Method and system for predictive assignment decision optimization of air fleet flight missions

Abstract

This invention discloses a predictive assignment decision optimization method and system for aircraft fleet flights, belonging to the field of intelligent decision-making in aviation operations and maintenance. This invention constructs a multi-dimensional collaborative constraint system integrating mission and maintenance, simultaneously embedding simulation logic and reinforcement learning reward and punishment mechanisms; it builds a multi-agent simulation environment based on AnyLogic, achieving closed-loop bidirectional data interaction between the simulation environment and the reinforcement learning model through a standardized bidirectional interactive interface; the state space of the reinforcement learning model incorporates aircraft fault prediction and health management data, and generates assignment decision schemes after multi-scenario training. This invention significantly improves the adaptability of assignment decisions to dynamic and uncertain environments, achieves multi-objective collaborative optimization, strengthens the engineering coupling of simulation and reinforcement learning, and enhances model generalization and deployability, making it applicable to civil aviation and aviation combat support scenarios.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to a predictive assignment decision optimization method and system for aircraft fleet flight missions. Background Technology

[0002] Flight mission assignment decisions for an aircraft fleet are a core decision-making process within the fleet operations and maintenance system. Connecting strategic operational goals with frontline execution, the effectiveness of these decisions directly determines the safety level, cost control, and operational efficiency of fleet operations and maintenance. Current research and applications for optimizing fleet mission decisions mainly fall into two categories: traditional decision-making methods and intelligent decision-making methods. Traditional decision-making methods, centered on rule engines, static programming, and heuristic algorithms, formulate assignment schemes through manually set fixed logic and mathematical programming models. They are widely used in basic task allocation, flight scheduling, and maintenance planning in the aviation industry and can solve fleet assignment problems in small-scale, deterministic scenarios.

[0003] Intelligent Decision-Making Methods: With the development of machine learning technology, reinforcement learning (RL) and deep reinforcement learning (DRL) have been gradually applied to the optimization of aircraft fleet decisions. Existing research generally abstracts problems such as fleet scheduling, flight recovery, and maintenance planning into Markov Decision Processes (MDPs), and uses algorithms such as Q-learning, PPO, and A2C to achieve sequential decision-making in dynamic environments. Meanwhile, discrete event simulation (DES) and multi-agent simulation (MAS) have become important tools for training and validating decision-making algorithms. Some studies have achieved a preliminary combination of simulation technology and single-agent reinforcement learning, validating the feasibility of intelligent methods in scenarios such as flight recovery for small fleets and optimization of single-aircraft maintenance strategies. Furthermore, existing technologies have begun to explore the application of multi-agent reinforcement learning (MARL) and hierarchical reinforcement learning (HRL) in the aviation field, attempting to solve the state-action space explosion problem of single-agent models, while expanding research objectives from simply minimizing delays / costs to multi-objective optimization directions such as maintenance interval balancing and carbon emission control.

[0004] While existing technologies provide a foundation for optimizing aircraft fleet assignment decisions, several key issues remain in practical engineering applications. The core shortcomings are as follows: Poor adaptability to dynamic environments: Traditional rule engines and static planning methods rely on fixed logic and cannot effectively cope with multi-source uncertainties in aviation operations and maintenance, such as sudden aircraft malfunctions, emergency mission insertions, changes in flight route weather, and temporary shortages of maintenance resources. This can easily lead to the failure of the original assignment plan, requiring repeated manual adjustments and resulting in low decision-making efficiency.

[0005] Imbalance in multi-objective optimization: Existing methods often focus on a single objective (such as minimizing mission delays or minimizing maintenance costs) as the core of optimization, making it difficult to simultaneously optimize multiple objectives such as mission completion rate, resource utilization, maintenance interval balance, and long-term fleet availability. Although some studies have attempted multi-objective optimization, they have not established an adaptive weight balancing mechanism, which can easily lead to the problem of "emphasizing short-term missions and neglecting long-term maintenance," resulting in overuse or over-maintenance of aircraft.

[0006] Low integration of simulation and reinforcement learning, and lack of engineering coupling: Existing research mostly relies on simplified simulation environments to train reinforcement learning models, failing to achieve high-fidelity simulation and reinforcement learning in an engineering-friendly closed-loop interaction; the interface design between simulation systems and reinforcement learning algorithms is imperfect, making it difficult to achieve real-time linkage of state input, action feedback, and reward evaluation, resulting in a disconnect between the trained strategy and real operation and maintenance scenarios, and poor applicability.

[0007] Weak generalization and deployability of models: Existing intelligent dispatch models are mostly designed for specific scenarios and fixed fleet sizes, lacking a standardized modeling and training framework. When scenarios change or fleet sizes are adjusted, the models need to be retrained, resulting in high adaptation costs and difficulty in quickly deploying them to different real-world scenarios such as civil aviation and aviation combat support.

[0008] The core reason for the above problems is that existing technologies have not achieved standardized modeling of all elements of fleet operation and maintenance, accurate simulation of multi-source uncertainties, engineering coupling of simulation and reinforcement learning, and innovation of reinforcement learning algorithms adapted to fleet assignment, which makes it difficult for intelligent assignment decision-making methods to meet the actual needs of complex aviation operation and maintenance scenarios. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a predictive assignment decision optimization method and system for flight missions of an aviation fleet. This invention addresses the core problems in the prior art, such as poor adaptability to dynamic environments, imbalance in multi-objective optimization, low integration of simulation and reinforcement learning, and weak model generalization and deployability. It enables predictive and intelligent assignment decision optimization for flight missions of an aviation fleet, taking into account both short-term mission execution and long-term operation and maintenance support, and improving the robustness, feasibility, and scenario adaptability of the assignment strategy.

[0010] This invention provides a method for optimizing predictive assignment decisions for aircraft fleet flight missions, comprising the following steps: A multi-dimensional collaborative constraint system integrating mission maintenance is constructed. The multi-dimensional collaborative constraint system includes time constraint rules, resource constraint rules, aircraft status constraint rules, priority constraint rules, and mission maintenance cross-constraint rules. The multi-dimensional collaborative constraint system is synchronously embedded into the simulation logic of the simulation environment. A multi-agent simulation environment is constructed, and a two-way interaction interface is set up. Through the two-way interaction interface, a closed-loop two-way data interaction is realized between the multi-agent simulation environment and the reinforcement learning model, including state data output, assigned action input, and reward value feedback. Based on the multi-dimensional collaborative constraint system, the multi-agent simulation environment, and the closed-loop bidirectional data interaction, the training of the reinforcement learning model is completed. The state space of the reinforcement learning model includes aircraft fault prediction and health management data, and an assignment decision scheme is generated.

[0011] Therefore, by establishing a multi-dimensional collaborative constraint system that integrates mission maintenance, the rules for assignment decision-making and operation and maintenance support are unified, avoiding conflicts between mission execution and maintenance plans. Through closed-loop two-way data interaction between the simulation environment and the reinforcement learning model, high-fidelity simulation and intelligent decision-making are coupled in an engineering manner, solving the problem of the disconnect between training strategies and real scenarios. By incorporating aircraft fault prediction and health management data into the state space, assignment decisions become predictive, enabling a shift from post-event adjustments to proactive operation and maintenance based on predictive decisions, thereby improving the safety and forward-looking nature of fleet operation and maintenance.

[0012] This involves standardizing and cleaning the collected raw fleet operation and maintenance data, filling in missing values, and verifying consistency to establish a unified fleet operation and maintenance data model. The data is then output in a unified format through a standardized interface.

[0013] Therefore, by standardizing the processing of raw fleet operation and maintenance data, a unified and standardized data foundation is provided for the subsequent construction of constraint systems, simulation environment setup, and reinforcement learning model training. This eliminates the format differences of multi-source heterogeneous data, ensures the consistency and accuracy of data flow between modules, and provides standardized data support for the generalization and adaptation of models.

[0014] The multi-dimensional collaborative constraint system is also equipped with a constraint conflict resolution sub-module, which automatically coordinates multiple constraint rules according to preset priorities when conflicts occur.

[0015] Therefore, the constraint conflict resolution submodule enables automatic coordination of multiple constraint rule conflicts, avoiding simulation logic anomalies and unexecutable decision schemes caused by constraint conflicts, improving the compliance and executability of assigned decision schemes, and reducing the cost of manual intervention.

[0016] The multi-agent simulation environment is built on the AnyLogic platform and includes a multi-agent modeling module and an uncertainty disturbance generation module. The multi-agent modeling module constructs all roles of fleet operation and maintenance agents and their corresponding interaction rules, while the uncertainty disturbance generation module embeds multi-source random disturbances during the simulation process.

[0017] Therefore, a multi-agent simulation environment built on the AnyLogic platform can achieve high-fidelity simulation of the entire fleet operation and maintenance process, accurately reproducing real assignment scenarios; the multi-agent modeling module fully characterizes the behavioral logic and interaction rules of each role in fleet operation and maintenance, ensuring the realism of the simulation environment; and the uncertainty perturbation generation module embeds multi-source random perturbations, enabling the reinforcement learning model to fully adapt to dynamic and uncertain scenarios during training, improving the robustness and environmental adaptability of the assignment strategy.

[0018] The multi-source random disturbances include sudden aircraft malfunctions, emergency mission insertions, and weather disturbances along flight routes.

[0019] Therefore, by covering multi-source random disturbances such as sudden aircraft failures, emergency mission insertions, and flight route weather disturbances, the most common uncertainty factors in aviation operation and maintenance scenarios are fully simulated, making the training scenario of the reinforcement learning model highly matched with the real engineering scenario, and further improving the practical applicability of the assignment strategy after training.

[0020] The bidirectional interaction interface is developed based on the Alpyne library and includes a state output interface, an action input interface, and a reward feedback interface. The state output interface abstracts the real-time state of the simulation environment into a standardized vector and outputs it to the reinforcement learning model. The action input interface receives the assigned actions output by the reinforcement learning model and converts them into execution instructions that the simulation environment can recognize. The reward feedback interface feeds back the operation and maintenance indicators of the simulation environment related to the reward value to the reinforcement learning model.

[0021] Therefore, a bidirectional interactive interface developed based on the Alpyne library enables seamless engineering integration between the AnyLogic simulation environment and the Python-based reinforcement learning model, solving the problem of low integration between simulation and reinforcement learning. Through three types of core interfaces, real-time closed-loop linkage of state, action, and reward is achieved, ensuring the real-time performance and accuracy of data interaction during reinforcement learning model training, and realizing a seamless transition between simulation training, policy verification, and engineering deployment.

[0022] The state output interface abstracts the real-time state of the simulation environment into a high-dimensional standardized vector and outputs it to the reinforcement learning model.

[0023] Therefore, by abstracting the real-time state of the simulation environment using high-dimensional standardized vectors, the core state information of fleet operation and maintenance can be fully covered while effectively controlling the state space dimension, avoiding the state space explosion problem in reinforcement learning models, and improving model training efficiency and convergence stability.

[0024] The reward function of the reinforcement learning model is set with multiple objective dimensions, which cover task completion rate, resource utilization rate, scheduling cost, maintenance interval balance, and long-term fleet availability.

[0025] The reward function of the reinforcement learning model is set to cover multiple objective dimensions such as task completion rate, resource utilization rate, scheduling cost, maintenance interval balance, and long-term fleet availability.

[0026] Therefore, by setting a reward function with multiple objectives, we can achieve multi-objective collaborative optimization of task completion rate, resource utilization rate, scheduling cost, maintenance interval balance, and long-term fleet availability. This solves the problem of imbalance in multi-objective optimization in existing technologies, balances the short-term task execution and long-term operation and maintenance support of the fleet, avoids overuse or over-maintenance of aircraft, and improves the operation and maintenance efficiency of the entire fleet life cycle.

[0027] The assignment decision scheme is evaluated quantitatively in multiple dimensions, and the evaluation results that fail to meet the preset optimization target are fed back to the multi-agent simulation environment and the reinforcement learning model to drive the iterative optimization of the assignment strategy.

[0028] Therefore, by using a multi-dimensional quantitative evaluation and reverse feedback mechanism for assignment decision-making schemes, the closed-loop iterative optimization of assignment strategies can be achieved, continuously improving the optimization effect of assignment decision-making schemes and ensuring that the schemes always match the actual needs and optimization goals of fleet operation and maintenance.

[0029] The present invention also provides a predictive assignment decision optimization system for flight missions of an aircraft fleet, comprising a fleet operation and maintenance element data layer, a multi-dimensional collaborative constraint layer, a multi-agent simulation layer, a reinforcement learning decision layer, and an assignment decision output layer that are connected bidirectionally in sequence; The fleet operation and maintenance element data layer is used to collect and standardize the raw data of fleet operation and maintenance from all dimensions, and output a fleet operation and maintenance dataset in a unified format. The multi-dimensional collaborative constraint layer has built-in multi-dimensional constraint rules for integrated task maintenance, which are used to synchronously embed the multi-dimensional constraint rules into the simulation logic of the multi-agent simulation layer and the reward and punishment mechanism of the reinforcement learning decision layer. The multi-agent simulation layer has a built-in interaction interface module, which enables closed-loop two-way data interaction with the reinforcement learning decision layer on state data, assigned actions, and operation and maintenance parameters related to rewards. The reinforcement learning decision layer has a built-in deep reinforcement learning model, which is used to generate the optimal assignment strategy after closed-loop training. The state space of the deep reinforcement learning model includes aircraft fault prediction and health management data. The assignment decision output layer is used to generate an executable assignment decision scheme based on the optimal assignment strategy.

[0030] Therefore, through the standardized design of the five-layer architecture, each module is connected bidirectionally through standardized interfaces, realizing modular and standardized control of the entire fleet assignment decision-making process. This improves the system's generalizability and deployability. When the fleet size is adjusted or the scenario is switched, only the parameters and weights of the corresponding modules need to be adjusted, without reconstructing the system, which greatly reduces the cost of scenario adaptation. The layers work together, with the data layer providing unified data support, the constraint layer providing unified rule boundaries, the simulation layer providing a high-fidelity training and verification environment, the reinforcement learning layer realizing intelligent decision generation, and the output layer realizing solution implementation and iterative optimization. Together, they solve the core technical problems of fleet assignment decision-making and achieve predictive and intelligent assignment decision optimization. Attached Figure Description

[0031] Figure 1 This is an overall architecture diagram of the predictive assignment decision optimization system for flight missions of an aircraft fleet involved in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the implementation of the predictive assignment decision optimization method for aircraft fleet flight missions according to an embodiment of the present invention. Detailed Implementation

[0032] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the following description, the same reference numerals are used for the same parts, and repeated descriptions are omitted. Furthermore, the drawings are merely schematic diagrams, and the proportions of the parts or the shapes of the parts may differ from the actual figures.

[0033] This embodiment uses a fleet of 20-24 aircraft and an average of 21 flights per week as an application scenario to specifically implement the method and system of the present invention.

[0034] The predictive assignment decision optimization method for aircraft fleet flight missions provided by this invention addresses the common technical problems of poor dynamic adaptability, multi-objective imbalance, disconnect between simulation and learning, and weak generalization in existing aircraft fleet assignment decision-making technologies through the deep synergy of four core features: a multi-dimensional collaborative constraint system, a high-fidelity multi-agent simulation environment, an engineered closed-loop interaction between simulation and reinforcement learning, and reinforcement learning decision-making incorporating PHM data. Specifically, the multi-dimensional collaborative constraint system provides a unified rule boundary for the simulation environment and the reinforcement learning model, avoiding a disconnect between decision-making and actual constraints; the multi-agent simulation environment provides a high-fidelity training and verification platform for the reinforcement learning model, simulating the uncertainties of real-world scenarios; the closed-loop bidirectional interactive interface achieves engineered coupling between simulation and reinforcement learning, ensuring consistency between the training process and real-world scenarios; and the reinforcement learning model incorporating PHM data enables predictive decision-making, balancing short-term tasks and long-term maintenance. These four features support and are coupled together to form an organic whole.

[0035] The predictive assignment decision optimization method for aircraft fleet flight missions provided by this invention has a core implementation process corresponding to... Figure 2 The complete process steps are as follows: The first step is to construct a multi-dimensional collaborative constraint system integrating task maintenance, corresponding to... Figure 2 Step S2. In this embodiment, the multi-dimensional collaborative constraint system includes time constraint rules, resource constraint rules, aircraft status constraint rules, priority constraint rules, and task-maintenance cross-constraint rules. Among them, time constraint rules are used to limit the time window for task execution and the maintenance window for aircraft maintenance; resource constraint rules are used to limit the capacity and availability of maintenance bays, personnel, and spare parts; aircraft status constraint rules are used to limit the task execution capability of aircraft in different health states; priority constraint rules are used to limit the execution priority of different levels of tasks and maintenance with different urgency levels; and task-maintenance cross-constraint rules are used to coordinate the temporal relationship between task execution and maintenance plans, avoiding time conflicts between tasks and maintenance. All the above constraint rules are synchronously embedded into the simulation logic of the simulation environment and the reward and punishment mechanism of the reinforcement learning model to ensure that the simulation process and decision generation follow unified constraint rules, fundamentally avoiding the problem that the decision scheme does not conform to the actual operation and maintenance constraints.

[0036] The second step is to construct a multi-agent simulation environment and set up a two-way interaction interface, corresponding to... Figure 2 Steps S3 and S4 are described. In this embodiment, a multi-agent simulation environment is built based on the AnyLogic platform to simulate the entire process of real fleet assignment and maintenance. At the same time, a two-way interactive interface is developed based on the Alpyne library. Through this interface, a closed-loop two-way data interaction is realized between the multi-agent simulation environment and the reinforcement learning model, including state data output, assignment action input, and reward value feedback. This makes the simulation environment a real-time training and verification carrier for the reinforcement learning model, realizing the engineering coupling of simulation and intelligent decision-making.

[0037] The third step is to complete the training of the reinforcement learning model and generate assignment decision schemes, corresponding to... Figure 2 Steps S5, S6, and S7 are described. In this embodiment, the state space of the reinforcement learning model includes aircraft fault prediction and health management data, specifically covering aircraft model, service life, cumulative flight hours, current health status, historical maintenance records, and remaining service life prediction data, enabling the model to make decisions based on the future health trend of the aircraft. Based on the aforementioned multi-dimensional collaborative constraint system, multi-agent simulation environment, and closed-loop bidirectional data interaction, a multi-scenario training process is designed. After completing the model training, an executable flight mission assignment decision scheme for the aircraft fleet is generated.

[0038] In this embodiment, its specific implementation corresponds to Figure 2 The S1 step, and Figure 1The fleet operation and maintenance element data layer is constructed by first collecting raw data on the airline's fleet operation and maintenance across all dimensions, including aircraft attribute data, flight mission information data, maintenance resource data, spare parts resource data, and environmental disturbance data. All raw data undergoes standardized cleaning, missing value imputation, and consistency verification to remove outliers. Data formats and dimensions are standardized to establish a unified fleet operation and maintenance data model, stored in Excel spreadsheets and structured databases. A standardized fleet operation and maintenance dataset is output through a standardized interface, providing a unified and standardized data foundation for subsequent constraint system construction, simulation environment setup, and reinforcement learning model training.

[0039] In this embodiment, its specific implementation corresponds to Figure 1 The constraint conflict resolution submodule is part of the multi-dimensional collaborative constraint layer. Within the multi-dimensional collaborative constraint system, a constraint conflict resolution submodule is configured with a preset priority order for constraint rules, from highest to lowest: aircraft state constraints related to flight safety, time constraints related to mission compliance, resource constraints related to resource availability, priority constraints related to mission execution, and cross-constraint rules related to operation and maintenance optimization. When multiple constraint rules conflict, the constraint conflict resolution submodule automatically coordinates according to the preset priority, prioritizing the satisfaction of higher-priority constraints and automatically adjusting the execution plan for lower-priority constraints. This avoids unexecutable decision-making schemes due to constraint conflicts, achieving automated resolution of constraint conflicts without manual intervention.

[0040] In this embodiment, its specific implementation corresponds to Figure 2 The S3 step, and Figure 1 The AnyLogic multi-agent simulation layer is used. The multi-agent simulation environment is built on the AnyLogic platform and includes a multi-agent modeling module and an uncertainty perturbation generation module. The multi-agent modeling module constructs all roles of the fleet maintenance agents and their corresponding interaction rules, specifically including aircraft agents (20 aircraft), mission command agents, maintenance resource agents, spare parts warehouse agents, and workstation agents. It designs the behavioral logic and interaction rules for each agent, such as the standby, mission execution, maintenance, and standby state transition logic for aircraft agents; the task allocation instruction issuance logic for mission command agents; and the workstation occupation and release logic for maintenance resource agents. Animation models are added to each agent to visualize the simulation process. The uncertainty perturbation generation module embeds multi-source random perturbations into the simulation process based on historical maintenance data and aviation expert knowledge, simulating uncertainties in real maintenance scenarios. This allows the reinforcement learning model to fully adapt to the dynamic environment during training, improving the robustness of the assignment strategy.

[0041] In this embodiment, the specific implementation corresponds to the functional implementation of the uncertainty disturbance generation module. Multi-source random disturbances include sudden aircraft malfunctions, emergency mission insertions, and route weather disturbances. Sudden aircraft malfunctions are randomly generated using AnyLogic's event function, such as engine failures or avionics system failures, with a failure rate set to 5%. Emergency mission insertions include temporarily added flight missions and emergency support missions, with an insertion rate set to 10%. Route weather disturbances include flight segment closures and flight delays caused by heavy rain, thunderstorms, and strong winds, with an occurrence rate set to 8%. These three types of disturbances cover the most common uncertainty factors in civil aviation operation and maintenance scenarios, ensuring a high degree of matching between the simulation environment and real engineering scenarios, and guaranteeing that the trained strategies can be directly implemented and applied.

[0042] In this embodiment, its specific implementation corresponds to Figure 2 The S4 step, and Figure 1 The AnyLogic multi-agent simulation layer includes a simulation reinforcement learning interaction interface module. This bidirectional interaction interface, developed using the Alpyne library, comprises three core modules: a state output interface, an action input interface, and a reward feedback interface. The state output interface abstracts the real-time state of the simulation environment, including aircraft state, task state, resource state, and environmental disturbance state, into standardized vectors and outputs them to the reinforcement learning model in Python. The action input interface receives the assigned actions (task aircraft allocation instructions) output by the Python reinforcement learning model, converts them into an instruction format recognizable by the simulation system, and sends them to the task command agent for execution. The reward feedback interface converts real-time operational metrics from the simulation environment, including task completion rate, delay time, and resource utilization, into corresponding reward values ​​and feeds them back to the reinforcement learning model in real-time for policy updates. Simultaneously, the AnyLogic simulation system is packaged into Java code, and the simulation environment can be invoked, paused, stepped, and terminated in Python using the Alpyne library, achieving a fully closed-loop real-time linkage between the simulation environment and the reinforcement learning model.

[0043] In this embodiment, the specific implementation corresponds to the functionality of the state output interface. The state output interface abstracts the real-time state of the simulation environment into a 128-dimensional standardized vector and outputs it to the reinforcement learning model. This 128-dimensional vector covers four core states: aircraft health status, mission attributes, resource availability, and disturbance information. While fully covering all elements of fleet operation and maintenance state information, it effectively controls the dimensionality of the state space, avoids the problem of state-action space explosion in the reinforcement learning model, and significantly improves the training efficiency and convergence stability of the model.

[0044] In this embodiment, its specific implementation corresponds to Figure 1The reward function module in the reinforcement learning decision layer is designed to address multiple objectives, including task completion rate, resource utilization, scheduling cost, maintenance interval balance, and long-term fleet availability. Each objective dimension is assigned a corresponding reward and penalty. Specifically, task completion rate corresponds to a positive reward, while incomplete tasks are penalized. Reasonable resource utilization is rewarded, while idle or overloaded resources are penalized. Scheduling cost below a threshold is rewarded, while exceeding the threshold is penalized. Maintenance interval balance meets preset requirements, while excessively concentrated or dispersed maintenance intervals are penalized. Maintaining long-term fleet availability at the target level is rewarded, while decreased availability is penalized. This multi-objective reward function design guides the reinforcement learning model to achieve collaborative optimization of multiple objectives, balancing short-term task execution with long-term maintenance support, and avoiding the problem of prioritizing short-term tasks over long-term maintenance.

[0045] In this embodiment, its specific implementation corresponds to Figure 2 The S8 steps, and Figure 1 The assignment decision output layer is constructed using a multi-dimensional quantitative evaluation system. This system evaluates the effectiveness of generated assignment decision schemes from multiple dimensions, including task completion rate, average delay time, on-time rate of high-priority tasks, resource utilization, maintenance interval balance, and fleet availability. Preset optimization target thresholds for each dimension. If the scheme evaluation result fails to meet the preset optimization target, the evaluation result is fed back to the multi-agent simulation environment and reinforcement learning model for retraining and policy optimization. This drives continuous iterative updates of the assignment strategy until the scheme meets the preset optimization target.

[0046] This invention also provides a predictive assignment decision optimization system for aircraft fleet flight missions, which is implemented as follows: This system comprises a fleet operation and maintenance element data layer, a multi-dimensional collaborative constraint layer, an AnyLogic multi-agent simulation layer, a reinforcement learning decision layer, and an assignment decision output layer. The five layers are interconnected in both directions, allowing for bidirectional data flow and forming an organic whole.

[0047] The fleet operation and maintenance element data layer serves as the underlying support for the system, providing data support for the entire system. It includes aircraft attribute modules, mission information modules, maintenance resource modules, spare parts resource modules, and environmental data modules. Each module integrates data through standardized data interfaces, outputting fleet operation and maintenance datasets in a unified format to the multi-dimensional collaborative constraint layer and the AnyLogic multi-agent simulation layer. Specifically, the aircraft attribute module stores data such as aircraft model, service life, cumulative flight hours, current status, and historical maintenance records; the mission information module stores data such as mission number, type, priority, time window, location, and aircraft requirements; the maintenance resource module stores data such as the quantity, status, and utilization rate of maintenance workstations, personnel, and equipment; the spare parts resource module stores data such as spare parts inventory, replenishment cycle, and unit cost; and the environmental data module stores data related to uncertainties such as emergencies.

[0048] The multi-dimensional collaborative constraint layer serves as the intermediate support layer of the system, providing rule constraints for simulation and decision-making. These constraints include time constraints, resource constraints, aircraft state constraints, priority constraints, and task / maintenance cross-constraints. These constraint rules are embedded into the simulation logic of the AnyLogic multi-agent simulation layer and the reward and punishment mechanism of the reinforcement learning decision layer. Each constraint module implements rule constraints such as maintenance window, workstation capacity, aircraft execution capability, task / maintenance priority, and task / maintenance timing coordination. A constraint conflict resolution submodule is also set up to automatically coordinate multiple constraints according to preset priorities when conflicts occur.

[0049] The AnyLogic multi-agent simulation layer serves as the core simulation environment of the system, providing a training and validation platform for reinforcement learning. Built upon AnyLogic simulation software, it includes a multi-agent modeling module, an uncertainty perturbation generation module, a simulation reinforcement learning interaction interface module, and a data log and indicator statistics module. It works in conjunction with a multi-dimensional collaborative constraint layer to implement constraint rules in simulation. Through the interaction interface module, it achieves closed-loop data interaction with the reinforcement learning decision layer, while simultaneously feeding back simulation logs and indicator data to the assignment decision output layer. The multi-agent modeling module constructs agents and agent groups such as aircraft, mission command, maintenance personnel, spare parts warehouse, and airfield, and designs... The behavioral logic and interaction rules of each intelligent agent simulate the real fleet operation and dispatch process; the uncertainty disturbance generation module generates multi-source random disturbances such as sudden failures, emergency task insertions, and weather effects based on historical data and domain expert knowledge, and embeds them into the simulation process; the simulation reinforcement learning interaction interface module is developed based on the Alpyne library, realizing the state output of the simulation environment, the action input of reinforcement learning agents, and operation and maintenance indicators related to rewards, which is the core of the engineering coupling of simulation and reinforcement learning; the data log and indicator statistics module records data such as task allocation, resource consumption, and delay time in real time during the simulation process, and statistically analyzes core indicators such as task completion rate and resource utilization rate.

[0050] The reinforcement learning decision layer is the core of the system's intelligent decision-making, learning the optimal assignment strategy. It includes a state / action space design module, a reinforcement learning model module, a reward function module, and a model training and optimization module. Through an interactive interface module, it acquires state data from the simulation layer, outputs assigned actions to the simulation layer, and obtains the optimal assignment strategy after training in multiple scenarios, which is then pushed to the assignment decision output layer. Specifically, the state space design module incorporates aircraft fault prediction and health management data into the state space to ensure the predictability of the decision; the reward function module covers multiple objective optimization dimensions to achieve a balance between short-term tasks and long-term maintenance; and the model training and optimization module designs multi-scenario training processes to ensure the model's convergence and robustness.

[0051] The assignment decision output layer is the top layer of the system, realizing the output and iterative optimization of decision schemes. It includes an assignment scheme generation module and a reinforcement learning model access module. It obtains the optimal assignment strategy from the reinforcement learning decision layer, generates executable flight mission assignment schemes, and evaluates the scheme effects from multiple dimensions. If the optimization goal is not achieved, the evaluation results are fed back to the simulation layer and the reinforcement learning layer to realize iterative updates of the strategy. The reinforcement learning model access module connects the reinforcement learning model trained by the decision layer into the system. The assignment scheme generation module outputs executable files such as mission aircraft allocation tables, maintenance plan coordination tables, and resource allocation schemes for direct use by assignment and dispatch personnel.

[0052] The system's five-layer architecture works in synergy. The fleet operation and maintenance element data layer provides a standardized data foundation for the multi-dimensional collaborative constraint layer and the AnyLogic multi-agent simulation layer. The multi-dimensional collaborative constraint layer provides unified constraint rules and reward / penalty boundaries for the AnyLogic multi-agent simulation layer and the reinforcement learning decision layer. The AnyLogic multi-agent simulation layer and the reinforcement learning decision layer achieve closed-loop bidirectional data flow of state, action, and reward through an interactive interface, realizing the engineering coupling of simulation and reinforcement learning. The reinforcement learning decision layer pushes the optimal strategy to the assignment decision output layer, which in turn feeds back the evaluation results to the AnyLogic multi-agent simulation layer and the reinforcement learning decision layer, enabling iterative optimization of the entire system and jointly achieving predictive and intelligent assignment decision optimization for aircraft fleet flight missions.

[0053] The embodiments described above do not constitute a limitation on the scope of protection of this technical solution. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the above embodiments should be included within the scope of protection of this technical solution.

Claims

1. An aircraft fleet flight mission predictive assignment decision optimization method, characterized by, Includes the following steps: A multi-dimensional collaborative constraint system integrating mission maintenance is constructed. The multi-dimensional collaborative constraint system includes time constraint rules, resource constraint rules, aircraft state constraint rules, priority constraint rules, and mission maintenance cross-constraint rules. The multi-dimensional collaborative constraint system is synchronously embedded into the simulation logic of the simulation environment and the reward and punishment mechanism of the reinforcement learning model. A multi-agent simulation environment is constructed, and a two-way interaction interface is set up. Through the two-way interaction interface, a closed-loop two-way data interaction is realized between the multi-agent simulation environment and the reinforcement learning model, including state data output, assigned action input, and reward value feedback. Based on the multi-dimensional collaborative constraint system, the multi-agent simulation environment, and the closed-loop bidirectional data interaction, the training of the reinforcement learning model is completed. The state space of the reinforcement learning model includes aircraft fault prediction and health management data, and an assignment decision scheme is generated.

2. The method according to claim 1, characterized in that, The collected raw fleet operation and maintenance data are standardized, cleaned, missing values ​​are filled, and consistency is verified to establish a unified fleet operation and maintenance data model. The fleet operation and maintenance dataset is output in a unified format through a standardized interface.

3. The method according to claim 1, characterized in that, The multi-dimensional collaborative constraint system is also equipped with a constraint conflict resolution sub-module, which automatically coordinates multiple constraint rules according to preset priorities when conflicts occur.

4. The method according to claim 1, characterized in that, The multi-agent simulation environment is built on the AnyLogic platform. The multi-agent simulation environment has a built-in multi-agent modeling module and an uncertainty disturbance generation module. The multi-agent modeling module constructs all roles of fleet operation and maintenance agents and corresponding interaction rules. The uncertainty disturbance generation module embeds multi-source random disturbances during the simulation process.

5. The method according to claim 4, characterized in that, The multi-source random disturbances include sudden aircraft malfunctions, emergency mission insertions, and weather disturbances along flight routes.

6. The method according to claim 1, characterized in that, The bidirectional interaction interface is developed based on the Alpyne library. The bidirectional interaction interface includes a state output interface, an action input interface, and a reward feedback interface. The state output interface abstracts the real-time state of the simulation environment into a standardized vector and outputs it to the reinforcement learning model. The action input interface receives the assigned actions output by the reinforcement learning model and converts them into execution instructions that the simulation environment can recognize. The reward feedback interface feeds back the operation and maintenance indicators related to rewards of the simulation environment to the reinforcement learning model.

7. The method according to claim 6, characterized in that, The state output interface abstracts the real-time state of the simulation environment into a high-dimensional standardized vector and outputs it to the reinforcement learning model.

8. The method according to claim 1, characterized in that, The reward function of the reinforcement learning model is set with multiple objective dimensions, which cover task completion rate, resource utilization rate, scheduling cost, maintenance interval balance, and long-term fleet availability.

9. The method according to claim 1, characterized in that, The assignment decision scheme is quantitatively evaluated in multiple dimensions, and the evaluation results that fail to meet the preset optimization target are fed back to the multi-agent simulation environment and the reinforcement learning model to drive the iterative optimization of the assignment strategy.

10. A predictive assignment decision optimization system for flight missions of an aircraft fleet, characterized in that, It includes a fleet operation and maintenance element data layer with sequential bidirectional data connectivity, a multi-dimensional collaborative constraint layer, a multi-agent simulation layer, a reinforcement learning decision layer, and an assignment decision output layer; The fleet operation and maintenance element data layer is used to collect and standardize the raw data of fleet operation and maintenance from all dimensions, and output a fleet operation and maintenance dataset in a unified format. The multi-dimensional collaborative constraint layer has built-in multi-dimensional constraint rules for integrated task maintenance, which are used to synchronously embed the multi-dimensional constraint rules into the simulation logic of the multi-agent simulation layer. The multi-agent simulation layer has a built-in interactive interface module, which enables closed-loop two-way data interaction between the reinforcement learning decision layer and the state data, assigned actions, and reward values. The reinforcement learning decision layer has a built-in deep reinforcement learning model, which is used to generate the optimal assignment strategy after closed-loop training. The state space of the deep reinforcement learning model includes aircraft fault prediction and health management data. The assignment decision output layer is used to generate an executable assignment decision scheme based on the optimal assignment strategy.

Images