Machine position distribution system and method based on vertical field light-weight large model
By adopting a stand allocation method based on a lightweight, large-scale model in a vertical domain, the challenges of airport stand allocation systems in handling complex data and multi-objective optimization have been solved. This has improved airport operational efficiency and service quality, provided an intelligent stand allocation solution that adapts to different operational scenarios and emergencies, and enhanced the system's flexibility and human-machine collaboration experience.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-29
- Publication Date
- 2026-04-07
AI Technical Summary
Existing airport gate allocation systems cannot effectively handle massive, multi-source, and dynamically changing airport operation data. They struggle to make reasonable gate allocation decisions when faced with complex situations such as flight schedule adjustments, sudden weather changes, and equipment failures. Furthermore, they lack comprehensive consideration of various influencing factors, making it difficult to balance multiple optimization objectives.
A lightweight, large-model-based gate allocation method based on vertical domains is adopted. Historical flight data is trained using a convolutional neural network (CNN) module and a Transformer architecture. Supervised learning and reinforcement learning are combined to design a multi-task output layer. Gate allocation schemes are generated through a hierarchical hybrid algorithm and the NSGA-II multi-objective evolutionary algorithm. The schemes are then optimized and adjusted through a human-computer interaction module.
It enables intelligent and efficient allocation of aircraft stands, improves the overall operational efficiency and service quality of airports, enhances the level of intelligence in airport operations, can quickly process massive amounts of data, accurately predict flight and aircraft stand demand, reduce manual intervention, enhance the flexibility and adaptability of the system, and improve the human-machine collaboration experience.
Smart Images
Figure CN121809897A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of airport operation and management technology, and in particular relates to a gate allocation system and method based on a lightweight large model in a vertical domain. Background Technology
[0002] In current airport operations management, gate allocation is a complex and critical task. Traditional gate allocation systems are mostly based on rules or simple algorithms, which have the following significant drawbacks: First, they cannot effectively handle massive, multi-source, and dynamically changing airport operations data. When faced with complex situations such as flight schedule adjustments, sudden weather changes, and equipment failures, they struggle to make quick and reasonable gate allocation decisions. Second, they lack comprehensive consideration of multiple influencing factors, making it difficult to balance multiple optimization objectives, such as maximizing gate utilization, minimizing passenger walking distances, and minimizing aircraft taxiing time.
[0003] With the development of artificial intelligence technology, lightweight large models in vertical domains have achieved remarkable results in fields such as natural language processing and image recognition. However, research and practice on applying lightweight large models in vertical domains to airport gate allocation are still relatively limited, and existing gate allocation systems cannot fully utilize the advantages of lightweight large models in vertical domains to solve complex practical problems. Summary of the Invention
[0004] To overcome the problems existing in related technologies, the present invention discloses an embodiment of a gate allocation system and method based on a lightweight large model in a vertical domain. The purpose of the present invention is to improve the intelligence and efficiency of the airport gate allocation system and optimize the overall operational efficiency of the airport by utilizing the powerful data processing, analysis and reasoning, and interactive experience capabilities of the lightweight large model in a vertical domain.
[0005] The technical solution is as follows: a camera allocation method based on a lightweight, large-scale model in a vertical domain, including the following steps: S1 collects real-time and historical flight data from various data sources at the airport and preprocesses the collected data; the various data sources include data from the flight information system, meteorological system, and equipment and facilities management system. S2 utilizes a lightweight large model in the vertical domain to train preprocessed historical flight data. The lightweight large model in the vertical domain adopts a Transformer architecture that incorporates a convolutional neural network (CNN) module. For the airport gate allocation scenario, the structure of the lightweight large model in the vertical domain is customized, and multi-task output is designed in the output layer of the lightweight large model in the vertical domain according to the gate allocation task. S3, based on the analysis results of the lightweight large model in the vertical domain, combined with the airport's business rules and actual situation, completes the balance optimization of multiple objectives in the complex airport operation environment, and uses a hierarchical hybrid algorithm to generate a preliminary gate allocation scheme. S4 optimizes and adjusts the generated preliminary aircraft station allocation plan, and promptly provides feedback and displays the adjustment results.
[0006] In step S2, the preprocessed historical flight data is trained using a lightweight, large model in the vertical domain. The training data includes the historical arrival and departure times of flights, gate usage, passenger flow, delay time, and weather conditions. During the training process, a combination of supervised learning and reinforcement learning is used to continuously optimize the model parameters.
[0007] Furthermore, the customized design of the lightweight large model structure for vertical domains includes: designing a multimodal data fusion module in the input layer of the lightweight large model for vertical domains to effectively fuse different types of data; In the output layer of the lightweight large model in the vertical domain, multi-task output is designed according to the task of the gate allocation. For flights with insufficient transit time, the shortest taxiing time gate is predicted, the best gate for normal flights is predicted, the estimated taxiing time is predicted, and the conflict risk is predicted. After training, the large model is deployed.
[0008] In step S3, a hierarchical hybrid algorithm is used to generate a preliminary station allocation scheme, including: First, constraint-based programming is used to initially screen the gate allocation schemes to ensure that the generated schemes meet the airport's basic business rules, including matching gate type with aircraft type, special requirements of flights, and meeting safety time interval requirements. During the generation process, the priority of flights is taken into account to obtain suitable gate resources. Then, the NSGA-II multi-objective evolutionary algorithm is introduced to optimize the preliminary selected schemes by maximizing the docking rate, minimizing the passenger walking distance, minimizing the aircraft taxiing time, and maximizing the gate turnover rate.
[0009] Furthermore, the initially screened solutions were further optimized in depth, including: Step 1: Problem definition and variable initialization; Define a flight set: Define the set of aircraft positions: For each flight Create variables This indicates the assigned machine position, and sets its field to "All Machine Positions". ; Collect input data: flight arrival time Departure time Aircraft size Machine location dimensions And location attributes including type or neighbor relationships; Step 2, specify constraints; List all hard constraints including time conflicts and size matching, and transform them into logical expressions for CP processing; apply a propagation mechanism to each constraint: before searching, reduce the domain of variables, and if the size of a gate is less than a set threshold, remove the gate from the domain of the relevant flight. Step 3, constraint solution and propagation; Constraint propagation is performed using the CP solver: constraints are iteratively checked, and the propagation domain is reduced to decrease the search space; if propagation results in an empty domain for any variable, the current partial solution is infeasible and is immediately discarded; the search process employs depth-first search (DFS) or finite backtracking to generate candidate assignment schemes, including tentative assignments. And propagate constraints to verify feasibility; Step 4: Output feasible solutions; Output the set of all feasible machine station allocation schemes that satisfy all constraints. Each solution is represented as a variable assignment vector. If there are multiple feasible solutions, the screening process ends; if there are no feasible solutions, the problem is reported as infeasible.
[0010] In step 2, the time conflict constraints include: If two flights have overlapping times, they cannot be assigned the same gate position. The expression is: ; In the formula, For the first One flight, For the first One flight, For flights, conditions Used to determine overlapping time intervals; In step 2, the machine position size matching constraints include: The aircraft size must be less than or equal to the parking space capacity, expressed as: ; In the formula, Indicates the camera position Capacity size.
[0011] Furthermore, the NSGA-II multi-objective evolutionary algorithm is introduced, with the optimization objectives of maximizing docking bridge rate, minimizing passenger walking distance, minimizing aircraft taxiing time, and maximizing gate turnaround rate, to perform in-depth optimization of the initially screened schemes; the specific steps include: Step 1, Problem Modeling and Coding; Decision variables: Encoded as integers for each flight Assigned to machine station , recorded as Population initialization: randomly generated A feasible solution is found, ensuring that the basic constraints are satisfied; Step 2, calculate the objective function; Four target values are calculated for each individual, and the bridge berth rate is maximized as follows: ; In the formula, To maximize bridge utilization, For the set of machine positions, For each flight Assigned to machine station The value, For the number of flights; Minimize the passenger walking distance as follows: ; In the formula, To minimize the walking distance for passengers, For camera position Distance to the center of the terminal building; The shortest taxiing time for an aircraft is represented as: ; In the formula, The shortest taxiing distance for an aircraft. From the runway to the parking position Average glide time; The highest machine turnover rate is represented as: ; In the formula, This represents the highest machine turnover rate. For camera position Effective usage time, For turnover cycle; Step 3, constraint handling; The expression for hard constraint gate-flight type matching is: ; In the formula, For flights Available machine location set; The expression for time-free operation is: ; In the formula, To ensure that flight times do not overlap, For aircraft size, For departure time, For the first Aircraft size for each flight For the first Departure time of each flight; Soft constraints, transit time: ; In the formula, For transit connection time, This is the minimum transit time.
[0012] Furthermore, the initially screened schemes are further optimized, including by continuously adjusting the parameters of the reinforcement learning algorithm based on real-time operational feedback, dynamically adapting to changes in airport operational status, and optimizing the allocation scheme.
[0013] In step S4, the generated preliminary aircraft stand allocation plan is optimized and adjusted, including: The system further optimizes and adjusts the initial gate allocation plan based on the input flight-specific requirements or feedback information in the user interface. This includes: real-time monitoring of airport operations, detecting gate conflicts, and initiating conflict detection and resolution mechanisms for optimization and adjustment; using a lightweight, large-scale model in the vertical domain based on historical conflict resolution experience and real-time data to model and analyze conflict situations, locating conflicting flights and gates; generating multiple conflict resolution strategies through deep learning algorithms and simulating and predicting the implementation effect of each strategy; selecting the optimal strategy, automatically adjusting the gate allocation plan, and providing timely feedback on the adjustment results.
[0014] Another objective of this invention is to provide a camera position allocation system based on a lightweight, large-scale model within a vertical domain. This system implements the aforementioned camera position allocation method based on a lightweight, large-scale model within a vertical domain. The system includes: The data acquisition and preprocessing module is used to collect real-time and historical flight data from various data sources at the airport, including flight information systems, meteorological systems, and equipment and facilities management systems, and to preprocess the collected data. The core processing module of the vertical domain lightweight large model is used to train the preprocessed historical flight data using the vertical domain lightweight large model. The vertical domain lightweight large model adopts the Transformer architecture with the integration of the convolutional neural network (CNN) module. For the airport gate allocation scenario, the structure of the vertical domain lightweight large model is customized, and multi-task output is designed in the output layer of the vertical domain lightweight large model according to the gate allocation task. The decision generation and optimization module is used to perform multi-objective balance optimization in the complex airport operation environment based on the analysis results of the lightweight large model in the vertical domain, combined with the airport's business rules and actual conditions, and to generate a preliminary gate allocation scheme using a hierarchical hybrid algorithm. The human-computer interaction module is used to optimize and adjust the generated preliminary camera position allocation plan, and to provide timely feedback and display of the adjustment results.
[0015] Furthermore, the aforementioned camera station allocation system based on a lightweight, large-scale vertical domain model is mounted on a computer device, which includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor. When the processor executes the computer program, it implements the functions of the aforementioned camera station allocation system based on a lightweight, large-scale vertical domain model.
[0016] Furthermore, the aforementioned station allocation system based on a lightweight, large-scale model of a vertical domain is mounted on a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can realize the functions of the aforementioned station allocation system based on a lightweight, large-scale model of a vertical domain.
[0017] Furthermore, the aforementioned station allocation system based on a lightweight, large-scale model of a vertical domain is mounted on an information data processing terminal. This information data processing terminal is used to provide a user input interface to implement the functions of the aforementioned station allocation system based on a lightweight, large-scale model of a vertical domain when executed on an electronic device.
[0018] Combining all the above technical solutions, the beneficial effects of this invention are as follows: it takes the lead in realizing the integrated development of artificial intelligence and civil aviation operations, and is an intelligent product and application with core competitiveness. As an auxiliary decision-making tool for intelligent optimization, it can effectively improve the level of intelligence in civil aviation airport operation support.
[0019] This invention enables intelligent and efficient allocation of aircraft parking positions through in-depth mining and analysis of multi-source data using a lightweight, large-scale model in a vertical domain, thereby improving the overall operational efficiency and service quality of airports.
[0020] The target of this invention is the top 40 domestic civil aviation airports with passenger throughput of 10 million or more. The commercial value is calculated as follows: the initial penetration rate reaches 25%, the price per airport is 2 million yuan, and the expected sales revenue in the first year is 20 million yuan. In the third year, the penetration rate increases to 50%, and the sales revenue exceeds 40 million yuan. Attached Figure Description
[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure; Figure 1 This is a flowchart of the camera position allocation method based on a lightweight large model in a vertical domain provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a machine position allocation system based on a lightweight large model in a vertical domain, provided in an embodiment of the present invention. Detailed Implementation
[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0023] The innovation of this invention lies in the fact that it utilizes the powerful data processing, analysis and reasoning, and interactive experience capabilities of a lightweight, large-scale model in a vertical domain to improve the intelligence and efficiency of the airport gate allocation system and optimize the overall operational efficiency of the airport.
[0024] Example 1, as Figure 1 As shown, a camera allocation method based on a lightweight, large-scale model in a vertical domain includes: S1 collects real-time and historical flight data from various data sources at the airport and preprocesses the collected data; the various data sources include data from the flight information system, meteorological system, and equipment and facilities management system. S2 utilizes a lightweight large model in the vertical domain to train preprocessed historical flight data. The lightweight large model in the vertical domain adopts a Transformer architecture that incorporates a convolutional neural network (CNN) module. For the airport gate allocation scenario, the structure of the lightweight large model in the vertical domain is customized, and multi-task output is designed in the output layer of the lightweight large model in the vertical domain according to the gate allocation task. S3, based on the analysis results of the lightweight large model in the vertical domain, combined with the airport's business rules and actual situation, completes the balance optimization of multiple objectives in the complex airport operation environment, and uses a hierarchical hybrid algorithm to generate a preliminary gate allocation scheme. S4 optimizes and adjusts the generated preliminary aircraft station allocation plan, and promptly provides feedback and displays the adjustment results.
[0025] Example 2, as Figure 2 As shown, the station allocation system based on a lightweight large model in a vertical domain provided in this embodiment of the invention includes a data acquisition and preprocessing module, a core processing module for a lightweight large model in a vertical domain, a decision generation and optimization module, and a human-computer interaction module.
[0026] The data acquisition and preprocessing module is responsible for collecting real-time and historical flight data from various data sources at the airport, including flight information systems, meteorological systems, and equipment and facilities management systems. It performs routine preprocessing operations such as cleaning, noise reduction, and normalization on the collected data to ensure data quality and consistency, providing a reliable data foundation for subsequent analysis and processing.
[0027] The core processing module of the lightweight, large-scale model for vertical domains adopts the Transformer architecture. As a cornerstone technology of modern large language models, the Transformer architecture, with its powerful self-attention mechanism, can effectively capture long-distance dependencies between different data, making it suitable for processing complex data in airport operations. To further improve model performance, a Convolutional Neural Network (CNN) module is integrated on top of the Transformer architecture. Leveraging the characteristics of CNN in efficient local feature extraction and translation invariance in spatial structure modeling, the combination of the two enables accurate modeling of "global semantics + local details," while optimizing computational efficiency to enhance the model's ability to extract local features.
[0028] The Transformer architecture is trained based on collected historical flight data, including historical arrival and departure times, gate availability, passenger flow, delay times, and weather conditions. During training, a combination of supervised learning and reinforcement learning is used to continuously optimize the model's parameters, improving the prediction accuracy and decision-making capabilities of this lightweight, large-scale model for vertical domains.
[0029] Furthermore, for airport gate allocation scenarios, a customized lightweight large-scale model structure for the vertical domain is designed. At the input layer of the lightweight large-scale model for the vertical domain, a data fusion module is designed to associate and fuse different types of data (including but not limited to flight dynamics information, gate occupancy status, runway usage status, passenger flow data, weather conditions, and equipment operating status). For example, it associates runway numbers with flights, flights with available gates, runway numbers with gates, and weather conditions with the operating speed of support vehicles to calculate support time.
[0030] In the output layer of the lightweight large model in the vertical domain, multi-task output is designed based on the task of gate allocation. This includes designing flight routes, aircraft types, arrival / departure ports, runway numbers, gate positions, and ground taxiing time matrices. For flights with insufficient turnaround time, the shortest taxiing time and gate position are predicted (by obtaining a set of gate positions from the ground taxiing time matrix and then selecting one from the set). For regular flights, the optimal gate position and estimated taxiing time are calculated (after gate allocation, based on the ground taxiing time matrix), along with potential conflict risks. After training, the large model is deployed to a high-performance server, using Docker for rapid deployment and migration. Simultaneously, Kubernetes is used for cluster management to achieve elastic scaling of the model service, handling high-concurrency requests during peak airport operation periods and providing core analysis and decision support for the system.
[0031] The decision generation and optimization module, based on the analysis results of a lightweight, large-scale model in the vertical domain, combined with the airport's business rules and actual conditions, including but not limited to flight dynamics, gate occupancy, runway usage status, passenger flow data, weather conditions, and equipment operating status, achieves multi-objective balanced optimization in the complex airport operation environment. The module employs a hierarchical hybrid algorithm to generate a preliminary gate allocation scheme. This includes: (1) First, the large model can generate an initial solution that meets the constraints based on the historical best solution and rule-based reasoning. This generates a preliminary gate allocation scheme, i.e., a high-quality initial population, and the generated scheme satisfies the airport's basic business rules, such as matching gate type with aircraft type, special flight requirements, and meeting safety time interval requirements. During the generation process, the special characteristics of flights are fully considered, prioritizing the allocation of special missions and international flights, such as emergency flights, live animal transport, VIP flights, and international flights, to ensure that special flights have priority access to gate resources.
[0032] For example, specifically including: Using constraint programming, preliminary screening of gate allocation schemes is conducted to ensure that the generated schemes meet the airport's basic business rules, including matching gate type with aircraft type, special requirements of flights, and meeting safety time interval requirements. During the generation process, appropriate gate resources are obtained by combining flight priorities. The initially screened solutions were then further optimized, including: Step 1: Problem definition and variable initialization; Define a flight set: Define the set of aircraft positions: For each flight Create variables This indicates the assigned machine position, and sets its field to "All Machine Positions". ; Collect input data: flight arrival time Departure time Aircraft size Machine location dimensions And location attributes including type or neighbor relationships; Step 2, specify constraints; List all hard constraints including time conflicts and size matching, and transform them into logical expressions for CP processing; apply a propagation mechanism to each constraint: before searching, reduce the domain of variables, and if the size of a gate is less than a set threshold, remove the gate from the domain of the relevant flight. Step 3, constraint solution and propagation; Constraint propagation is performed using the CP solver: constraints are iteratively checked, and the propagation domain is reduced to decrease the search space; if propagation results in an empty domain for any variable, the current partial solution is infeasible and is immediately discarded; the search process employs depth-first search (DFS) or finite backtracking to generate candidate assignment schemes, including tentative assignments. And propagate constraints to verify feasibility; Step 4: Output feasible solutions; Output the set of all feasible machine station allocation schemes that satisfy all constraints. Each solution is represented as a variable assignment vector. If there are multiple feasible solutions, the screening process ends; if there are no feasible solutions, the problem is reported as infeasible.
[0033] Using constraint-based programming, a preliminary screening of gate allocation schemes is conducted, including: Step 1: Problem definition and variable initialization; Define a flight set: Define the set of aircraft positions: For each flight Create variables This indicates the assigned machine position, and sets its field to "All Machine Positions". ; Collect input data: flight arrival time Departure time Aircraft size Machine location dimensions And location attributes including type or neighbor relationships; Step 2, specify constraints; List all hard constraints including time conflicts and size matching, and transform them into logical expressions for CP processing; apply a propagation mechanism to each constraint: before searching, reduce the domain of variables, and if the size of a gate is less than a set threshold, remove the gate from the domain of the relevant flight. Step 3, constraint solution and propagation; Constraint propagation is performed using the CP solver: constraints are iteratively checked, and the propagation domain is reduced to decrease the search space; if propagation results in an empty domain for any variable, the current partial solution is infeasible and is immediately discarded; the search process employs depth-first search (DFS) or finite backtracking to generate candidate assignment schemes, including tentative assignments. And propagate constraints to verify feasibility; Step 4: Output feasible solutions; Output the set of all feasible machine station allocation schemes that satisfy all constraints. Each solution is represented as a variable assignment vector. If there are multiple feasible solutions, the screening process ends; if there are no feasible solutions, the problem is reported as infeasible.
[0034] For example, the time complexity of this process depends on the number of constraints and the size of the problem, but the propagation mechanism of CP can filter out invalid solutions early and improve efficiency. For example, in the actual airport problem, when n and m are in the hundreds, CP can complete the initial screening in seconds. In step 2, the time conflict constraints (non-overlapping constraints) include: If two flights have overlapping times, they cannot be assigned the same gate. ; In the formula, For the first One flight, For the first One flight, For flights, conditions Used to determine overlapping time intervals; Machine position size matching constraints include: The aircraft size must be less than or equal to the parking space capacity; ; In the formula, Indicates the camera position Capacity size, aircraft size (For example, 1: small, 2: medium, 3: large) Variables: Gate allocation variables for each flight; Domains: The range of possible values for each variable; Constraints: Logical or mathematical conditions describing the relationships between variables (e.g., flights with overlapping times cannot be assigned to the same gate). The CP solver uses propagation and search algorithms (such as backtracking or branch and bound) to find all feasible solutions or verify that no solution exists. The initial screening phase focuses primarily on quickly identifying feasible solutions rather than optimizing the objective (such as minimizing delays), therefore the emphasis is on constraint definition and propagation.
[0035] The NSGA-II multi-objective evolutionary algorithm is introduced to optimize the preliminary selected schemes by maximizing the docking rate, minimizing passenger walking distance, minimizing aircraft taxiing time, and maximizing gate turnover rate.
[0036] Specifically, this includes: Step 1, problem modeling and coding; Decision variables: Encoded as integers for each flight Assigned to machine station , recorded as ; Population initialization: Randomly generated There are 1 feasible solution (each solution is a gate allocation scheme) to ensure that the basic constraints are met (such as gate type matching and no flight time conflict).
[0037] Step 2, calculate the objective function; Four target values are calculated for each individual: To maximize the bridge utilization rate: ; In the formula, To maximize bridge utilization, For the set of machine positions, For each flight Assigned to machine station The value, For the number of flights; Minimize the passenger walking distance as follows: ; In the formula, To minimize the walking distance for passengers, For camera position Distance to the center of the terminal building; The shortest taxiing time for an aircraft is represented as: ; In the formula, The shortest taxiing distance for an aircraft. From the runway to the parking position Average glide time; The highest machine turnover rate is represented as: ; In the formula, This represents the highest machine turnover rate. For camera position Effective usage time, For turnover cycle; Step 3, constraint handling; The expression for hard constraint gate-flight type matching is: ; In the formula, For flights Available machine location set; The expression for time-free operation is: ; In the formula, To ensure that flight times do not overlap, For aircraft size, For departure time, For the first Aircraft size for each flight For the first Departure time of each flight; Soft constraints, transit time: ; In the formula, For transit connection time, This is the minimum transit time.
[0038] For example, the NSGA-II multi-objective evolutionary algorithm is introduced to form a collaborative mechanism of "large model providing knowledge guidance + NSGA-II efficiently searching for the optimal solution". The combination of the two technologies is suitable for handling scenarios with complex constraints and conflicting objectives. The optimization objectives are to maximize the docking rate, minimize the passenger walking distance, minimize the aircraft taxiing time, and maximize the gate turnover rate, and to deeply optimize the solutions after the initial screening.
[0039] NSGA-II performs non-dominated sorting and crowding calculation on the initial solutions generated by the large model, retaining high-quality individuals. It expands the solution space through crossover and mutation operations, filtering invalid solutions using constraint rules (such as aircraft type adaptation and safety clearance). It iterative optimization gradually approaches the Pareto optimal solution set, periodically feeding intermediate solutions back to the large model for adjustment suggestions. The large model analyzes the NSGA-II search process in real time and dynamically adjusts the search direction. When the algorithm gets stuck in a local optimum (e.g., the bridge utilization rate is generally low in the Pareto solution set), it outputs guidance information (e.g., if the bridge utilization rate of narrow-body aircraft is too high in the current solution set, it suggests prioritizing the allocation of bridges to wide-body aircraft). NSGA-II adjusts its crossover and mutation strategies accordingly (e.g., increasing the mutation probability of the wide-body aircraft-bridge gene).
[0040] Furthermore, by employing reinforcement learning algorithms, the system can continuously adjust the parameters of the reinforcement learning algorithm based on real-time operational feedback, dynamically adapt to changes in airport operational status, and optimize the allocation scheme.
[0041] The system is set by default to ensure consistent data values for multiple objectives in order to achieve a balanced proportion among objectives. On the front-end interface, users can set the target value for each optimization item as needed, such as docking rate, passenger walking distance, aircraft taxiing time, and gate turnover rate.
[0042] The human-computer interaction module provides airport staff with a user-friendly interface, supporting functions such as data querying, scheme adjustment, and result display. Staff can input special requirements or feedback information through this module, and the system will further optimize and adjust the allocation scheme based on this information. During scheme execution, the system monitors the airport's operational status in real time. Once it detects potential gate conflicts such as "the same resource being occupied by multiple requests at the same time" or "resource usage not complying with safety and operational rules," it immediately activates the conflict detection and resolution mechanism for optimization and adjustment, such as optimizing flights from remote gates to near gates, towing flights from near gates to remote gates after passengers have disembarked at the bridge, or adjusting gates to the nearest available location. A lightweight, large-scale model for the vertical domain analyzes conflict situations based on historical conflict resolution experience and real-time data. By determining whether the same resource is occupied by multiple flights at the same time or whether resource usage does not comply with safety and operational rules, it quickly locates conflicting flights and gates. Through deep learning algorithms, multiple conflict resolution strategies are generated from historical conflict resolution experience and real-time data, and the implementation effect of each strategy is simulated and predicted. The effectiveness of the simulation prediction is mainly assessed based on whether it meets the target parameter settings. The system selects the gate allocation strategy for flights based on the weight of each multi-objective parameter set, automatically adjusts the gate allocation plan, and promptly feeds back the adjustment results to relevant departments.
[0043] To further illustrate the effects of the embodiments of the present invention, the following experiment was conducted. A large-scale model simulation experiment was carried out using operational data from a major airport as the background. This airport's gate resources can meet the needs of peak-hour flight operations. The goal is to achieve globally optimal scheduling of gate resources based on a lightweight large-scale model in the civil aviation vertical field, maximizing jet bridge utilization, minimizing total flight taxiing time, and improving on-time performance. The main work includes modeling flight and gate information, using Transformer, Deep Reinforcement Learning (DRL), and Graph Neural Networks (GNN) as the technical foundation. Model training, simulation, and inference share a single high-performance server as the computing node. The CPU is an Intel Xeon processor (64 cores / 128 threads), equipped with two NVIDIA H100 or A100 GPUs (80GB VRAM), 1024GB of RAM, and a high-speed NVMe SSD storage pool (TB level). This airport handles an average of 530 inbound and outbound flights per day. The model training used nearly three years of historical data, totaling approximately 60GB.
[0044] Because of the relatively small amount of data, the simulation speed is much faster than real-time when running on the inference server; the simulation model can achieve 50-100 times the real-time speed. That is, simulating the operation of an airport for 24 hours would only take about 0.12-0.24 hours in the real world. Under normal scheduling conditions, generating the optimal resource allocation plan for a batch of flights in the next 4 hours takes less than 5 minutes.
[0045] The success of the experiment's ultimate goal (metrics) will be verified through improvements in the following key performance indicators (KPIs): Increase the boarding bridge docking rate: from the current 80% to 82%-85%. Reduce the average taxiing-in / taxiing-out time of flights: reduce by 10%-15%.
[0046] Improved on-time departure rate: Based on the baseline stipulated by the Civil Aviation Administration (excluding June-August, the on-time departure rate of airports with a passenger throughput of 10 million passengers (inclusive) or more shall not be less than 85% in the current month; during June-August, the on-time departure rate of airports with a passenger throughput of 10 million passengers (inclusive) or more shall not be less than 80% in the current month), the rate will be increased by 3-5 percentage points.
[0047] As can be seen from the above, the present invention improves allocation efficiency and accuracy: the lightweight large model in the vertical domain can quickly process massive amounts of data, accurately predict flight and gate demand, realize fast and accurate allocation of gates, reduce manual intervention, and improve work efficiency.
[0048] Multi-objective optimization has been achieved: by comprehensively considering various influencing factors, such as flight type, passenger flow, and aircraft size, a balance is achieved between multiple optimization objectives, such as maximizing the docking rate, minimizing passenger walking distance, and minimizing aircraft taxiing time, thereby improving the overall operational efficiency and service quality of the airport.
[0049] The system's flexibility and adaptability have been enhanced: the lightweight, large-scale model in the vertical domain has powerful learning capabilities and can automatically adjust the allocation strategy according to changes in airport operations, adapting to different operational scenarios and emergencies.
[0050] Enhanced human-computer collaboration experience: A user-friendly human-computer interaction interface facilitates operation and management by staff, while also supporting manual intervention and adjustments to achieve efficient decision-making through human-computer collaboration.
[0051] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.
Claims
1. A method for allocating camera positions based on a lightweight, large-scale model in a vertical domain, characterized in that, The method includes the following steps: S1 collects real-time and historical flight data from various data sources at the airport and preprocesses the collected data; the various data sources include data from the flight information system, meteorological system, and equipment and facilities management system. S2 utilizes a lightweight large model in the vertical domain to train preprocessed historical flight data. The lightweight large model in the vertical domain adopts a Transformer architecture that incorporates a convolutional neural network (CNN) module. For the airport gate allocation scenario, the structure of the lightweight large model in the vertical domain is customized, and multi-task output is designed in the output layer of the lightweight large model in the vertical domain according to the gate allocation task. S3, based on the analysis results of the lightweight large model in the vertical domain, combined with the airport's business rules and actual situation, completes the balance optimization of multiple objectives in the complex airport operation environment, and uses a hierarchical hybrid algorithm to generate a preliminary gate allocation scheme. S4 optimizes and adjusts the generated preliminary aircraft station allocation plan, and promptly provides feedback and displays the adjustment results.
2. The camera position allocation method based on a lightweight large model in a vertical domain according to claim 1, characterized in that, In step S2, the preprocessed historical flight data is trained using a lightweight, large model in the vertical domain. The training data includes the historical arrival and departure times of flights, gate usage, passenger flow, delay time, and weather conditions. During the training process, a combination of supervised learning and reinforcement learning is used to continuously optimize the model parameters.
3. The camera position allocation method based on a lightweight large model in a vertical domain according to claim 2, characterized in that, Customized design of lightweight large model structure for vertical domain includes: designing a multimodal data fusion module in the input layer of the lightweight large model for vertical domain to effectively fuse different types of data; In the output layer of the lightweight large model in the vertical domain, multi-task output is designed according to the task of the gate allocation. For flights with insufficient transit time, the shortest taxiing time gate is predicted, the best gate for normal flights is predicted, the estimated taxiing time is predicted, and the conflict risk is predicted. After training, the large model is deployed.
4. The camera position allocation method based on a lightweight large model in a vertical domain according to claim 1, characterized in that, In step S3, a hierarchical hybrid algorithm is used to generate a preliminary station allocation scheme, including: First, constraint-based programming is used to initially screen the gate allocation schemes to ensure that the generated schemes meet the airport's basic business rules, including matching gate type with aircraft type, special requirements of flights, and meeting safety time interval requirements. During the generation process, the priority of flights is taken into account to obtain suitable gate resources. Then, the NSGA-II multi-objective evolutionary algorithm is introduced to optimize the preliminary selected schemes by maximizing the docking rate, minimizing the passenger walking distance, minimizing the aircraft taxiing time, and maximizing the gate turnover rate.
5. The camera position allocation method based on a lightweight large model in a vertical domain according to claim 4, characterized in that, The initially screened solutions were then further optimized, including: Step 1: Problem definition and variable initialization; Define a flight set: Define the set of aircraft positions: For each flight Create variables This indicates the assigned machine position, and sets its field to "All Machine Positions". ; Collect input data: flight arrival time Departure time Aircraft size Machine location dimensions And location attributes including type or neighbor relationships; Step 2, specify constraints; List all hard constraints including time conflicts and size matching, and transform them into logical expressions for CP processing; apply a propagation mechanism to each constraint: before searching, reduce the domain of variables, and if the size of a gate is less than a set threshold, remove the gate from the domain of the relevant flight. Step 3, constraint solution and propagation; Constraint propagation is performed using the CP solver: constraints are iteratively checked, and the propagation domain is reduced to decrease the search space; if propagation results in an empty domain for any variable, the current partial solution is infeasible and is immediately discarded; the search process employs depth-first search (DFS) or finite backtracking to generate candidate assignment schemes, including tentative assignments. And propagate constraints to verify feasibility; Step 4: Output feasible solutions; Output the set of all feasible machine station allocation schemes that satisfy all constraints. Each solution is represented as a variable assignment vector. If there are multiple feasible solutions, the screening process ends; if there are no feasible solutions, the problem is reported as infeasible.
6. The camera position allocation method based on a lightweight large model in a vertical domain according to claim 5, characterized in that, In step 2, the time conflict constraints include: If two flights have overlapping times, they cannot be assigned the same gate position. The expression is: ； In the formula, For the first One flight, For the first One flight, For flights, conditions Used to determine overlapping time intervals; In step 2, the machine position size matching constraints include: The aircraft size must be less than or equal to the parking space capacity, expressed as: ； In the formula, Indicates the camera position Capacity size.
7. The camera position allocation method based on a lightweight large model in a vertical domain according to claim 4, characterized in that, The NSGA-II multi-objective evolutionary algorithm is introduced to optimize the preliminary selected schemes by maximizing the docking rate, minimizing passenger walking distance, minimizing aircraft taxiing time, and maximizing gate turnover rate. The specific steps include: Step 1, Problem Modeling and Coding; Decision variables: Encoded as integers for each flight Assigned to machine station , recorded as Population initialization: randomly generated A feasible solution is found, ensuring that the basic constraints are satisfied; Step 2, calculate the objective function; Four target values are calculated for each individual, and the bridge berth rate is maximized as follows: ； In the formula, To maximize bridge utilization, For the set of machine positions, For each flight Assigned to machine station The value, For the number of flights; Minimize the passenger walking distance as follows: ； In the formula, To minimize the walking distance for passengers, For camera position Distance to the center of the terminal building; The shortest taxiing time for an aircraft is represented as: ； In the formula, The shortest taxiing distance for an aircraft. From the runway to the parking position Average glide time; The highest machine turnover rate is represented as: ； In the formula, This represents the highest machine turnover rate. For camera position Effective usage time, For turnover cycle; Step 3, constraint handling; The expression for hard constraint gate-flight type matching is: ； In the formula, For flights Available machine location set; The expression for time-free operation is: ； In the formula, To ensure that flight times do not overlap, For aircraft size, For departure time, For the first Aircraft size for each flight For the first Departure time of each flight; Soft constraints, transit time: ； In the formula, For transit connection time, This is the minimum transit time.
8. The camera position allocation method based on a lightweight large model in a vertical domain according to claim 4, characterized in that, The initial screening of the proposed solutions was further optimized, including by continuously adjusting the parameters of the reinforcement learning algorithm based on real-time operational feedback to dynamically adapt to changes in the airport's operational status and optimize the allocation scheme.
9. The camera position allocation method based on a lightweight large model in a vertical domain according to claim 1, characterized in that, In step S4, the generated preliminary aircraft stand allocation plan is optimized and adjusted, including: The system further optimizes and adjusts the initial gate allocation plan based on the input flight-specific requirements or feedback information in the user interface. This includes: real-time monitoring of airport operations, detecting gate conflicts, and initiating conflict detection and resolution mechanisms for optimization and adjustment; using a lightweight, large-scale model in the vertical domain based on historical conflict resolution experience and real-time data to model and analyze conflict situations, locating conflicting flights and gates; generating multiple conflict resolution strategies through deep learning algorithms and simulating and predicting the implementation effect of each strategy; selecting the optimal strategy, automatically adjusting the gate allocation plan, and providing timely feedback on the adjustment results.
10. A machine station allocation system based on a lightweight, large-scale model in a vertical domain, characterized in that, The system implements the camera position allocation method based on a lightweight, large model in a vertical domain as described in any one of claims 1-9, and the system includes: The data acquisition and preprocessing module is used to collect real-time and historical flight data from various data sources at the airport, including flight information systems, meteorological systems, and equipment and facilities management systems, and to preprocess the collected data. The core processing module of the vertical domain lightweight large model is used to train the preprocessed historical flight data using the vertical domain lightweight large model. The vertical domain lightweight large model adopts the Transformer architecture with the integration of the convolutional neural network (CNN) module. For the airport gate allocation scenario, the structure of the vertical domain lightweight large model is customized, and multi-task output is designed in the output layer of the vertical domain lightweight large model according to the gate allocation task. The decision generation and optimization module is used to perform multi-objective balance optimization in the complex airport operation environment based on the analysis results of the lightweight large model in the vertical domain, combined with the airport's business rules and actual conditions, and to generate a preliminary gate allocation scheme using a hierarchical hybrid algorithm. The human-computer interaction module is used to optimize and adjust the generated preliminary camera position allocation plan, and to provide timely feedback and display of the adjustment results.