruction source markers distinguish between model-generated and unreplaced instructions, rule engine-replaced instructions, and instructions generated by system downgrading or rearrangement. Risk tags, derived from trigger location tags and conflict type statistics in compliance checks and offset records, are used to identify critical flights, critical gates, high-conflict areas, and safety-sensitive time windows. In model-instruction-priority mode, the system prioritizes "preserving effective model decisions as much as possible," extracting model-sourced shuttle bus instructions that have not been replaced by rules from the instruction set as candidate instructions. Set-level consistency checks are performed on these candidate instructions to ensure the time window sequence for the same vehicle is continuous and free from mutual exclusion conflicts. If a vehicle's model instructions have breaks in the sequence or potential conflicts with safety-sensitive constraints, the system adopts a partial rollback strategy for that vehicle, switching only the conflicting instruction segment corresponding to that vehicle to rule engine-sourced instructions without affecting the issuance of model instructions for other vehicles, thus achieving fine-grained control at the vehicle level. In hybrid instruction mode, the system recombines model-sourced instructions and rule engine-sourced instructions according to preset hybrid rules.

[0095] The hybrid rules include at least three types of constraints: The first type is the critical task priority rule, which prioritizes instructions involving critical flights or security-sensitive gates using the rule engine source to reduce risk; the second type is the consistency rule, which maintains the same source strategy for the same vehicle's task chain as much as possible to avoid frequent switching that could cause unstable execution rhythm, allowing for local insertion or replacement only when necessary; the third type is the proportion achievement rule, which, while satisfying the aforementioned constraints, ensures that the final proportion of model source instructions is as close as possible to the takeover ratio, guaranteeing that the gradual implementation goal is actually executed. The proportion achievement calculation process is based on a "replaceable set": the system first identifies which instructions can select a model or rule version without triggering hard constraints, forming a replaceable set; then, within the replaceable set, instructions are sorted from low to high risk, prioritizing the use of model instructions to fill the low-risk portion, until the proportion of model instructions reaches the target takeover ratio or the safety constraint upper limit. The sorting is determined by risk labels and historical offset sensitivity. Historical offset sensitivity comes from statistics on the offset frequency in similar scenarios; higher offset frequency indicates a more unstable model in that scenario, resulting in a lower ranking. In this way, the hybrid reorganization can reflect the takeover ratio without introducing excessive model takeover of high-risk scenarios due to mechanical proportional sampling.

[0096] After extraction or hybrid recombination, the system sends model instructions or hybrid instructions to the vehicle terminal interface to drive the vehicle to perform actions, and collects the vehicle position coordinate data and task status change data generated after execution. The vehicle terminal interface can be the vehicle control interface provided by the rigid business system or the standard interface of the vehicle dispatching system. Before issuing, a terminal-parsable message carrier is generated for each instruction, and a unique instruction identifier, vehicle identifier, action parameters, and time window parameters are written in it. At the same time, a message sequence number is written so that the terminal can perform sequential control.

[0097] The message sequence number is determined by sequentially incrementing the planned start time for the same vehicle within the same operating cycle to ensure that the terminal can still execute in order even in the event of network out-of-order delivery. A confirmation mechanism is used for message delivery: after sending, the system waits for the terminal to return a receipt, which includes at least the vehicle identifier, a unique instruction identifier, the reception time, and the reception result code. If no receipt is received within a preset waiting time, the system retransmits the message according to a retry strategy. The number of retries and the retry interval are determined by link reliability statistics, with the principle being to cover common packet loss scenarios without causing link congestion. If the retry still fails, the system marks the instruction as a delivery anomaly and triggers a degradation strategy. The degradation strategy typically involves using a rule engine instruction or entering a safe waiting point instruction for the vehicle to ensure that the vehicle does not enter an uncontrollable state due to missing instructions. During vehicle execution, the terminal continuously or by event reports vehicle location coordinate data and task status change data. The location coordinate data includes at least coordinates and timestamps, and the task status change data includes at least status codes such as task start, execution, completion, cancellation, or exception, as well as the time of occurrence. It should also carry a task identifier or receipt identifier that can be associated with a unique identifier of the instruction to ensure that the feedback corresponds one-to-one with the instruction.

[0098] After collecting vehicle location coordinate data and task status change data, the real-time vehicle mapping status is updated based on this data. The update process includes data cleaning, location mapping, status fusion, and anomaly identification. Data cleaning is used to handle positioning jitter, duplicate reporting, and short-term missing reports: the system deduplicates location data in chronological order, with the deduplication window determined by the location reporting cycle statistics; jitter is smoothed, with the smoothing amplitude determined by the upper limit of vehicle positioning error statistics, so that short-term jumps are not misjudged as vehicle instantaneous movement; missing reports are filled with the most recent valid value for a short time, but the upper limit of the filling time is determined by the longest reporting interval of positioning data under normal operation. If the upper limit is exceeded, the vehicle location is marked as unreliable and an alarm is triggered. Location mapping is used to map coordinates to airport spatial units or stand area indexes. The mapping is determined based on the airport spatial grid or stand area boundary configuration. If the coordinates fall outside the boundary, they are processed according to a preset correction strategy. The allowable distance of the correction strategy is determined by historical positioning drift statistics to avoid mapping obviously abnormal coordinates as valid locations. State fusion is used to uniformly write task state changes and location mappings into the vehicle real-time mapping state table: The system updates the vehicle's current space ownership, current task occupancy, task stage and expected completion status using the vehicle identifier as an index, and performs consistency verification between the task state change time and the location time. If there is a contradiction where the task is completed but the location is still far from the target area, it is recorded as a consistency anomaly and enters the review queue.

[0099] Anomaly identification is used to identify situations such as "failure to execute on time, execution timeout, deviation from the route, entry into prohibited areas, and repeated task switching." The "timeout" threshold is determined by the highest percentile of historical execution times for similar tasks, combined with network and on-site execution latency margins. The "deviation from the route" determination is based on the relationship between the permitted area of ​​the road network and the vehicle's current location, and a spatial buffer is set in conjunction with the upper bound of the positioning error to avoid false positives. All anomalies are written to the alarm queue along with the associated instruction unique identifier, vehicle identifier, and trigger time, so that they can be reflected in the status snapshot in subsequent decision-making cycles and drive a more conservative takeover strategy.

[0100] S7 includes receiving real-time vehicle location coordinates and task status change data, and analyzing the vehicle road network distribution status using the airport road network topology model; calculating the driving distance value based on the vehicle road network distribution status and flight arrival events; generating a task queue to be matched if the driving distance value meets the preset shuttle service range; obtaining the vehicle driving trajectory data and remaining battery information associated with the task queue to be matched, and constructing a multi-dimensional data matrix by combining the vehicle driving trajectory data and the remaining battery information; performing spatiotemporal dimension standardization processing on the multi-dimensional data matrix to generate a second operating status snapshot containing complete vehicle location and flight event information.

[0101] In this implementation, the system first receives real-time vehicle location coordinates and task status change data, and then organizes the data using vehicle identifiers and timestamps as primary indexes. Location coordinate data includes at least the vehicle identifier, coordinate values, timestamp, and positioning quality marker; task status change data includes at least the vehicle identifier, task identifier, status code, status occurrence time, and association information with the instruction's unique identifier. Before entering the road network parsing stage, the system performs time alignment, deduplication, and anomaly suppression on the data. Time alignment is achieved by uniformly placing all location points and status change points into a preset time slice granularity. The time slice granularity is not arbitrarily set but is determined jointly by the reliable reporting period of vehicle locations and the minimum step size allowed by the closed-loop scheduling system, typically choosing the larger of the two, to avoid overly fine time slices leading to amplified coordinate jitter and overly coarse time slices causing delayed expression of status changes.

[0102] The deduplication method is as follows: for multiple location reports of the same vehicle within the same time slice, only the one with the latest timestamp is retained, and the number of discarded records is counted as a link quality indicator. The anomaly suppression method is as follows: the displacement of the location points in two consecutive time slices is calculated, and the maximum reasonable movement capability allowed by the vehicle within the field is used as the upper bound constraint. This upper bound is jointly determined by the speed limit rules of the vehicle type, the configuration of the road speed limit zone, and the length of the time slice, and is superimposed with the buffer margin of the upper bound of the positioning error. The upper bound of the positioning error is obtained from the reconciliation statistics of historical positioning deviations, usually using the high quantile level as the upper bound, to ensure that the buffer covers anomaly drift without masking real anomalies. If the displacement exceeds the upper bound, the system does not directly use the point in the road network analysis, but marks it as untrusted and triggers a correction strategy. The correction strategy prioritizes selecting a reasonably adjacent location point with a reasonable road network connectivity relationship with the previous trusted point. If correction cannot be performed, the previous trusted point is retained and the time slice is marked as missing to avoid erroneous coordinates polluting subsequent distance calculations.

[0103] After data cleaning, the vehicle road network distribution is obtained by analyzing the airport road network topology model. The airport road network topology model consists of a set of road nodes, a set of road segments, road segment connectivity, directional constraints, restricted areas, temporary control zones, speed limit zones, and key access point configurations. The road nodes and road segments are derived from the airport's basic map data, while the restricted and control configurations are derived from operational management strategies. Road network analysis is achieved through "coordinate projection onto road segments": the system calculates the spatial distance from each reliable location point to each candidate road segment. The selection of candidate road segments does not involve a full traversal but rather narrows down the range using spatial indices, allowing the calculation to be completed under real-time constraints. After projection, the relative position of the vehicle to its road segment and on that road segment is obtained.

[0104] To prevent positioning jitter from causing vehicles to jump back and forth between adjacent road segments, the system performs a continuity check on the projection results. This continuity check is based on whether the projection is connected to the road segment of the previous time slice. If connected, the current projection is accepted; otherwise, an alternative road segment connected to the previous road segment and with a projection distance within the allowable range is searched within the current location's neighborhood. This allowable range is determined by the upper bound of the positioning error and the minimum distance between road segments, ensuring that substitution only occurs within the explainable positioning drift range. Ultimately, the vehicle road network distribution status includes at least the road segment identifier, the road segment's travel direction, the vehicle's relative position on the road segment, and a road segment-level congestion marker. The congestion marker is generated by counting the number of vehicles mapped to the same or adjacent road segments within the same time slice, and classifying them based on the road segment's capacity configuration. The capacity configuration is determined by basic parameters such as road width, one-way restrictions, and intersection capacity. When the count exceeds the typical carrying capacity of the road segment during the same time period, it is marked as congested. Typical load levels are obtained from historical road network vehicle distribution statistics. Thresholds are usually set in layers according to time periods to avoid misjudgment caused by using the same threshold for peak and off-peak periods.

[0105] After obtaining the vehicle road network distribution status, the system calculates the travel distance by combining flight arrival events. Flight arrival events are provided by the flight event stream and include at least the flight identifier, arrival time, arrival gate or arrival area identifier, and connection demand marker. The system first maps the arrival gate or arrival area to road network target points. These target points generally correspond to gate entrance nodes or road nodes adjacent to the boarding gate. The mapping relationship is provided by a binding table between gates and road nodes, which is derived from the airport infrastructure configuration and fixed before going live. The travel distance is calculated based on the shortest feasible path of the road network and incorporates restrictions such as traffic restrictions, one-way traffic, and temporary traffic control to ensure that the distance reflects true accessibility. To improve consistency with actual travel time, the system can convert road segment speed limit levels, congestion markers, and control levels into "road segment travel costs," so that the shortest path is not simply the geometric shortest, but the feasible path with the lowest cost.

[0106] The determination of road segment toll costs follows a unified standard: speed limit levels correspond to basic costs, congestion markers correspond to additional costs, and control levels correspond to prohibition or extremely high costs. The additional cost coefficient is not subjectively set but is obtained through fitting historical traffic data. Specifically, it compares the average travel time of the same road segment under different congestion levels and sets the cost multiplier according to the travel time multiplier, making the cost sensitive to congestion changes while remaining monotonic. In cases where real-time congestion data is lacking, the system degenerates into obtaining a basic distance by summing the geometric lengths of the roads, and adds a safety margin to this basic distance. The size of the safety margin is determined by the historical statistical upper bound of the "difference between straight-line distance and actual road network distance," ensuring a conservative assessment of accessibility even in the absence of road condition information. To avoid abnormally amplified distance results due to abnormal road network mapping, the system sets a maximum clipping upper limit for the distance results. This upper limit is determined based on historical statistics of the longest reasonable detour route of the airport road network, typically using a high quantile level as the upper limit, covering extreme detours but excluding obvious erroneous mappings.

[0107] When the travel distance meets the preset shuttle service range, a queue of tasks to be matched is generated. The preset shuttle service range is used to filter out vehicles that can handle the shuttle service for this flight from the entire fleet of vehicles. Its determination is based on the timeliness requirements of the shuttle service. In practice, the system configures the shuttle service range as a distance upper bound. This distance upper bound is not a fixed constant, but is calculated by combining the "allowed arrival time upper limit" and the "typical road speed lower limit," with a safety margin added to cover congestion and execution delays. The allowed arrival time upper limit is determined by the flight shuttle service specifications, such as the maximum tolerable waiting time for arriving passengers. The typical road speed lower limit is obtained from the historical lowest stable throughput capacity of vehicles under speed limit and congestion conditions, usually using a low percentile to maintain conservatism. The safety margin is determined by network latency, terminal execution latency, and path deviation caused by positioning errors, with a high percentile used to cover abnormal fluctuations. During the screening process, the system not only compares distances but also filters based on vehicle task status: vehicles in an uninterrupted operation phase are removed, while those in a standby or transferable phase are retained. The determination of a vehicle's transferability is based on the task status code and the estimated remaining task duration. The remaining task duration is calculated from the task status change time and the standard working time model. The standard working time model is derived from historical stratified statistics and includes a robust central value and fluctuation range to reduce the impact of abnormal delays.

[0108] For vehicles that meet the service range and availability requirements, the system generates candidate matching entries and sorts them according to flight arrival time. Simultaneously, it adjusts the queue based on critical flight priority. Critical flight priority is configured by the airport's operational strategy and can be determined based on factors such as flight class, international / domestic attribute, and delay sensitivity, and is permanently stored as configurable tags in the system. For multiple candidate vehicles on the same flight, the system sorts them from nearest to farthest, prioritizing vehicles with higher battery levels or whose trajectory displays are closer to the target direction when distances are similar, thus making the candidate queue more achievable.

[0109] Subsequently, the system acquires vehicle trajectory data and remaining battery information associated with the task queue to be matched, and constructs a multi-dimensional data matrix by combining the vehicle trajectory data and remaining battery information. Vehicle trajectory data is formed from the continuous location reports of vehicles over a recent period. The system extracts a sequence of trajectory points according to a preset trajectory window. The length of the trajectory window is determined based on the principle of "sufficient to reflect short-term movement trends without introducing excessive historical noise." Its value is jointly determined by the vehicle location reporting cycle and the typical transition time: the reporting cycle determines the trajectory point density, and the typical transition time determines the time span to be covered. Typically, a length sufficient to cover the typical time it takes for a vehicle to transition from one location to an adjacent area is selected to reflect the vehicle's current direction of movement and accessibility trends.

[0110] Remaining battery power information is sourced from the vehicle terminal or energy management module, and includes at least the battery percentage and estimated available range. The battery percentage is typically a directly reported value, while the estimated available range is calculated from a historical energy consumption model. This model is stratified by vehicle type, passenger load, air conditioning load, and road condition level, using a robust central value as a benchmark to ensure consistent range estimation across different operating conditions. The system writes the current road network location, distance to the target, trajectory point sequence summary, remaining battery power, estimated range, and associated flight arrival event information for each candidate vehicle into a unified structure, forming a multi-dimensional data matrix. To control data volume and maintain usability, the trajectory point sequence can be further compressed into key summaries, such as the road segment sequence, average speed, number of stops, and number of direction changes for the most recent time slices. These summary indicators are calculated based on the road network mapping results of the trajectory points, preserving motion trend information without retaining all original points.

[0111] The multidimensional data matrix is ​​standardized in terms of its spatiotemporal dimensions to generate a second operational status snapshot containing complete vehicle location and flight event information. Temporal standardization is achieved by aligning the trajectories and states of different vehicles to the same set of standard time slices, and merging trajectory points falling between time slices according to the nearest time slice. Short-term interpolation is allowed for missing time slices, but the interpolation length is constrained by the maximum allowable gap; exceeding this constraint marks the trajectory as incomplete. The maximum allowable gap is determined based on location reporting reliability statistics, typically using the high quantile of the longest consecutive missing reporting duration under normal link quality as the upper limit to avoid masking the true loss of connection with excessively long interpolations.

[0112] Spatial standardization is achieved by unifying all spatial representations into road segment identifiers, relative road segment positions, and corresponding aircraft stand neighborhood markers. The aircraft stand neighborhood markers are provided by the adjacency list of road segments and aircraft stand regions, enabling the model to understand the relative spatial relationship between vehicles and aircraft stands. For points where mapping is uncertain or where correction occurs, the system adds a trusted marker. The trusted marker is generated based on the positioning quality marker, whether the continuity check passes, and whether correction occurs, ensuring that subsequent decisions can adopt a more conservative strategy for low-confidence data. After completing spatiotemporal standardization, the system encapsulates the vehicle road network distribution status, driving distance values, queue of tasks to be matched, trajectory summary, battery and range summary, and flight arrival event summary into a second operational state snapshot, and generates a unique snapshot identifier and time index for this snapshot.

[0113] S8 includes acquiring policy offset records and constructing an offset status dataset based on the policy offset records; performing time-series feature analysis on the offset status dataset to generate an offset trend feature vector; calculating the rule conflict rate value based on the offset trend feature vector; and generating a takeover ratio increment instruction if the rule conflict rate value is lower than a preset threshold for multiple consecutive periods, so as to automatically increase the takeover ratio of the shuttle bus scheduling function according to the takeover ratio increment instruction.

[0114] In this implementation, policy offset records are first acquired, and an offset state dataset is constructed based on these records. The policy offset records are derived from the accumulated results of the first policy offset records formed in step S4 within the time-series index. Each record contains at least a unique instruction identifier, the issuing agent identifier, the decision cycle timestamp of the offset occurrence, the offset cause category, the rule conflict type, the offset intensity level, the trigger location label, and a summary of the resource state context at the time of the offset occurrence. When constructing the offset state dataset, the system filters based on shuttle bus scheduling, using either "resource type is shuttle bus" or "instruction target is related to shuttle bus scheduling functions," thereby eliminating interference from irrelevant offsets such as those from cleaning teams and baggage tractors on the takeover ratio decision.

[0115] After filtering, the system aggregates data according to the decision cycle, generating a cycle-level offset summary record for each cycle. This summary record includes at least the total number of shuttle bus model instructions within that cycle, the number of instructions corrected, the number of instructions that underwent rule conflict corrections, the frequency of each conflict type, the distribution of offset intensity levels, and a summary of the offset trigger locations. The statistical definition of the total number of model instructions is based on "instructions whose source is marked as a model and which have entered the compliance verification process," avoiding the inclusion of instructions directly generated by the rule engine in the denominator, which would dilute the conflict rate. To avoid incomparability caused by differences in the number of instructions across different cycles, the system simultaneously calculates percentage-type indicators and retains the original count-type indicators. The count is used for capacity assessment, while the percentage is used for stability comparison.

[0116] After constructing the offset state dataset, the system performs time-series feature analysis on the dataset to generate offset trend feature vectors. The time-series feature analysis is performed in units of a preset analysis window, determined by both the window length and the sliding step size. The principle for determining the window length is: it must cover a complete typical fluctuation segment in airport operations to avoid misjudging long-term stability based solely on short-term stability; at the same time, it should not be too long to prevent the increase in takeover ratio from lagging behind the actual improvement in the model. A common practice is to select a number of periods that cover at least one peak period or at least one flight span as the window length, based on the decision cycle length, and allowing different window configurations for peak and off-peak periods. The sliding step size controls the update frequency; its value is generally equal to one decision cycle to ensure that the trend judgment is updated in each cycle, or a larger step size can be used to reduce computational overhead when computing power is limited.

[0117] Multiple time-series features are extracted within each window and arranged in a fixed order to form an offset trend feature vector. These features include at least the window mean of the proportion of rule-conflict offsets, the strength of the upward or downward trend within the window, the fluctuation amplitude within the window, the length of consecutive high offset segments, and the concentration of conflict types. The trend strength is calculated by observing the change in the proportion of conflicts in chronological order within the window, determining whether the overall trend is downward, upward, or approximately flat, and using the stability of the change as the quantification of strength. The fluctuation amplitude is used to characterize whether there are large fluctuations; the system obtains the fluctuation level by summarizing the deviation between the proportion of conflicts in each period within the window and the window mean. The concentration of conflict types is used to determine whether the offset is concentrated on a few rules. The system statistically analyzes the proportion of each conflict type within the window; when a certain type dominates for a long period, the concentration increases, indicating that the model may not be fully adapted to the boundaries of specific rules. To ensure comparability between different windows, the system performs range normalization and hierarchical encoding on all features. The normalization scale is obtained from historical sample statistics and is maintained consistently through version management.

[0118] After obtaining the offset trend feature vector, the rule conflict rate is calculated based on the offset trend feature vector. The rule conflict rate is used as the core metric for increasing the takeover ratio, and its calculation method must be strictly consistent. The system first calculates the cycle conflict rate within each decision cycle, and then forms the window conflict rate within the analysis window. The calculation of the cycle conflict rate is based on the ratio between the number of rule conflict type offset events and the number of shuttle bus model instructions, where rule conflict type offsets include at least flight shortest turnaround time conflicts, gate occupancy conflicts, safety boundary conflicts, and resource mutual exclusion conflicts.

[0119] To prioritize safety, the system allows for the configuration of risk weights for different conflict types. These weights are determined based on operational safety levels and the priority of hard constraints: safety boundary conflicts have the highest weight, followed by station occupancy conflicts, and then minimum turnaround time conflicts and resource exclusion conflicts. The weights are calculated by converting operational risk stratification requirements into numerical proportions and calibrating them using the operational impact of different conflicts in historical reviews. This ensures that the impact of an equal number of safety boundary conflicts on the conflict rate is significantly higher than that of general conflicts. The window conflict rate is calculated by summarizing the periodic conflict rates within the window. During summarization, the principle of "proximate causes are more sensitive" is applied, assigning higher impact to data from the current period to ensure the system can respond promptly to recent improvements or degradations. The strength of this impact is controlled by a preset attenuation coefficient. The attenuation coefficient is determined to avoid overemphasizing occasional changes in a single period or excessive averaging that leads to sluggishness. It is typically determined based on the autocorrelation periodic statistics of historical conflict rates, making the window conflict rate sensitive to genuine trend changes but insensitive to short-term noise. The final rule-based conflict rate value is written into the periodic evaluation results and used for subsequent threshold determination.

[0120] When the rule conflict rate is detected to be lower than a preset threshold for multiple consecutive periods, an incremental takeover ratio instruction is generated, and the takeover ratio of the shuttle bus dispatching function is automatically increased based on this incremental instruction. The preset threshold is used to define the "stable level at which the takeover ratio can be increased." Its determination is not based on subjective experience, but rather on a combination of operational risk tolerance and historical stable operational data. Specifically, the determination method is as follows: select periods in historical operational data that are considered to be operating smoothly and with acceptable rule conflicts, statistically analyze the distribution of rule conflict rates within these periods, and take the quantile that represents the "low conflict" level as the threshold benchmark; at the same time, set stricter thresholds in more risk-sensitive scenarios (such as peak hours, severe weather, and periods with more regulations) so that the same conflict rate does not trigger an increase in high-risk scenarios.

[0121] Thresholds can be configured in tiers based on time period, airport operation level, proportion of critical flights, or gate congestion level. The tier boundaries are determined by the historical distribution of the aforementioned operational status indicators and control strategies to ensure that the thresholds remain consistent despite changes in risk. The decision length for multiple consecutive cycles is used to suppress aggressive upward adjustments caused by occasional low conflict rates. The principle for determining this length is that it must cover multiple cycles of at least one typical fluctuation segment to ensure statistical stability; at the same time, it should not be too long to avoid slow upward adjustments. Typically, the decision cycle length is used as a basis, selecting the number of cycles that can cover at least one typical congestion change or several key nodes within at least one flight as the continuous decision length, and allowing for an increase in this length during high-risk periods for a more conservative approach. The system adopts a "slight tolerance" rule when determining continuity, meaning that a few minor exceedances of the threshold for one cycle are allowed within a continuous interval, but sustained rebounds are not permitted. The allowed number of such tolerances is determined statistically by historical noise levels to avoid long-term non-triggering of upward adjustments due to abnormal noise in a single cycle disrupting continuity.

[0122] Upon meeting the triggering conditions, an incremental takeover ratio instruction is generated. This instruction must include at least the increment magnitude and an effective period identifier, and may also include a scope marker to limit its effect to shuttle bus dispatching functions only. The determination of the increment magnitude follows the principle of "smooth increase and controllable risk," prohibiting large, one-time increases. The system sets a maximum single-period growth limit for the increment magnitude. This limit is determined by the operational safety strategy, based on a sensitivity assessment of historical takeover ratio changes to conflict rate rebounds. Specifically, it observes in historical playback whether the conflict rate significantly increases when the takeover ratio increases too rapidly, thus setting the upper limit as the increment that "does not trigger a significant rebound." The increment magnitude can also be correlated with the margin between the current conflict rate and the threshold: when the conflict rate is far below the threshold and the trend continues to decline, a larger increment, but still not exceeding the upper limit, can be used; when the conflict rate is close to the threshold or fluctuates significantly, a smaller increment is used to reduce the risk of a rebound.

[0123] The effective period identifier is used to specify from which decision period the incremental instruction takes effect, typically the next decision period, to ensure a complete record chain between instruction generation and execution. After generating an incremental instruction, the system writes it to the progressive import control parameter storage and updates the current takeover ratio. Simultaneously, it performs boundary pruning on the updated takeover ratio to ensure it does not exceed the maximum takeover ratio and does not breach safety margin requirements. The maximum takeover ratio is fixed by the operational safety policy, ensuring that even if the model performs well, the fallback ratio of the rule engine or hybrid channel is retained. To prevent immediate rebound after an increase, the system activates a protection period strategy for several periods after the increase. The length of the protection period is determined by statistical analysis of typical lag periods in historical rebounds. If the conflict rate increases significantly during the protection period, a freeze or rollback is triggered. The trigger threshold for freeze or rollback can be set to a threshold higher than the increase threshold to create different threshold bands and avoid frequent back-and-forth oscillations.

[0124] 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. An airport ground handling task scheduling method based on cooperative multi-agent reinforcement learning, executed in an airport operation system configured with progressively introduced control parameters and equipped with a command coordination channel, characterized in that, include: S1. Obtain real-time flight location, gate occupancy status, special vehicle operation progress, and personnel check-in data from the rigid business system, and merge them to form a first snapshot of the operating status containing timestamps; S2. Input the first snapshot of the running state into the trained multi-agent reinforcement learning model. The model outputs the first set of decision instructions for resource scheduling. Each instruction carries the identifier of the issuing agent. S3. Perform compliance judgment on the first set of decision instructions according to the preset rules. If all instructions comply with the preset rules, generate the first verification pass signal. If there is an instruction that violates the preset rules, replace the instruction with the instruction generated by the rigid business system rule engine, generate the first set of corrected instructions, and record the rule conflict type and the source of the corrected instruction. S4. A monitoring mechanism is used to compare the instruction differences between the first decision instruction set and the first corrected instruction set. The instruction with differences is marked with the offset reason. The difference data, rule conflict type, associated resource status and first running state snapshot are associated and stored to form a first policy offset record containing the agent's behavior trajectory and rule violation context.

2. The airport ground handling task scheduling method based on cooperative multi-agent reinforcement learning according to claim 1, characterized in that: S1 includes: Obtain standardized business data streams, which are obtained by cleaning and transforming heterogeneous source data from rigid business systems. A spatial status dataset is generated based on standardized business data flow. The spatial status dataset is generated based on the spatial mapping of real-time flight location coordinates and the determination of gate occupancy status codes. Work chain data is constructed based on spatial state dataset. The work chain data is established by matching the work progress messages of special vehicles with the work node identifiers in the personnel sign-in logs. Time correction and multi-dimensional fusion are performed on the job chain data to generate the first snapshot of the running state.

3. The airport ground handling task scheduling method based on cooperative multi-agent reinforcement learning according to claim 1, characterized in that: S2 includes: Obtain a snapshot of the first running state, and input the snapshot of the first running state into the multi-agent reinforcement learning model to obtain the probability distribution of candidate actions for the shuttle bus, the cleaning team and the baggage towing vehicle; Joint action deduction is performed based on the probability distribution of candidate actions. If there is a spatiotemporal conflict in the deduction results, the policy gradient is updated by exchanging local observation information until a conflict-free joint action sequence is generated. The conflict-free joint action sequence is parsed, and the actions in the conflict-free joint action sequence are converted into scheduling operation codes and associated with the agent identifiers to generate the first set of decision instructions for resource scheduling.

4. The airport ground handling task scheduling method based on cooperative multi-agent reinforcement learning according to claim 1, characterized in that: S3 includes: Obtain a first set of decision instructions containing scheduling information, and map the first set of decision instructions to the preset shortest flight turnaround time rules and gate conflict safety boundaries to construct a compliance verification matrix; Extract the unique identifier of the violation instruction from the compliance verification matrix to generate a list of violation instructions; Input the list of non-compliant instructions into the rigid business system's rule engine to generate rigid alternative instructions that comply with security specifications; The corresponding instruction in the first decision instruction set is replaced by a rigid substitution instruction to generate the first modified instruction set.

5. The airport ground handling task scheduling method based on cooperative multi-agent reinforcement learning according to claim 1, characterized in that: S4 includes: Obtain the first set of decision instructions and the first set of corrected instructions, extract the list of instruction difference data, and retrieve the rule conflict type based on the list of instruction difference data to generate a difference dataset with cause labels; Retrieve the first runtime state snapshot based on the differential dataset with cause labels, extract the associated resource status and perform a structured association with the first runtime state snapshot; The intelligent agent's behavior trajectory is constructed by parsing the structured and associated dataset, and the rule conflict type and associated resource status are mapped to the intelligent agent's behavior trajectory to form a first policy offset record containing the intelligent agent's behavior trajectory and the rule violation context.

6. The airport ground handling task scheduling method based on cooperative multi-agent reinforcement learning according to claim 1, characterized in that, This also includes S5, which determines the proportion of shuttle bus scheduling functions taken over by model decision-making within the current operating cycle based on the historical offset sequence and offset frequency statistics in the progressively imported control parameters. Specifically, this includes: Obtain the historical offset sequence and offset frequency statistics from the progressive import control parameters, and extract the shuttle bus scheduling task feature data within the current operating cycle; The scheduling instruction stability index is generated based on the historical offset sequence, and the offset frequency statistics are converted into business risk level values. The initial takeover ratio coefficient is obtained by inputting the scheduling instruction stability index and the business risk level value into the progressive import control model. The initial takeover ratio coefficient is corrected using the model decision confidence level to generate a dynamic weight factor. Based on the dynamic weight factor, the proportion of shuttle bus scheduling functions taken over by the model decision in the current operating cycle is determined.

7. The airport ground handling task scheduling method based on cooperative multi-agent reinforcement learning according to claim 6, characterized in that, It also includes S6, where the instruction coordination channel receives the first verification pass signal or the first corrected instruction set, and selectively sends model instructions or hybrid instructions to the vehicle terminal of the rigid business system according to the current shuttle bus dispatching function takeover ratio. The system executes the instructions and updates the vehicle location and task status, specifically including: The first corrected instruction set transmitted via the instruction coordination channel and the percentage of shuttle bus dispatching function takeover; Based on the comparison between the shuttle bus dispatching function takeover ratio and the preset threshold, model instructions are extracted from the first corrected instruction set or recombined to generate hybrid instructions. The model instructions or hybrid instructions are sent to the vehicle terminal interface to drive the vehicle to perform actions, and the vehicle position coordinate data and task status change data generated after execution are collected. The real-time vehicle mapping status is updated based on vehicle location coordinate data and task status change data.

8. The airport ground handling task scheduling method based on cooperative multi-agent reinforcement learning according to claim 7, characterized in that, This also includes S7, which, in the next decision cycle, acquires updated vehicle location and mission status data, and combines this with new flight arrival events to form a second operational status snapshot, specifically including: Receive real-time vehicle location coordinates and task status change data, and analyze the vehicle road network distribution status using the airport road network topology model. The driving distance is calculated based on the vehicle network distribution and flight arrival events. If the driving distance value meets the preset shuttle service range, a queue of tasks to be matched is generated; Obtain vehicle driving trajectory data and remaining battery information associated with the queue of tasks to be matched, and construct a multi-dimensional data matrix by combining the vehicle driving trajectory data and the remaining battery information; The multidimensional data matrix is ​​standardized in terms of spatiotemporal dimensions to generate a second operational snapshot containing complete vehicle location and flight event information.

9. The airport ground handling task scheduling method based on cooperative multi-agent reinforcement learning according to claim 8, characterized in that, It also includes S8 and continuously accumulating strategy offset records, analyzing the offset trend within the time series index. If the analysis results show that the conflict rate of the model decision regarding the shortest turnaround time rule for flights is lower than a preset threshold over multiple consecutive periods, the takeover ratio of the shuttle bus scheduling function during the gradual import process is automatically increased, specifically including: Obtain the policy offset records and construct the offset state dataset based on the policy offset records; Perform time-series feature analysis on the offset state dataset to generate offset trend feature vectors.

10. The airport ground handling task scheduling method based on cooperative multi-agent reinforcement learning according to claim 9, characterized in that: S8 further includes: Calculate the rule conflict rate value based on the offset trend feature vector; If the rule conflict rate is lower than the preset threshold for several consecutive periods, an incremental takeover ratio instruction is generated to automatically increase the takeover ratio of the shuttle bus scheduling function.

Images