Aerial signing system and device, training method, electronic equipment and storage medium

Abstract

The invention discloses an aviation signing system, an aviation signing generation device, a training method of an aviation generation type large model, electronic equipment and a non-instantaneous computer readable storage medium. The aviation dispatching system comprises an acquisition module, an auxiliary decision-making module and a user interaction module, the acquisition module is configured to acquire navigation notification information and airway weather information; the auxiliary decision-making module is configured to generate auxiliary decision-making information through an aviation generation type large model and a navigation announcement knowledge graph based on the navigation announcement information and the airway weather information; and the user interaction module is configured to display the auxiliary decision-making information. The aviation signing system can improve the quality and reliability of aviation signing work.

Description

Technical Field

[0001] Embodiments of the present disclosure relate to an aviation dispatch system, an aviation dispatch generation device, a training method for an aviation generative large model, an electronic device, and a non-transitory computer-readable storage medium. Background Art

[0002] With the rapid development of the air transport industry, NOTAM data and en-route weather data are complex and diverse. Traditional manual processing methods are inefficient and prone to errors. Existing technologies are also insufficient in data fusion and intelligent analysis, making it difficult to meet the requirements for efficient and safe air dispatch decision-making. For example, NOTAM parsing relies on manual annotation, which is time-consuming and has limited accuracy. Route planning and dispatch decisions lack comprehensive intelligent analysis of multi-source data, making it difficult to respond promptly to complex situations such as severe weather and airspace restrictions. Summary of the Invention

[0003] At least one embodiment of the present disclosure provides an aviation dispatch system, comprising an acquisition module, a decision support module, and a user interaction module. The acquisition module is configured to acquire NOTAM information and en-route weather information; the decision support module is configured to generate decision support information based on the NOTAM information and en-route weather information using an aviation generative large model and a NOTAM knowledge graph; and the user interaction module is configured to display the decision support information.

[0004] For example, in the aviation dispatch system provided in one embodiment of the present disclosure, the auxiliary decision information includes first auxiliary decision information and second auxiliary decision information; the auxiliary decision module is further configured to generate the first auxiliary decision information and the second auxiliary decision information based on the navigation notice information and the route weather information through the aviation generative big model and the navigation notice knowledge graph, wherein the first auxiliary decision information includes route plan information, and the second auxiliary decision information includes release decision recommendation information.

[0005] For example, in the aviation dispatch system provided in one embodiment of the present disclosure, the aviation generative big model includes a risk assessment model; the auxiliary decision module is further configured to generate the first auxiliary decision information based on the navigation notice information and the enroute weather information through a search algorithm, a risk assessment model and / or the navigation notice knowledge graph.

[0006] For example, in the aviation dispatch system provided in one embodiment of the present disclosure, the auxiliary decision module is further configured to generate the second auxiliary decision information based on the navigation notice information and the enroute weather information through aviation business rules, the aviation generative big model and / or the navigation notice knowledge graph.

[0007] For example, in the aviation dispatch system provided in one embodiment of the present disclosure, the user interaction module is further configured to generate and display a route plan in response to the user's operation on the operation interface through the code interactive route design tool and the aviation generative large model.

[0008] For example, in the aviation dispatch system provided by an embodiment of the present disclosure, the code-interactive route design tool is developed through the aviation generative big model, geographic information system and / or flight performance database.

[0009] For example, in the aviation dispatch system provided in an embodiment of the present disclosure, the user interaction module is further configured to provide the data adjusted by the user operation to the aviation generative big model for model training and optimization.

[0010] For example, in the aviation dispatch system provided in one embodiment of the present disclosure, the NOTAM knowledge graph is based on multi-source NOTAM data and multi-source en-route weather data, and is optimized through the aviation generative big model.

[0011] At least one embodiment of the present disclosure provides an aviation dispatch generation device, comprising: a data processing module and a knowledge graph module. The data processing module is configured to generate object entity information based on multi-source NOTAM data and multi-source en-route weather data using an aviation generative big model; and the knowledge graph module is configured to generate a NOTAM knowledge graph based on the object entity information.

[0012] For example, in the air dispatch generation device provided in an embodiment of the present disclosure, the object entity information includes at least one of geographic location and facility type, airspace structure type, time condition type or event type entity information.

[0013] For example, in the aviation dispatch generation device provided in one embodiment of the present disclosure, the knowledge graph module is further configured to generate the navigation notice knowledge graph based on the object entity information and the object entity relationship information, wherein the object entity relationship information includes at least one of spatial relationship information, temporal relationship information or impact relationship information.

[0014] At least one embodiment of the present disclosure provides a training method for an aviation generative big model, the training method comprising: obtaining a training data set determined based on multi-source NOTAM data, enroute weather data, or aircraft parameter data; and performing multi-task training on the aviation generative big model based on the training data set, wherein the aviation generative big model adopts a deep learning framework.

[0015] For example, in the training method of the aviation generative large model provided in an embodiment of the present disclosure, the training method further includes: determining the training data set, wherein the multi-source NOTAM data is parsed based on an aviation notice corpus including the object entity information annotations.

[0016] For example, in the training method of the aviation generative large model provided in an embodiment of the present disclosure, the aviation generative large model includes an encoding layer, which is used to convert the object entity information into a vector representation corresponding to the aviation generative large model.

[0017] For example, in the training method of the aviation generative large model provided in an embodiment of the present disclosure, the encoding layer includes a sub-encoding layer for aviation terminology and / or geospatial information.

[0018] For example, in the training method of the aviation generative large model provided in one embodiment of the present disclosure, the performance indicators of the aviation generative large model are evaluated based on the leave-one-out method or the cross-validation method, wherein the performance indicators include: aviation notice parsing accuracy, decision recommendation rationality and / or route code generation compliance rate.

[0019] For example, in the training method of the aviation generative large model provided in one embodiment of the present disclosure, the multiple tasks include a navigation notice parsing task, a decision support task, a risk analysis task and / or a route design task.

[0020] For example, in the training method of the aviation generative large model provided in one embodiment of the present disclosure, the aviation generative large model is a large model based on the Transformer architecture constructed using a deep learning framework.

[0021] At least one embodiment of the present disclosure provides an electronic device, including the aviation dispatch system provided by any embodiment of the present disclosure and / or the aviation dispatch generation device provided by any embodiment of the present disclosure.

[0022] At least one embodiment of the present disclosure provides an electronic device, comprising: a processor and a memory, wherein the memory stores at least one computer program, and when the at least one computer program is executed by the processor, the training method of the aviation generative large model provided by any embodiment of the present disclosure is implemented.

[0023] At least one embodiment of the present disclosure provides a non-transitory computer-readable storage medium for non-temporarily storing computer-executable instructions. When the computer-executable instructions are executed by a computer, the training method for the aviation generative large model provided by any embodiment of the present disclosure is implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings of the embodiments will be briefly introduced below. Obviously, the drawings in the following description only relate to some embodiments of the present disclosure, rather than limiting the present disclosure.

[0025] Figure 1 A schematic block diagram of an aviation dispatch system provided by at least one embodiment of the present disclosure is shown;

[0026] Figure 2 A schematic block diagram of an aviation dispatch generating device provided by at least one embodiment of the present disclosure is shown;

[0027] Figure 3 A flowchart of a method for training an aviation generative large model provided by at least one embodiment of the present disclosure is shown;

[0028] Figure 4 A schematic diagram illustrating an application of an aviation dispatch system provided by at least one embodiment of the present disclosure is shown;

[0029] Figure 5 A schematic block diagram of a training device for an aviation generative large model provided by at least one embodiment of the present disclosure is shown;

[0030] Figure 6 A schematic block diagram of an electronic device provided by at least one embodiment of the present disclosure is shown;

[0031] Figure 7 A schematic block diagram showing another electronic device provided by at least one embodiment of the present disclosure; and

[0032] Figure 8 A schematic diagram of a computer-readable storage medium provided by at least one embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0033] To make the purpose, technical solutions, and advantages of the embodiments of the present disclosure more clear, the technical solutions of the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the described embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0034] Unless otherwise defined, the technical or scientific terms used in this disclosure should have the usual meanings understood by people with ordinary skills in the field to which this disclosure belongs. The words "first", "second" and similar words used in this disclosure do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one", "an" or "the" do not indicate a quantity limitation, but rather indicate the existence of at least one. Words such as "include" or "comprise" mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0035] In the field of air transportation, flight dispatch and release are key to ensuring flight safety and punctuality. With the rapid development of the aviation industry, the number of flights continues to increase, and the operating environment becomes increasingly complex, placing higher demands on flight dispatch work.

[0036] Currently, the processing of NOTAMs mainly relies on manual methods or simple information systems. NOTAMs come from a wide range of sources, including various notices issued by aviation management departments and airport authorities in various countries. Their formats and contents are diverse, covering information such as airport facility changes, runway maintenance, temporary airspace control, and navigation equipment failures. Traditional processing methods make it difficult to quickly and accurately extract key information from a large number of text-based notices and effectively integrate them. For example, dispatchers need to spend a lot of time reading and analyzing NOTAM documents in different formats, manually screening information related to specific flights. This is not only inefficient, but also prone to omissions or misunderstandings due to human negligence, which in turn affects the formulation and adjustment of flight plans.

[0037] In terms of monitoring and analyzing en-route meteorological conditions, although meteorological data can be obtained through various means such as meteorological observation stations and meteorological satellites, there are still many deficiencies in applying meteorological information to aviation dispatch decisions. For example, the processing and analysis of meteorological data may lack deep integration with route planning. Existing systems provide discrete meteorological observation point data or large-scale weather forecasts, which make it difficult to accurately assess the impact of meteorological conditions on specific routes. For example, when faced with complex meteorological systems such as fronts and cyclones, it is impossible to accurately judge the real-time impact of their movement paths and intensity changes on flight routes, making it difficult for dispatchers to formulate reasonable response strategies, increasing the risk of flights encountering dangerous weather conditions.

[0038] For example, traditional route planning methods are typically designed based on fixed route networks and standard flight procedures, without adequate consideration of real-time navigation notices and dynamic weather conditions. When temporary flight restrictions or sudden severe weather conditions occur, existing route planning methods struggle to quickly adjust and provide suitable alternative routes. For example, when an airport is temporarily closed or airspace is controlled, traditional route planning is unable to promptly re-plan safe and efficient routes based on these changes, potentially leading to flight delays, excessive detours, wasted fuel, or increased operating costs.

[0039] For example, when making flight dispatch decisions, the dispatcher must rely on their professional experience and a limited rulebook. Dispatchers need to comprehensively consider multiple factors, including aircraft performance, crew qualifications, fuel reserves, destination airport support capabilities, NOTAMs, and weather conditions. However, due to the lack of effective intelligent auxiliary tools, this manual experience-based decision-making process struggles to conduct comprehensive and accurate quantitative analysis of complex, multi-factor situations. For example, when assessing the dispatch risk of a flight under specific weather conditions, it is difficult to accurately determine the additional fuel reserves, appropriate alternate airports, and risk points during flight. This can lead to overly conservative or risky dispatch decisions, impacting the safety and economy of aviation operations.

[0040] The inventors of the present disclosure have noticed that aviation dispatch and release technology has many limitations in aspects such as NOTAM processing, enroute weather analysis, route planning and release decision-making, and is unable to meet the needs of efficient, safe and economical operation of the modern aviation industry. There is an urgent need for a system to improve the overall level of aviation dispatch work.

[0041] For example, regarding the efficient processing and precise integration of navigation notices, navigation notices come from a wide range of sources, have complex formats, and contain massive amounts of content. How to use appropriate technical means to automatically, quickly, and accurately extract relevant information, and build a structured, highly correlated knowledge graph to achieve efficient integration and intelligent query of notice information, thereby assisting dispatchers to grasp relevant information in a timely and comprehensive manner, and minimize dispatch errors caused by missing information or untimely processing, is an urgent problem to be solved.

[0042] For example, the challenge of accurately assessing and integrating en-route weather conditions is complex and ever-changing. Existing meteorological data processing methods struggle to transform dispersed meteorological observation and forecast data into precise risk assessment information for specific routes and effectively integrate it with other dispatch decision-making factors. Therefore, a meteorological data model and assessment mechanism is needed that can analyze the impact of meteorological factors on different routes in real time, provide a reliable meteorological basis for route planning and dispatch decisions, and mitigate the negative impact of meteorological factors on flight safety and operational efficiency.

[0043] For example, when it comes to dynamic multi-constraint route planning, traditional route planning methods lack flexibility and dynamic optimization capabilities under the dual constraints of NOTAM restrictions and real-time weather changes, and are unable to quickly generate optimal route solutions that meet multiple constraints. Therefore, a route planning algorithm is needed that can comprehensively consider information such as airspace restrictions and airport conditions in NOTAMs, as well as factors such as severe weather areas and wind direction and speed in meteorological conditions, and adjust routes in real time during flight operations to minimize flight delays, fuel consumption, and safety risks.

[0044] For example, regarding the lack of intelligent, quantitative decision-making support for dispatch, current dispatch decisions rely primarily on dispatcher experience, making it difficult to conduct comprehensive, accurate, and quantitative analysis of numerous complex and interrelated factors. There is an urgent need for a dispatch decision-making model based on big data and intelligent algorithms. This model can integrate multiple factors, including aircraft performance data, crew qualification information, fuel reserve status, destination airport support capabilities, NOTAMs, and weather conditions, to provide dispatchers with scientific, quantitative dispatch decision recommendations, improve the accuracy and reliability of dispatch decisions, and ensure the safety and economy of aviation operations.

[0045] Embodiments of the present disclosure provide an aviation dispatch system, an aviation dispatch generation device, a training method for an aviation generative large model, an electronic device, and a non-transitory computer-readable storage medium.

[0046] The aviation dispatch system includes an acquisition module, a decision support module, and a user interaction module. The acquisition module is configured to acquire NOTAM information and en-route weather information; the decision support module is configured to generate decision support information based on the NOTAM information and en-route weather information using an aviation generative large model and a NOTAM knowledge graph; and the user interaction module is configured to display the decision support information.

[0047] The aviation dispatch system provided by at least one embodiment of the present disclosure can improve the quality and reliability of aviation dispatch work from multiple aspects such as NOTAM processing, decision support, risk analysis, and route design, and can effectively reduce the risk of safety accidents caused by factors such as human error, meteorological disasters, or navigation restrictions. At least one embodiment of the present disclosure can significantly reduce aviation operating costs and increase flight operating revenue by reducing flight delays and cancellations, optimizing fuel consumption, and improving dispatch efficiency. At least one embodiment of the present disclosure can also accurately calculate fuel and plan routes to save a large amount of fuel costs, and timely decision adjustments can reduce economic losses caused by flight delays or cancellations.

[0048] Figure 1 A schematic block diagram of an aviation dispatch system provided by at least one embodiment of the present disclosure is shown.

[0049] like Figure 1As shown, the airline dispatch system 100 includes an acquisition module 110 , an auxiliary decision module 120 and a user interaction module 130 .

[0050] The acquisition module 110 is configured to acquire NOTAM information and en-route weather information.

[0051] The decision support module 120 is configured to generate decision support information based on NOTAM information and en-route weather information through an aviation generative big model and a NOTAM knowledge graph.

[0052] The user interaction module 130 is configured to display decision-supporting information.

[0053] For example, the acquisition module 110 may acquire navigation notices and en-route weather information in real time from multiple source channels through a data acquisition interface.

[0054] For example, the acquisition module 110 may be configured to acquire NOTAM information and en-route weather information in real time from a multi-source data interface, and perform pre-processing and standardization.

[0055] For example, NOTAM information may include notices related to aviation operations issued by aviation management departments, airport authorities, etc. NOTAM information may cover changes in airport facilities (such as runway maintenance), temporary airspace control, navigation equipment failure or other restrictions, etc.

[0056] For example, en-route weather information may include real-time weather data along the route (such as temperature, humidity, wind speed, wind direction, air pressure, etc.) and weather forecast data (such as thunderstorm and heavy fog warnings, etc.).

[0057] For example, data related to route weather information can come from meteorological satellites, ground observation stations, weather radars, etc. These data can be cleaned and integrated into structured data to be used to evaluate the impact of weather on flight safety.

[0058] For example, raw data can be obtained from relevant channels through data interface tools or data extraction tools.

[0059] For example, natural language processing (NLP) may be used to parse the NOTAM text and extract entity information therein (eg, airport, time, restriction type, etc., which is not limited in the embodiments of the present disclosure).

[0060] For example, meteorological data can be denoised and normalized, and organized by temporal and spatial dimensions for subsequent analysis.

[0061] For example, the decision support module 120 may be configured to generate route planning, risk warnings, and / or clearance decision recommendations based on NOTAM and meteorological data, in combination with the aviation generative big model and the NOTAM knowledge graph.

[0062] For example, aviation generative big models can be used to parse NOTAM text and correlate it with historical data, predict the impact of weather changes on air routes, or generate route code snippets to avoid restricted areas.

[0063] For example, aviation generative big models can be used in conjunction with meteorological data to predict the impact of severe weather on air routes (such as visibility, turbulence probability, etc.).

[0064] For example, aviation generative big models can be used to call flight performance databases (such as aircraft fuel efficiency and maximum range) to generate multiple sets of candidate route plans.

[0065] For example, aviation generative big models can calculate the quantitative risks of route plans (such as delay probability, diversion requirements, etc.), and can combine risk models trained with historical data to evaluate fuel consumption and safety redundancy (such as additional fuel reserves).

[0066] For example, an aviation generative big model may include an artificial intelligence model based on deep learning, which is specially trained for aviation field data and has the capabilities of semantic understanding, logical reasoning and code generation.

[0067] For example, an aviation generative big model may include a big model based on the Transformer architecture built using a deep learning framework.

[0068] For example, the aviation generative big model can be based on an aviation generative big model architecture such as Qwen or LLaMA. Specifically, the aviation generative big model can be based on the aviation generative big model architecture of Qwen 2.5 or LLaMA 3. For example, Qwen 2.5 supports multiple languages ​​and long contexts, including function calling, and is suitable for multilingual parsing of aviation terminology (such as Chinese and English NOTAMs, Notice to Airmen).

[0069] For example, Function Calling can adapt to terminology-intensive scenarios and is suitable for parsing flight plans, NOTAMs, and AIXMs (Aeronautical Information Exchange Models), generating aviation standard codes. For example, Function Calling allows large models to call predefined functions, mapping natural language input (such as "Design a route from PEK to ZSPD, considering NOTAMs and fuel efficiency") to structured tasks.

[0070] For example, LLaMA3 can include a Transformer architecture that can adapt to the fusion of aviation terminology and geospatial data, suitable for real-time route optimization and cross-modal data processing.

[0071] For example, PyTorch can dynamically compute graphs and perform distributed training to adapt to aviation big data.

[0072] For example, a NOTAM knowledge graph can include a structured database for storing entity information and entity relationship information in NOTAMs. For example, semantic association and reasoning can be used to support fast query and risk warning.

[0073] For example, the “nodes” in the NOTAM knowledge graph may include airport A, airspace B or restricted time; the “edges” may include airport A → affected routes → airspace B that needs to be detoured.

[0074] For example, for dynamic route planning, GIS maps and flight performance data can be combined to generate corresponding route codes (such as avoiding thunderstorm areas, optimizing fuel consumption, etc.).

[0075] For example, for risk quantification analysis, historical data can be used to train risk assessment models, calculate delay probabilities, diversion requirements, etc., and output risk levels and response strategies.

[0076] For example, when generating decision recommendations, factors such as aircraft performance, crew qualifications, and fuel reserves can be comprehensively considered to recommend a dispatch plan (such as alternate airport selection, additional fuel quantity, etc.).

[0077] For example, the NOTAM knowledge graph can process multi-source NOTAM data to extract object entity information such as geographic location and facility classes, airspace structure classes, time condition classes, event classes, as well as object entity relationship information such as spatial relationships, temporal relationships, and influence relationships. It constructs a structured knowledge network to represent various entities and their relationships in NOTAM.

[0078] For example, in an aviation dispatch system, the NOTAM knowledge graph can be combined with the aviation generative big model to provide support for the decision support module 120. The NOTAM knowledge graph can help analyze NOTAM information and enroute weather information to generate auxiliary decision-making information such as route plans and dispatch decision recommendations. This facilitates intelligent route planning, risk analysis, and dispatch decision-making, improving the intelligence and accuracy of aviation dispatch and ensuring flight safety and efficiency.

[0079] For example, the user interaction module 130 may be configured to display auxiliary decision information in a visual interface, support users (eg, dispatchers, etc.) to review and modify decision suggestions, and collect feedback to optimize the model.

[0080] For example, auxiliary decision-making information may include flight plan adjustment suggestions (such as delay time, route changes, etc.), risk warnings (such as thunderstorm area risk level, etc.), fuel calculations (such as preferred fuel reserves based on weather and route, etc.), etc.

[0081] For example, the visualization interface may integrate a map view (e.g., displaying routes and weather coverage, etc.), a table view (e.g., flight plan details, etc.), or a risk dashboard, etc., and the embodiments of the present disclosure are not limited to this.

[0082] For example, modifications made by users (e.g., dispatchers, etc.) to system recommendations (e.g., alternate airports selected after user adjustments) can be recorded for iterative training of large aviation generative models.

[0083] In at least one embodiment of the present disclosure, the use of an aviation-generative big model to parse NOTAMs allows for the rapid and accurate extraction of key information from complex and diverse NOTAM texts, such as airport conditions, airspace restrictions, and navigation facility changes, and automatically constructs a structured knowledge graph. Compared to traditional NOTAM processing methods that rely on manual labor or simple information systems, this significantly improves the speed and accuracy of information extraction, reduces the risk of dispatch errors due to information omissions or misunderstandings, and enhances dispatch efficiency. Furthermore, based on the semantic understanding capabilities of the knowledge graph and big model, the aviation dispatch system can automatically identify potential connections between NOTAMs and their combined impact on flight operations. For example, when multiple related NOTAMs are issued, the chain reaction of a series of navigation restrictions within a certain area on a specific group of flights can be quickly inferred, and timely warnings can be issued to dispatchers, allowing them to prepare countermeasures in advance, thereby enhancing the safety and planning of aviation operations.

[0084] In some embodiments of the present disclosure, the auxiliary decision information may include first auxiliary decision information and second auxiliary decision information. The auxiliary decision module 120 may be further configured to generate the first auxiliary decision information and the second auxiliary decision information based on the NOTAM information and the en-route weather information using an aviation generative large model and a NOTAM knowledge graph.

[0085] For example, the first auxiliary decision information may include route plan information, and the second auxiliary decision information may include release decision recommendation information.

[0086] For example, the route plan information may include flight path planning, and may include specific waypoints, flight altitudes, speed limits, and avoidance strategies, etc., to provide an executable route selection plan for the flight, which is not limited in the embodiments of the present disclosure.

[0087] For example, the release decision recommendation information may include decision recommendation information on whether the flight can be released, when to release it, what measures need to be taken (such as adjusting the take-off time, fuel reserves, and alternate airport selection), etc. The embodiments of the present disclosure do not limit this.

[0088] For example, based on route plan information (such as information from the first auxiliary decision-making information and multi-source data integration (for example, aircraft performance (such as maximum take-off weight, fuel capacity), crew qualifications (such as whether they are qualified to fly in complex weather conditions) or destination airport support capabilities (such as runway length, alternate landing facilities)), etc.), the corresponding release decision recommendation information can be generated through the aviation generative big model and the navigation notice knowledge graph.

[0089] For example, the decision support module 120 can be configured to receive the navigation notice information and route weather information (such as wind speed, thunderstorm distribution, etc.) provided by the acquisition module 110; perform deep semantic understanding and feature extraction on the received information through the aviation generative large model, such as identifying object entity information such as "runway 05R is closed" from the navigation notice text, and extracting key factors affecting the flight (such as strong wind area) from the route weather information; various object entity information (such as geographic location, event entity) and their relationships (such as spatial relationship, influencing Relationships), forming a structured knowledge network, for example, associating "runway closure events" with "affiliated airports" and "affected airspaces"; using the aviation generative big model combined with information such as airspace restrictions and meteorological risk areas in the NOTAM knowledge graph to generate the first auxiliary decision-making information; based on the rules and historical data such as airport status (whether the runway is available) and meteorological conditions (whether visibility and wind speed meet the standards) in the NOTAM knowledge graph, the aviation generative big model can be used to analyze and determine whether the flight can be released, when to release it, or whether the alternate airport needs to be adjusted, etc., to generate the second auxiliary decision-making information.

[0090] For example, the first auxiliary decision information and the second auxiliary decision information can work together to achieve dynamic linkage.

[0091] For example, if the route plan needs to be adjusted due to sudden weather changes, the dispatch decision recommendation can be updated synchronously (such as recalculating fuel requirements).

[0092] For example, users can adjust waypoints based on a visual display of the route plan (GIS (Geographic Information System) map), and the system will provide real-time feedback on changes to the release recommendation.

[0093] In at least one embodiment of the present disclosure, the aviation dispatch system outputs route plans and flight release recommendations, providing both specific operational paths and risk assessments, thereby reducing dispatcher workload. Leveraging the real-time data processing capabilities of the aviation generative large-scale model, the primary and secondary decision-making support information can be dynamically adjusted based on NOTAM and weather changes.

[0094] In at least one embodiment of the present disclosure, the aviation generative model can comprehensively integrate multiple factors, such as navigation notices, en route weather conditions, aircraft performance, and crew qualifications, to provide dispatchers with scientific, quantitative decision-making assistance. Compared with traditional empirical decision-making, this avoids decision-making errors caused by personal experience limitations or incomplete information, significantly improving the accuracy and reliability of flight release decisions. For example, under complex weather conditions and navigation restrictions, it can accurately calculate the required fuel reserves, suitable alternate airports, and optimal takeoff time for a flight, effectively reducing operating costs and ensuring flight safety.

[0095] In at least one embodiment of the present disclosure, the large-scale model can rapidly reassess and adjust its decision recommendations as NOTAMs and weather conditions change in real time. During flight operations, if a flight encounters sudden severe weather or temporary flight restrictions, the dispatcher can be provided with timely response strategies, such as adjusting the route, changing the flight altitude, or changing the speed. This ensures that the flight can flexibly respond to various emergencies, reduces the probability of flight delays and cancellations, and improves the adaptability and flexibility of aviation operations.

[0096] In some embodiments of the present disclosure, the aviation generative model includes a risk assessment model. The decision support module 120 can be further configured to generate first decision support information based on NOTAM information and en-route weather information through a search algorithm, a risk assessment model, and / or a NOTAM knowledge graph.

[0097] For example, risk assessment models can be trained based on historical flight data, accident statistics, meteorological disaster records, etc., to quantitatively assess the safety risks of route plans (such as delay probability, diversion requirements, fuel shortage risk, etc.).

[0098] For example, search algorithms can include heuristic algorithms (such as A* algorithm, genetic algorithm) or deep learning-driven path planning algorithms.

[0099] For example, data preprocessing can be performed on NOTAM information and en-route weather information input into risk assessment models. For example, NOTAM text can be parsed, key entities (such as restricted area coordinates and effective time) extracted, and mapped to knowledge graph nodes. For example, real-time meteorological data (such as thunderstorm area polygon coordinates and wind speed profiles) can be integrated with forecast information to construct a spatiotemporal multidimensional data cube.

[0100] For example, for candidate route generation, constraint modeling can be performed, and NOTAM restrictions (such as no-fly zones) and weather constraints (such as maximum permissible turbulence levels) can be converted into boundary conditions of the search space. For example, the coordinate range of the thunderstorm area is set as the search restricted area, and the airspace control area is set as the obstacle that the path needs to bypass. For example, path search and optimization can be performed, with fuel efficiency and flight time as optimization goals, to generate initial candidate routes. For example, assuming that the initial path from starting point A to end point B is a straight line, all NOTAM restriction areas and weather restricted areas can be traversed, the cost of the bypass path (such as increased range and fuel consumption) can be calculated, and the corresponding candidate route plan can be generated.

[0101] For example, risk indicator calculations can assess the corresponding risk dimensions for each candidate route. These could include meteorological risk (e.g., thunderstorm penetration probability, turbulence intensity forecast), notification risk (e.g., whether the detour path is adjacent to other restricted areas (which can be queried through knowledge graph associations), fuel risk (e.g., the impact of the increased distance of the detour on fuel margin), or model inputs (e.g., route coordinates, meteorological data, and aircraft performance parameters (e.g., fuel efficiency).

[0102] For example, for route optimization driven by the NOTAM knowledge graph, semantic association analysis can be performed, such as querying the knowledge graph for entities related to candidate routes (such as alternate airports and navigation equipment status). The NOTAM knowledge graph can also be used to detect whether candidate routes contain implicit restrictions (such as implicit no-fly zones caused by the overlap of multiple NOTAMs).

[0103] For example, for the generated route plans, the candidate routes can be ranked by multi-objective decision-making, such as comprehensive risk score, fuel consumption, flight time and other indicators.

[0104] For example, after the search algorithm generates candidate routes, the risk assessment model can provide real-time risk feedback and drive the algorithm to iteratively adjust the route (such as avoiding high-risk areas).

[0105] In at least one embodiment of the present disclosure, the limitations of traditional single-source risk assessments can be overcome by integrating multi-dimensional information such as meteorological data, NOTAM restrictions, and aircraft performance. A NOTAM knowledge graph supports semantic-level correlation analysis of NOTAMs, enabling the identification of implicit risks (such as the cumulative effects of multiple NOTAMs). Furthermore, the search algorithm can dynamically adjust the weights of heuristic functions based on risk assessment results, balancing safety and efficiency objectives.

[0106] In at least one embodiment of the present disclosure, by learning and analyzing large amounts of historical data, the aviation generative model can establish a precise risk assessment model, transforming complex factors such as weather conditions and NOTAMs into quantifiable risk indicators. For example, it can accurately calculate the probability of a flight being delayed, diverted, or experiencing an accident under specific routes and weather conditions, providing dispatchers with intuitive and clear risk assessment results, enabling them to implement targeted risk mitigation measures and reduce aviation operational risks.

[0107] In at least one embodiment of the present disclosure, based on risk assessment results, the air dispatch system can automatically generate corresponding risk warning information and response strategy recommendations. When the risk level exceeds a preset threshold, the dispatcher is promptly alerted and provided with specific response measures, such as increasing fuel reserves, selecting a safer route, or selecting an alternate airport, effectively improving the safety of aviation operations and emergency response capabilities.

[0108] In some embodiments of the present disclosure, the auxiliary decision module 120 can be further configured to generate second auxiliary decision information based on the NOTAM information and the enroute weather information through aviation business rules, aviation generative big models and / or NOTAM knowledge graphs.

[0109] For example, aviation business rules may include regulations, standard operating procedures (SOPs), and company policies that must be followed during aviation operations. Examples include fuel margin rules, crew qualification rules, and airport weather minimums, though the embodiments of this disclosure do not limit these.

[0110] For example, to generate secondary decision-making information, input data can be integrated. For example, for NOTAM information, key restrictions (such as airport closures and airspace restrictions) can be parsed and extracted, and mapped to knowledge graph nodes. For en-route weather information, real-time weather data (such as en-route wind speed and visibility) and forecast information (such as the probability of thunderstorms in the next three hours) can be integrated to construct a spatiotemporal multidimensional feature matrix.

[0111] For example, infeasible options can be initially filtered out based on predefined aviation business rules. For example, if the visibility at the destination airport is below the minimum standard (e.g., <800 meters), the flight is prohibited from being released.

[0112] For example, quantitative analysis and prediction using large-scale aviation generative models can enable multi-dimensional risk assessments. Inputs to these models can include route plans, weather data, or aircraft performance. These models can output metrics such as delay probability (based on historical delay statistics under similar weather conditions), diversion requirement probability (combining the risk of sudden weather changes along the route and the availability of alternate airports), and fuel consumption forecasts (calculating precise fuel requirements based on range, wind speed, and altitude). Risk assessment results can also generate flight dispatch recommendations (e.g., "Permit dispatch, but increase diversion fuel to 2,000 kg").

[0113] For example, the NOTAM knowledge graph can be used to query the real-time status of alternate airports (such as capacity, equipment status, and weather conditions). For example, multiple objectives can be optimized and ranked, combining aviation business rule compliance, risk assessment results, and knowledge graph reasoning to rank candidate options and generate decision recommendations with detailed parameters and evidence.

[0114] In some embodiments of the present disclosure, when a new NOTAM is issued, the air dispatch system can input the NOTAM information into an aviation generative model. The aviation generative model, combined with the NOTAM knowledge graph, quickly analyzes the scope and extent of the NOTAM's impact on existing flight plans. For example, if a runway maintenance notice is issued at a particular airport, the aviation generative model can predict which flights require departure time adjustments and rerouting based on the airport's operational data, flight traffic distribution, and other information, and provide corresponding adjustment recommendations.

[0115] For example, during the flight planning phase, the aviation generative big model can comprehensively consider factors such as NOTAMs, weather conditions, aircraft performance, and crew qualifications. For example, based on the aircraft's maximum range and fuel consumption rate, combined with the impact of wind direction and speed on fuel consumption from meteorological data, and the impact of airspace restrictions in NOTAMs on air routes, the model can calculate fuel requirements and flight times for different flight planning scenarios. For example, when encountering unexpected situations (such as severe weather or temporary NOTAMs), the aviation generative big model can reassess the feasibility of existing flight plans. For example, if an airport suddenly issues a heavy fog weather advisory, the aviation generative big model can analyze information such as visibility and estimated duration of the fog. Combined with the airport's takeoff and landing standards and flight queues, it can recommend which flights should be delayed and which flights can choose alternate airports. It also provides the basis for selecting alternate airports (such as the airport's support capabilities and distance from the destination airport).

[0116] For example, flight planning involves multiple factors (such as NOTAMs, weather conditions, aircraft performance, and crew qualifications). The impact of each factor on fuel requirements and flight time can be quantified and weighted to optimize decision-making. The embodiments of this disclosure do not limit the specific weight assignments. Multimodal data fusion can also be used to capture interactive relationships.

[0117] For example, empirical weighting can be used. For example, expert knowledge and historical data analysis can be used to determine the degree of influence of each factor on the final outcome. For example, aircraft performance may have a greater impact on flight time, while weather conditions may have a more significant impact on fuel consumption. For different decision factors (such as wind speed, wind direction, and airspace restrictions), preliminary weights can be set based on historical cases or the opinions of domain experts. For example, weights can be continuously optimized through regression analysis, A / B testing, and other methods.

[0118] For example, data-driven weight learning can be used. For example, machine learning models can be trained to learn the weights of different factors influencing flight schedules. For feature selection, feature engineering can be used to identify key factors influencing flight schedules. Machine learning algorithms (such as decision trees, random forests, and neural networks) can then be used to calculate the relative importance of each factor. Methods such as gradient descent and optimization algorithms can be used to automatically adjust the weight of each factor, ensuring that the model's output predictions are as accurate as possible.

[0119] For example, multimodal data fusion can combine information from different sources (such as meteorological data, NOTAM data, and aircraft performance data) to create a model. This data can be fused and weighted using deep learning models (such as multi-input neural networks) or ensemble methods. Each input data source (such as meteorological data and NOTAM data) can be processed using a separate neural network or model. The results are then fused through weighted summation, attention mechanisms, and other methods to produce the final decision output.

[0120] Regarding converting meteorological influences into fuel calculations, since meteorological conditions (such as wind speed, wind direction, temperature, etc.) have a significant impact on fuel consumption, it is necessary to effectively convert meteorological data into input for the fuel calculation model.

[0121] For example, a fuel consumption model based on meteorological conditions can use a physical model to describe the relationship between meteorological conditions and fuel consumption. For example, a physical model can include wind speed and fuel consumption, temperature and air density, or flight altitude.

[0122] For example, Wind Speed ​​and Fuel Consumption can include calculating fuel consumption based on the effects of wind speed and direction. For example, flying into a headwind will increase an aircraft's fuel consumption, while flying with a tailwind will reduce fuel consumption. This can be modeled using simple linear or nonlinear relationships. For example, Temperature and Air Density can include how temperature affects aircraft engine efficiency and air density, thus accounting for the effects of temperature in fuel consumption calculations. For example, Flight Altitude can include how flight altitude directly affects air density and drag, which in turn affects fuel consumption. Flying at higher altitudes is generally more efficient.

[0123] For example, at least one embodiment of the present disclosure may also enable data-driven fuel consumption predictions. For example, a machine learning model (such as a regression model, a support vector machine, etc.) may be used to learn the impact of meteorological data and aircraft performance parameters on fuel consumption, and the complex nonlinear impact of meteorological conditions on fuel consumption may be captured by training on historical data.

[0124] For example, at least one embodiment of the present disclosure can also use historical flight data, combined with en-route weather data (wind speed, wind direction, temperature, etc.) and flight parameters (such as altitude and speed), to construct a training set. Regression models (such as random forest regression and XGBoost) can be trained to predict fuel consumption under different weather conditions.

[0125] For example, at least one embodiment of the present disclosure can also utilize hybrid models for fuel consumption forecasting, combining physical models with data-driven models to simultaneously consider the laws of physics and patterns in historical data. For example, a physical model can be used to first calculate a baseline value for fuel consumption, and then a data-driven model can be used to modify or fine-tune this baseline value, resulting in a more accurate forecast.

[0126] For example, at least one embodiment of the present disclosure can also combine physical models with data-driven methods, which can not only utilize the known physical laws of meteorological conditions, but also learn potential nonlinear relationships from historical data, thereby improving the accuracy and flexibility of the model.

[0127] For example, multi-model integration can combine physical models with machine learning models into a single integrated system. For example, a physical model can be used for preliminary fuel consumption calculations, while a machine learning model can be used to fine-tune fuel consumption in specific situations. For example, a physical model can be used to calculate a baseline fuel consumption value. A machine learning model can then modify this baseline value based on historical flight data and weather changes. For example, if fuel consumption under certain weather conditions exhibits anomalies, the machine learning model can detect and correct these deviations.

[0128] For example, the parameters of the physical model and the data-driven model can be optimized simultaneously to enable them to work together. Using optimization algorithms (such as genetic algorithms and particle swarm optimization), the outputs of the physical and data-driven models can be aligned, thereby providing more accurate fuel consumption predictions. For example, when training the data-driven model, the predicted values ​​output by the physical model can be used as training targets, while also considering the constraints of the physical model (such as flight altitude and speed).

[0129] In at least one embodiment of the present disclosure, a large, code-based, interactive route design system, utilizing a generative aviation model in conjunction with a geographic information system and a flight performance database, can generate optimal route plans based on real-time weather conditions and NOTAM restrictions. For example, by accurately calculating the impact of wind direction and speed on flight, route turning points and flight altitudes can be appropriately adjusted, effectively reducing fuel consumption and operating costs while improving flight efficiency.

[0130] In some embodiments of the present disclosure, en-route weather data and NOTAM information can be input into a risk analysis model built based on an aviation-generative macromodel. The macromodel comprehensively considers the direct impact of meteorological conditions on flight safety (such as severe weather such as strong winds, heavy rain, and thunderstorms) as well as the indirect risks brought about by NOTAMs (such as the risk of detours caused by temporary airspace control). By learning from a large amount of historical data using a machine learning algorithm, it quantifies the risk level. For example, under specific meteorological conditions and NOTAM restrictions, the model can calculate the probability of a flight being delayed, diverted, or even involved in an accident, and provide corresponding response strategies for different risk levels, such as increasing fuel reserves and selecting a suitable alternate airport.

[0131] For example, building a risk analysis model can use multi-source meteorological data and multi-source NOTAM data as input variables, and risk events such as flight delays, diversions, and accidents as output variables. The aviation generative large model uses a large amount of historical data to learn the risk probability distribution under different meteorological conditions and NOTAM combinations. For example, when analyzing the risk of a flight flying a specific route, the model considers meteorological changes along the route (such as whether it passes through thunderstorm areas, strong wind belts, etc.) and potential risks in the NOTAM (such as the complexity of temporary control areas). Through a comprehensive evaluation of these factors, the model outputs the risk level of the flight at different stages (takeoff, cruising, landing) and provides corresponding risk mitigation measures, such as adjusting the flight altitude, speed, or changing the route.

[0132] In at least one embodiment of the present disclosure, a risk analysis model may be constructed by adopting strategies such as hybrid modeling methods, route segmentation modeling, or domain-specific feature engineering.

[0133] For example, for hybrid modeling methods, risk analysis models can be built based on rule-learning hybrid models and statistical modeling combined with machine learning.

[0134] For example, in a rule-based learning hybrid model, a rule-based system can be introduced in the pre-processing and post-processing stages of the model. The rules can be based on aviation safety standards or expert experience (e.g., flight risk increases in thunderstorm areas, and the speed must be reduced when the wind speed in low-altitude airspace exceeds a certain threshold).

[0135] For example, when combining statistical modeling with machine learning, we can first use a statistical regression model based on historical data (such as logistic regression) to estimate the preliminary association between meteorological and advisory data and risks, and then use a deep learning model (such as the Transformer model) to explore more complex nonlinear relationships.

[0136] For example, for Segmented Route Modeling, since routes are spatially structured, aviation risks often manifest differently within specific areas. In particular, route segments (such as takeoff, cruising, and landing) have a significant impact on risk assessment. For example, routes can be segmented, and the entire journey can be divided into multiple stages (such as takeoff, cruising, and landing). Each stage is modeled separately based on different environmental variables (such as weather and navigation notices). For example, during the cruising stage, the main concerns may be factors such as airspace restrictions, weather changes, and temporary control of flight routes, while during the landing stage, more consideration may be given to weather conditions, airport busyness, runway congestion, etc. A dedicated model is established for each segment to conduct risk assessment, and the outputs of each stage are combined to obtain a comprehensive risk score for the entire journey.

[0137] For example, in the calculation of risk probability, the probability distribution model can include Bayesian inference and generative models.

[0138] For example, Bayesian inference can be used to update the risk probability of a flight by combining historical data with real-time weather data using Bayes’ theorem. For example, by modeling the frequency of accidents under similar historical weather conditions and NOTAM combinations, the conditional probability under the current flight mission can be obtained.

[0139] For example, generative models can learn the probability distribution of risk events based on meteorological data, NOTAMs, and other inputs using generative large models such as variational autoencoders (VAEs) or generative adversarial networks (GANs).

[0140] For example, based on event analysis and taking into account the scarcity of aviation events, the probability of accidents or major events can be estimated through rare event models (such as extreme value theory), and data imbalance can be considered in the probability calculation.

[0141] For example, the output of risk probability can include a risk classification. For example, this classification can be performed by applying thresholds to the calculated probability distribution to generate risk levels for flights at various stages (e.g., low risk, medium risk, and high risk). Each risk level is associated with a different response strategy. For example, a low risk risk may require no action; a medium risk risk may suggest route adjustments or altitude adjustments; and a high risk risk risk may suggest an alternate landing or temporary airspace avoidance.

[0142] For example, the classification can be based on the probability distribution of the model output and set thresholds, such as 0-30% for low risk, 30-70% for medium risk, and 70-100% for high risk.

[0143] For example, risk control strategies can define a series of pre-set risk mitigation measures based on different risk levels. For example, low risk: maintain the current flight plan; medium risk: adjust the flight altitude, speed, or route; high risk: require immediate emergency measures such as route deviation or diversion.

[0144] For example, in at least one embodiment of the present disclosure, in order to effectively convert the risk level output by the large model into an action plan, the response strategy can be closely associated with the model prediction results.

[0145] For example, when generating risk mitigation strategies, the system can assign them to corresponding flight phases based on the risk level calculated by the model. For example, during takeoff, when the risk is high, it may be recommended to delay takeoff or select a different runway to avoid adverse weather conditions. For example, during cruising, it may be recommended to change flight altitude to avoid strong winds or thunderstorms. For example, during landing, when the risk is high, it may be necessary to change the landing airport or increase emergency preparedness.

[0146] For example, to optimize decision-making, a decision support system (DSS) combined with a multi-task learning model can provide comprehensive decision support throughout the flight. For example, the model can generate dynamic response strategies based on current weather data, NOTAMs, and historical flight data, and update flight risks in real time.

[0147] In at least one embodiment of the present disclosure, a risk analysis model is combined with aviation domain knowledge, and the accuracy of the model is improved through strategies such as hybrid modeling and route segmentation modeling. In addition, domain-specific techniques (such as Bayesian inference and extreme value theory) are introduced into the calculation of risk probability distribution.

[0148] In some embodiments of the present disclosure, the user interaction module 130 is further configured to generate and display a route plan through a code-interactive route design tool and an aviation generative large model in response to user operations on the operation interface.

[0149] For example, the user interface can provide a visual interface for users (such as dispatchers) to input parameters, adjust route design, and view real-time feedback.

[0150] For example, the code-interactive route design tool can define route logic through code snippets and support collaborative optimization with large models.

[0151] For example, users can enter flight parameters such as the origin, destination, aircraft model, weather conditions, and NOTAM restrictions on the user interface. Based on these parameters, the aviation generative model combines aircraft performance data from the flight performance database (e.g., fuel consumption at different altitudes and speeds, flight stability), and geospatial information from the GIS (e.g., geographical obstacles such as mountains, oceans, and cities) to generate initial route code snippets. For example, the aviation generative model can determine the optimal cruising altitude and flight direction based on wind direction and speed information from meteorological data to minimize fuel consumption. It can also generate route code that avoids restricted areas based on airspace restrictions in NOTAMs. Users can interactively modify and optimize the generated route code on the user interface. For example, based on their experience and familiarity with the local airspace, they can adjust route parameters such as turning points, flight altitude, or speed restrictions. The user interaction module 130 can then provide real-time feedback on the impact of these modified route parameters on flight operations (e.g., changes in fuel consumption, flight time, and risk level), helping users determine the optimal final route plan.

[0152] In some embodiments of the present disclosure, an intuitive user interface can be provided to dispatchers, displaying information such as flight plans, NOTAM details, weather conditions, route planning results, and intelligent decision-making recommendations. Dispatchers can review, modify, and confirm system-generated decision recommendations on this interface. For example, if a dispatcher believes the system-recommended alternate airport is inappropriate, they can select another alternate airport on the interface based on their experience and knowledge of local airports. The system will then reassess the risks and adjust the relevant plan based on the new selection.

[0153] For example, dispatcher operational behaviors and feedback can be collected and used as new data to feed into the aviation generative big model for model training and optimization. For example, if a dispatcher modifies the system's recommended route under a specific circumstance, the system will record this modification and the relevant data at the time (such as weather conditions, NOTAMs, and aircraft performance), allowing the big model to continuously improve the accuracy and rationality of its decisions during subsequent learning.

[0154] In at least one embodiment of the present disclosure, users can finely control route design through code or user interface operations, while simultaneously integrating with the macro model to ensure safety and efficiency. Users can interactively modify and optimize the route code generated by the macro model on the route design platform, flexibly adjusting route parameters based on their own experience and actual circumstances. This personalized design approach fully leverages the advantages of human-machine collaboration, making route planning more aligned with actual operational needs and further improving the refined management of aviation operations.

[0155] In some embodiments of the present disclosure, a code-interactive route planning tool is developed using an aviation generative big model, a geographic information system, and / or a flight performance database.

[0156] For example, leveraging the code generation capabilities of the aviation generative big model, combined with a geographic information system (GIS) and a flight performance database, a code-based interactive route planning tool has been developed. Dispatchers can enter parameters such as the flight's origin and destination, aircraft type, weather conditions, and NOTAM restrictions into the tool, and the big model will generate a route code snippet that meets these requirements.

[0157] For example, based on the wind direction and speed information in the meteorological data, the large model can optimize the turning points and flight altitude of the route to reduce fuel consumption; at the same time, based on the airspace restriction information in the navigation notice, it generates a route path code that avoids the restricted area and displays it visually on the GIS interface. Dispatchers can modify and optimize the code according to actual conditions and ultimately determine the optimal route plan.

[0158] For example, a dedicated route planning software platform could be developed that integrates a large aviation generative model, a geographic information system (GIS), and a flight performance database. Dispatchers would then input flight parameters such as the origin and destination, aircraft type, weather conditions, and NOTAM restrictions.

[0159] For example, the aviation generative large-scale model generates initial route code snippets based on these parameters, combined with aircraft performance data from the flight performance database (such as fuel consumption and flight stability at different altitudes and speeds) and geospatial information from the GIS (such as geographical obstacles such as mountains, oceans, and cities). For example, the model determines the optimal cruising altitude and flight direction to minimize fuel consumption based on wind direction and speed information from meteorological data; and generates route code to avoid restricted areas based on airspace restrictions in NOTAMs.

[0160] For example, dispatchers can interactively modify and optimize generated route codes on the platform. For example, based on their experience and familiarity with the local airspace, they can adjust route parameters such as turning points, flight altitudes, or speed limits. The platform provides real-time feedback on the impact of modified route parameters on flight operations (such as changes in fuel consumption, flight time, and risk level), helping dispatchers determine the optimal final route plan.

[0161] For example, a multilingual terminology knowledge base covering the International Civil Aviation Organization (ICAO), civil aviation administration regulations, and airline manuals can be constructed and mapped into vector form to serve as embedding signals for the input model. For example, terms can be weighted by frequency and importance, with significant weights assigned to terms such as "runway closed" and "hazardous airspace."

[0162] For example, in at least one embodiment of the present disclosure, a predefined mapping rule library can be used to convert fuzzy expressions in natural language (such as "the weather is bad" and "fly faster") into standard aviation parameters, such as flight level (FL), meteorological code (METAR / TAF) (i.e., Meteorological Aerodrome Report / Terminal Aerodrome Forecast), and speed limit (Mach Number).

[0163] For example, the output results are based on the aviation-specific domain language (DSL) to ensure that the output conforms to the standard format of the flight control system or mission planning system, making it easier for downstream systems such as geographic information system (GIS) platforms and route simulation platforms to directly call and execute.

[0164] For example, real-time linkage logic between GIS, performance databases, and large models can be implemented. This allows for hierarchical linkage of data architectures. GIS can provide a spatial reference foundation, including airport coordinates, route paths, terrain elevation, and airspace structure. The performance database can provide dynamic data inputs, such as weather (wind speed / direction / air pressure) and flight characteristics (range / maximum altitude / fuel curve). During inference, the large model can dynamically pull this information, enabling "on-demand inference," improving generation efficiency and decision accuracy. For example, a linkage process can be designed. Upon receiving a natural language task, the model triggers an external data interface. Geographic information and navigation parameters can be injected into the model context. The model dynamically invokes spatial and performance factors (such as "current waypoint wind speed" and "maximum climbable altitude") at each computational layer for path determination and generation control. For example, a knowledge graph intermediary mechanism could be implemented. GIS and performance databases could standardize naming, definition, and query interfaces through an intermediate knowledge graph (KG). The model and external systems could be semantically linked through the KG, avoiding data format coupling and improving system flexibility.

[0165] For example, for optimization methods of geospatial computing, spatial embedding modeling can be performed. For example, spherical coordinate encoding can be used to convert spatial entities such as waypoints, airports, and obstacles into vector representations that can be recognized by the model, supporting the calculation of spatial relationships such as flight distance, angle, and direction. Track segmentation modeling can also be performed, which can divide the complete route into computable segments (such as waypoint pairs, climb segments, cruise segments, and descent segments), and their fuel consumption, time, weather impact and other attributes can be modeled separately to improve accuracy. A fast path evaluation mechanism can also be implemented, which can introduce heuristic searches based on graph computing (such as A*, Dijkstra, etc.) combined with deep network valuation functions to quickly converge to a set of candidate routes for large model optimization selection.

[0166] For example, the basis for selecting quantitative indicators and optimization methods for route optimization, such as the optimality indicator system (which can be dynamically weighted by mission type), can include: fuel consumption, which is calculated based on the flight phase performance model; flight timeliness, which is the flight time that takes into account the influence of flight distance and wind field; safety redundancy, which takes into account the ability to fly around and the distance to alternate airports; airspace conflict, which avoids congested routes and restricted airspace; environmental meteorological risks, such as avoiding strong tropospheric winds and thunderstorms. For example, the basis for selecting optimization algorithms can include: for deterministic scenarios (such as fixed weather + fixed routes), graph search + heuristic methods can be used; for high-dimensional dynamic constraint scenarios, reinforcement learning or reinforcement search methods (such as genetic algorithms and Monte Carlo tree search) are introduced to explore and evaluate candidate routes; for scenarios where the user's expression is unclear, the large model first generates candidate mission intentions and objective functions, and then the system automatically selects the appropriate optimization path.

[0167] In at least one embodiment of the present disclosure, the deep integration of user interaction module 130 with the aviation generative large model enables a closed-loop learning process from user operation to model optimization. This not only improves route planning accuracy and dispatcher efficiency, but also enhances the model's adaptability and intelligence through continuous data feedback.

[0168] In some embodiments of the present disclosure, the NOTAM knowledge graph is based on multi-source NOTAM data and multi-source en-route weather data, and is optimized through an aviation generative big model.

[0169] For example, multi-source NOTAM data may include historical NOTAMs obtained from multiple sources such as aviation management departments and airport authorities of various countries, covering information such as airport facility changes, runway maintenance, and temporary airspace control. The embodiments of the present disclosure are not limited to this.

[0170] For example, multi-source route weather data may include temperature, humidity, air pressure, wind speed, wind direction, and severe weather data such as thunderstorms and typhoons from sources such as integrated weather observation stations and weather satellites, and the embodiments of the present disclosure are not limited to this.

[0171] For example, multi-source en-route weather data can be pre-processed and standardized. For example, natural language processing (NLP) can be used to parse unstructured NOTAMs, extracting key entities (airport, restriction type, time) and relationships (e.g., "Airport A → Runway Closed → Impacts Flight B"). For example, weather data can be tied to en-route spatial coordinates for spatiotemporal alignment, such as mapping thunderstorm area polygons to latitude and longitude ranges in a geographic information system (GIS).

[0172] For example, in at least one embodiment of the present disclosure, optimizing the NOTAM knowledge graph through an aviation generative large model may include generating synthetic data through the aviation generative model to fill sparse areas (such as simulating the impact of extreme weather events on routes), receiving new notices and weather data, and performing incremental learning on the aviation generative model to update the NOTAM knowledge graph.

[0173] In at least one embodiment of the present disclosure, the safety, efficiency, and intelligence of aviation operations are improved by integrating multi-source NOTAM and meteorological data and utilizing aviation generative big models for dynamic optimization.

[0174] For example, in the embodiments of the present disclosure, the acquisition module 110, the decision support module 120, and the user interaction module 130 can be hardware, software, firmware, or any feasible combination thereof. For example, the acquisition module 110, the decision support module 120, and the user interaction module 130 can be dedicated or general-purpose circuits, chips, or devices, or can be a combination of a processor and memory. The embodiments of the present disclosure do not limit the specific implementation of each of the above modules.

[0175] It should be noted that the above Figure 1 The components and structures of the aviation dispatch system 100 shown are merely exemplary and non-limiting. The aviation dispatch system 100 may further include other components and structures as needed.

[0176] Figure 2 A schematic block diagram of an aviation dispatch generation device provided by at least one embodiment of the present disclosure is shown.

[0177] like Figure 2 As shown, the aviation dispatch generation device 200 includes a data processing module 210 and a knowledge graph module 220.

[0178] The data processing module 210 is configured to generate object entity information based on multi-source NOTAM data and multi-source en-route weather data through an aviation generative big model.

[0179] The knowledge graph module 220 is configured to generate a knowledge graph of NOTAMs based on object entity information.

[0180] For example, the data processing module 210 can be configured to perform text parsing, such as using an aviation generative large model for named entity recognition (NER), extracting entities such as airports (such as "ZBAA"), runways (such as "05R"), and times (such as "2504151200"), and performing nested entity parsing (such as "temporary no-fly zone radius 5 nautical miles" to extract "no-fly zone" and "5 nautical miles").

[0181] For example, the data processing module 210 can be configured to perform data cleansing, such as standardizing the time format to UTC (Coordinated Universal Time) (e.g., "2504151200" is converted to "2025-04-15T12:00Z"), and verifying the validity of the ICAO code (checking the consistency of airport coordinates through the AIXM (Aeronautical Information Exchange Model) database).

[0182] For example, the object entity information includes at least one of geographic location and facility type, airspace structure type, time condition type, or event type entity information.

[0183] For example, geographic location and facility object entity information may include an airport (Airport), whose attributes include the ICAO code (primary key), the IATA code, coordinates (such as ZBAA 40.07° north latitude, 116.59° east longitude), and elevation (such as Beijing Capital Airport elevation 35 meters); a runway (Runway), whose attributes include the runway number (such as "05R / 23L"), length (3800 meters), surface type (asphalt), and the airport to which it belongs (Link to Airport), etc. The embodiments of the present disclosure do not limit this.

[0184] For example, the entity information of the airspace structure object class may include a flight information region (FIR), such as ZSHF (Shanghai Flight Information Region), with attributes including vertical range (ground to FL660) and boundary coordinate string; a prohibited area (ProhibitedArea), such as "temporary prohibited area center 4900N00230E, radius 5NM (Nautical Mile)", with attributes including activity time, affected altitude layer, etc. The embodiments of the present disclosure are not limited to this.

[0185] For example, the entity information of the time condition object may include a time range (TimeRange): effective time (Start: 2025-04-15T10:00Z, End: 2025-04-15T16:00Z), supporting recurrence (such as 08:00-17:00 every day), etc., which is not limited by the embodiments of the present disclosure; environmental conditions (Condition), such as "runway slippery" and "low visibility", and associated impact areas (such as the surface of a certain airport runway).

[0186] For example, event-class object entity information may include closure events (Closure), such as AirportClosure and RunwayClosure, associated affected entities (such as runway 05R) and time ranges; restriction events (Restriction), such as AltitudeRestriction (prohibited from falling below 10,000 feet), SpeedRestriction (speed limit of 250 knots), etc., which are not limited in the embodiments of the present disclosure.

[0187] For example, a large aviation generative model can adopt a hybrid architecture of Qwen2.5 and LLaMA3, optimize the embedding layer for aviation terms (such as "SID (Standard Instrument Departure)" and "RVSM (Reduced Vertical Separation Minimum)"), and associate entities with meteorological data through a cross-modal attention mechanism (such as the "thunderstorm" entity to enhance attention to wind speed and radar reflectivity).

[0188] For example, the object entity relationship information includes at least one of spatial relationship information, temporal relationship information, or impact relationship information.

[0189] For example, a spatial relationship (Structural / Positional) can include "part of": a runway belongs to an airport (Runway → partOf → Airport), such as "05R" belonging to "ZBAA"; "locatedIn": a navigation aid is located in a flight information region (Navaid → locatedIn → FIR), such as "PEK VOR" located in "ZSHF"; and "connectsTo": a taxiway connects to a runway (Taxiway → connectsTo → Runway), such as "Taxiway A1 connects to runway 05R." The present disclosure does not limit this. Here, VOR refers to a VHF Omnidirectional Range (VHF Omnidirectional Range).

[0190] For example, the temporal relationship (Temporal) may include validDuring: the event is valid during a certain time period (Event→validDuring→TimeRange), such as the effective time range associated with the "runway closure event"; scheduledFor: the planned time of the engineering activity (WorkInProgress→scheduledFor→TimeRange), such as "runway maintenance is scheduled to start at 2025-04-15T02:00Z", etc. The embodiments of the present disclosure do not limit this.

[0191] For example, the impact relationship (Impact / Status) may include affects: event impact entity (Event→affects→Runway), such as "thunderstorm weather affects takeoff and landing on runway 05R"; impacts: event impact operation (Event→impacts→Procedure), such as "airspace control affects standard instrument departure procedure (SID)", etc. The embodiments of the present disclosure are not limited to this.

[0192] In some embodiments of the present disclosure, the knowledge graph module 220 may be further configured to generate a NOTAM knowledge graph based on object entity information and object entity relationship information.

[0193] For example, to generate a knowledge graph of NOTAMs, node creation can be performed. For example, object entity information can be obtained from the data processing module 210, and nodes can be created by type (such as Airport node, Runway node), with the primary key being the ICAO code or event ID (to ensure uniqueness).

[0194] For example, exemplary nodes may include: Airport node, label "Airport", attributes {icao: "ZBAA", name: "Beijing Capital Airport"}; RunwayClosure node, label "Event", attributes {type: "Closure", affected_runway: "05R"}.

[0195] For example, edge (relationship) generation can be performed. For example, edges can be automatically added based on logical associations between entities. For example, an "affects" edge can be added from the "RunwayClosure" event node to the "ZBAA" airport node to indicate that the event affects the airport; and a "partOf" edge can be added from the "05R Runway" node to the "ZBAA" airport node to indicate that the runway belongs to the airport.

[0196] For example, a relationship property example may include: a "validDuring" edge includes a time range object {start: "2025-04-15T08:00Z", end: "2025-04-15T12:00Z"}.

[0197] For example, when optimizing the knowledge graph for NOTAMs, the logical reasoning capabilities of the aviation generative model can be leveraged to supplement missing relationships. For example, if the NOTAM mentions "Runway 05R closure causing flight diversion," the model automatically infers the "causes" relationship between the "Runway Closure" event and the "alternate airport" (this requires training with historical case studies).

[0198] For example, a graph neural network (GNN) can be used to perform relationship verification so that the spatial relationship between airspace restrictions and waypoints conforms to the AIXM standard (for example, the distance between the no-fly zone boundary coordinates and the waypoint is ≥5NM).

[0199] For example, a NoSQL graph database (such as Neo4j) can be used to store and query NOTAM knowledge graphs. Nodes store entity types and attributes, while edges store relationship types and weights (such as the severity of the impact). For example, a query might look for "all airports affected by the temporary airspace restrictions in Shanghai between 2015-04-15T10:00Z and 16:00Z."

[0200] In at least one embodiment of the present disclosure, by fusing text parsing with numerical data, for example, from "visibility 800 meters, runway 05R closed", the "low visibility" condition entity and the "runway closed" event entity are extracted simultaneously and associated with the same airport node. And through large-model semantic reasoning, the causal relationship between "airport closure" and "flight delay" is automatically completed without explicit rule definition. Moreover, based on the AIXM standard verification of the coordinate compliance of the airspace entities in the graph, the "no-fly zone" polygon can be made not to overlap with the waypoints, thereby improving the reliability of decision-making. A dynamic knowledge graph can be quickly constructed to provide structured data support for route planning and risk analysis, which is more than 80% more efficient than traditional manual processing and has an entity parsing accuracy of 95%.

[0201] In some embodiments of the present disclosure, the construction of the NOTAM knowledge graph may be achieved through the following method.

[0202] For example, regarding the schema design of a NOTAM chart, node types can include NOTAM, Airport, Runway, Restricted Area, Waypoint / Airway, NavAid, TimeRange, Airspace, and Equipment. Edge types can include AFFECTS: a notification affects a facility or area, CLOSES: closing a relationship, such as a notification closing a runway, ESTABLISHES: establishing, such as establishing a new waypoint or restricted airspace, REVOKES: revoking a previous notification or facility status, LOCATED_AT: indicating the geographical location of an entity, such as a navigation station located at an airport, EFFECTIVE_DURING: a notification or status is effective for a certain time period, etc.

[0203] For example, when generating a knowledge graph for NOTAMs, triples can be extracted using a hybrid approach based on rules and large models. For example, simple relationships (e.g., runway closure, navigation station failure) can be extracted using regularization or template matching; complex relationships (e.g., conditional activation, specific heading, time period diversity) can be extracted using a fine-tuned instruction model or the ReAct inference model. Multiple representations can be normalized, such as grouping "Runway 18L unavailable" and "RWY18L CLOSED" into the same semantic node.

[0204] For example, you can use Neo4j (for property graphs) or RDF / GraphDB (for semantic ontologies) to implement graph database design; you can manage hierarchical namespaces for different entity types (such as NOTAM::ZBAA::20240401); and you can set tracking attributes such as source_doc, last_updated, and status (active / invalid) for entities.

[0205] For example, the update mechanism of the knowledge graph of NOTAM.

[0206] For example, the identification of notification change types can include, for example, new notifications: directly generate new triples and write them into the graph; change notifications: if the number is the same but the text is different → it is considered a "change", and the original entities and relationships are marked as historical versions; cancellation notifications: if the text contains CANCEL NOTAM or REVOKES → mark the original notification relationship as status=invalid; resend notifications: only adjust the format or timestamp of the text → the original notification can be directly overwritten and updated.

[0207] For example, for the announcement time comparison strategy, the first_seen_time and last_updated_time can be saved in the graph for each announcement; if multiple announcements hit the same object (such as the same runway), the triplet with the most recent effective time is retained as the primary version; all versions can be retained to facilitate "tracing back the graph status by time period" in the future.

[0208] For example, the knowledge graph update method for NOTAMs may include scheduled notification parsing → entity / relationship extraction → graph snapshot comparison → conflict resolution → batch storage into GraphDB.

[0209] For example, for the association rule design between the NOTAM knowledge graph and aviation data, association rule classification can be performed.

[0210] For example, association rule classification may include airport entity alignment, runway entity mapping, navigation facility alignment, airspace boundary positioning, and announcement time period and flight matching.

[0211] For example, airport entity alignment can include the requirement that the ICAO / IATA airport mentioned in the notice must exist in the airport database, otherwise it is a temporary airport. For example, ZBAA → "Beijing Capital International Airport".

[0212] For example, runway entity mapping can include runway number normalization, and graph node normalization by RWY+number. For example, Runway 36L → RWY36L.

[0213] For example, the alignment of the navigation facility can include aligning the navigation station identifier to the location, type and other attributes in the AIXM. For example, VOR / DME ZBN→ZBN navigation station entity.

[0214] For example, locating airspace boundaries can include matching an airspace database to calculate its coverage area if the notification contains boundary coordinates or path points. For example, a polygonal airspace can be used to calculate whether it covers a route segment.

[0215] For example, matching the notification time period with the flight may include checking with the flight database whether it affects the scheduled flight. For example, the scheduled take-off and landing time of flight CA123 conflicts with the closed runway.

[0216] For example, in the embodiments of the present disclosure, the data processing module 210 and the knowledge graph module 220 can be hardware, software, firmware, or any feasible combination thereof. For example, the data processing module 210 and the knowledge graph module 220 can be dedicated or general-purpose circuits, chips, or devices, or can be a combination of a processor and memory. The embodiments of the present disclosure do not limit the specific implementation of each of the above modules.

[0217] It should be noted that, in the embodiment of the present disclosure, each module of the air dispatch generating device 200 may correspond to each step of the air dispatch generating method provided by the present disclosure (described below). Figure 2 The components and structures of the aviation dispatch generating device 200 shown are merely exemplary and non-limiting. The aviation dispatch generating device 200 may further include other components and structures as needed.

[0218] Figure 3 A flowchart of a method for training an aviation generative large model provided by at least one embodiment of the present disclosure is shown.

[0219] At least one embodiment of the present disclosure provides a training method for an aviation generative large model, such as Figure 3 As shown, the training method includes steps S300 to S310.

[0220] Step S300: Acquire a training data set determined based on multi-source NOTAM data, en-route weather data, or aircraft parameter data.

[0221] Step S310: Based on the training data set, multi-task training is performed on the aviation generative large model, wherein the aviation generative large model adopts a deep learning framework.

[0222] For example, aircraft parameter data may include performance parameters such as maximum range, fuel consumption rate, climb rate, cruising speed, etc. of different types of aircraft (such as BADA database data), and the embodiments of the present disclosure are not limited to this.

[0223] For example, the multi-source NOTAM data, route weather data or aircraft parameter data can be cleaned, denoised and standardized, such as converting unstructured NOTAM text into structured data and unifying the time and space dimensions of weather data.

[0224] For example, a deep learning framework (such as PyTorch) can be used to build a large aviation generative model, which can be optimized based on the Transformer architecture and support multi-task parallel training. The embodiments of the present disclosure are not limited to this.

[0225] For example, the multiple tasks may include a NOTAM parsing task, a decision support task, a risk analysis task, and / or a route design task.

[0226] For example, the navigation notice parsing task may include training a model to extract key entities (such as airport ICAO code, restricted time, affected area, etc., which are not limited in the embodiments of the present disclosure) from the notice text and constructing knowledge graph nodes.

[0227] For example, auxiliary decision-making tasks may include learning decision logic such as flight delay prediction and diversion demand assessment, and the input includes data such as flight plan, crew qualifications, and fuel reserves. The embodiments of the present disclosure do not limit this.

[0228] For example, a risk analysis task may include learning the risk probability distribution under the combination of meteorological conditions (such as thunderstorms and crosswinds) and navigation notices (such as no-fly zones) through historical data, and outputting risk levels such as delays, diversions, and accidents. The embodiments of the present disclosure are not limited to this.

[0229] For example, route planning tasks can include combining GIS and flight performance data to generate route codes that comply with standards (such as ARINC 424) to optimize fuel efficiency and flight time.

[0230] For example, during the training process, evaluation indicators such as the cross entropy loss function can be used to continuously adjust the model parameters to improve the accuracy and generalization ability of the model.

[0231] For example, loss function design for airline dispatch can include improvements to standard cross-entropy. For example, cost-sensitive cross-entropy can be constructed for flight delay prediction, assigning different loss factors to false positives and false negatives (e.g., increasing the penalty weight for missing high-risk events to 5x). Focal Loss is used for risk assessment tasks to improve the model's recognition capabilities for minority classes (e.g., extreme weather and high-risk advisories). For example, structural constraint loss can also be included in route generation, adding path legitimacy constraints such as path closure checks and segment height continuity. An "illegal route structure penalty" can be added to training to prevent the model from generating code that does not comply with chart specifications.

[0232] For example, to balance the differences in multi-task training, weighted mechanisms such as static task weight initialization, dynamic imbalance adjustment strategy and loss function fusion expression can be adopted.

[0233] Exemplarily, static task weight initialization can be configured as risk analysis task: weight is set to 0.4 (task safety criticality is high); auxiliary decision task: weight 0.3; route design task: weight 0.2; navigation notice parsing task: weight 0.1 (as a basic task). The embodiments of the present disclosure do not limit the specific weight configuration.

[0234] For example, for the dynamic imbalance adjustment strategy, if a task converges quickly on the validation set, its weight can be reduced to increase the training opportunities of the generalization task; you can refer to multi-task balancing algorithms such as GradNorm to dynamically adjust the gradient amplitude of each task to an approximate balance to reduce the dominance of a certain task in the training process.

[0235] For example, the loss function fusion expression can be expressed as:

[0236]

[0237] Among them, λi\lambda_i is the dynamic weight of each type of task, LiL_i is the corresponding loss function, and i is a positive integer.

[0238] For example, the training method of the aeronautical generative large model can include a staged training mechanism.

[0239] For example, the first phase can include a general pre-training phase. For example, multiple open corpora (such as aviation encyclopedias, flight manuals, and news information) can be used for language modeling training. The goals of the first phase can include learning general language representations and basic domain concepts.

[0240] For example, the second stage can include a fine-tuning phase for aerial tasks. For example, supervised training can be performed using a well-labeled multi-task dataset; each task has its own loss function and weights, and a parameter-sharing architecture can be used.

[0241] For example, the third stage (optional) can include task-specific fine-tuning. For example, single-task fine-tuning can be performed on a small, high-value dataset for a critical task (such as risk assessment) to improve reliability and decision consistency in specific application scenarios.

[0242] For example, for NOTAM parsing, evaluation indicators may include F1-score (field level) and structural integrity rate, which can be used to determine whether core fields such as valid restricted areas and time periods are extracted.

[0243] For example, for delay prediction, evaluation indicators may include AUC and false alarm rate / missing alarm rate, for example, the model's ability to predict critical flights can be examined.

[0244] For example, for risk assessment, evaluation metrics may include high-risk recall rate and risk explanation coverage, which can be used to judge the recognition ability of a small number of high-risk samples.

[0245] For example, for route design, evaluation indicators may include route legality rate, planned fuel consumption difference, and airspace conflict rate. For example, the actual flight cost gap and compliance between the generated route and the standard path of the scheduling system can be compared.

[0246] In some embodiments of the present disclosure, the training method further includes: determining a training data set. The training data set can actually be collected in advance and annotated as necessary.

[0247] For example, based on the aviation notice corpus including object entity information annotation, multi-source NOTAM data is parsed.

[0248] For example, a corpus of aviation notices can be constructed. For example, historical navigation notice texts can be collected and manually annotated with object entity information.

[0249] For example, the corpus size can include multiple (e.g., more than 100,000) annotated data and can support multi-language (Chinese and English) term mapping (such as "QNH" corresponding to "corrected sea level pressure").

[0250] For example, to parse multi-source NOTAM data, a named entity recognition (NER) model can be trained using an annotated corpus to extract object entity information from new NOTAMs, such as identifying the airport code (ZBAA), runway number (05R), and closing time (2025-04-15 12:00Z) from “ZBAA RWY 05R CLSD UNTIL 202504151200”.

[0251] For example, the parsed structured data (such as JSON format) can be associated with the knowledge graph nodes to form the input features in the training dataset.

[0252] In some embodiments of the present disclosure, the aeronautical generative big model may include an encoding layer, where the encoding layer is used to convert object entity information into a vector representation corresponding to the aeronautical generative big model.

[0253] For example, for object entity information encoding, the encoding layer can receive pre-processed structured data (such as airport coordinates, notice terms, meteorological parameters) and convert it into a high-dimensional vector representation through the embedding layer.

[0254] For example, for multimodal fusion coding, the encoding layer can support cross-modal interaction based on the multimodal characteristics of aerial data (text, numerical, spatial).

[0255] For example, a coding layer includes multiple sub-coding layers, such as sub-coding layers for aviation terminology and / or geospatial information.

[0256] For example, the implementation of aviation terminology coding layer may include terminology library construction and coding methods.

[0257] For example, for the construction of the terminology library of the aviation terminology coding layer, the terminology sources may include: ICAO standard terminology, airline manuals, NOTAM, AIXM, etc., and may include general aviation, commercial aviation, and ATC fields.

[0258] For example, for multilingual processing of terminology bases, it can include Chinese-English mapping, such as based on ICAO English and Civil Aviation Administration Chinese specifications (such as "QNH" "Corrected Sea Pressure"), combined with the multilingual capabilities of Qwen2.5, to build a bilingual dictionary; storage, such as Elasticsearch, supports dynamic queries of terms, definitions, context, and language; updates support new NOTAM terms (such as temporary airspace restrictions).

[0259] For example, the encoding method for the aviation term encoding layer may include an aviation term embedding layer or a term attention mechanism.

[0260] For example, embedding aviation terms can involve adding an embedding layer to the input layer of a Transformer to map the terms into high-dimensional vectors (e.g., 512 dimensions).

[0261] For example, the term attention mechanism can include adding a term-specific attention layer in the Transformer encoder to highlight key terms.

[0262] For example, the data for the geospatial information coding layer can come from AIXM (waypoints, airspace, XML format), meteorological data (METAR / TAF), and terrain data (OpenStreetMap).

[0263] For example, to implement the geospatial information encoding layer, a geospatial embedding layer is designed to map coordinate, meteorological, and airspace data into vectors. For example, preprocessing can include: AIXM parsing to extract waypoints (latitude / longitude) and airspace restrictions; meteorological data preprocessing to normalize wind speed, air pressure, and other features into numerical features. For example, embedding can include: coordinates, converting longitude and latitude into spherical embeddings (sin / cos transformations) to preserve spatial relationships; meteorological, using a multi-layer perceptron (MLP) to map wind speed and air pressure into vectors; and airspace, using a graph neural network (GNN) to encode airspace topology (such as no-fly zone adjacency). For example, training can include: initialization using LLaMA3-based MLP weights; fine-tuning using AIXM and meteorological data to optimize spatial semantics. For example, output can include: feeding both the geospatial vector and the term vector into a Transformer.

[0264] For example, the fusion of the aviation terminology encoding layer and the geospatial information encoding layer can dynamically associate aviation terminology and geospatial information in a shared embedding space.

[0265] For example, in a shared embedding space, terms and geographic data are mapped to the same dimension (512) and interact through an attention mechanism. For example, for training tasks, term-coordinate associations can be performed, such as predicting the geographic location corresponding to a term (e.g., "Beijing Capital Airport" → coordinates). Route generation can also be performed, such as generating ARINC 424 routes based on terms and spatial data. For example, outputs can include alignment of terms and geographic information in the embedding space (e.g., associating "crosswind limit" with wind speed data).

[0266] For example, for cross-modal interactions, a cross-modal attention layer can be added to the Transformer encoder. For example, term vectors can be used as queries and spatial vectors as keys / values ​​to calculate interaction weights. Spatial data can be weighted based on NOTAM priorities (e.g., increasing the weight of spatial data for the term "urgent"). Output terms can be deeply integrated with geographic information to improve route generation accuracy.

[0267] In some embodiments of the present disclosure, the performance indicators of the aviation generative large model are evaluated based on the leave-one-out method or the cross-validation method.

[0268] For example, the leave-one-out method uses every sample in a dataset as a test set, with the remaining samples used as training sets. This approach is suitable for small-scale, high-value data. For aviation data, it can be used to verify rare events or test new formats. For rare events, such as extreme weather events (typhoons), the sample size is small but the impact is significant. For new format testing, when new NOTAM formats or AIXM versions are released, the sample size is limited and individual verification is necessary.

[0269] For example, the data can be grouped by event type (e.g., "runway closure," "thunderstorm") so that each type of event is tested at least once.

[0270] For example, you can sort by timestamp and prioritize testing the latest NOTAM (such as the no-fly zone on 2025-04-15) to simulate real-time scenarios.

[0271] For example, the training set can contain only the NOTAM preceding the test sample.

[0272] For example, for multimodal processing, each sample can include multimodal data (such as NOTAM text, AIXM coordinates, and meteorological values). During testing, the evaluation model can comprehensively analyze all modes (such as the correlation between NOTAM "thunderstorm" and meteorological wind speed).

[0273] For example, high-risk samples (such as no-fly zones NOTAM) can be given a higher verification weight to achieve zero false negatives. Example: If a no-fly zone is missed, it will be recorded as a 10-fold error.

[0274] For example, if the performance of a certain type of sample (such as "urgent NOTAM") is lower than the threshold (such as F1<0.95), adjust the training data ratio and retrain.

[0275] For example, K-Fold cross-validation can split a dataset into K parts, alternating between using one part for testing and K-1 parts for training. This is suitable for large-scale data. For aviation data, it can be applied to general mission verification or scenarios with diverse geographic or temporal distributions.

[0276] For example, mission validation can also include NOTAM parsing and route design, where large amounts of data are required to assess generalization capabilities. For example, regional / temporal diversity allows the model to adapt to different airspaces and time periods.

[0277] For example, for the selection of K value, K=5 (80% training, 20% testing) can be selected conventionally, and K=10 can be selected for rare events to increase test coverage and reduce deviation. The embodiments of the present disclosure do not limit the specific selection of K value.

[0278] For example, stratified sampling can be performed by task type or data characteristics. For example, stratification by task type can ensure that each fold contains samples from tasks such as NOTAM analysis and risk analysis. For example, stratification by data characteristics can ensure that different airspaces have a certain proportion, data at different times is balanced, or samples from different events are evenly distributed. For example, 1 million data items are divided into 5 folds, with each fold containing 200,000 items, including 2% rare events.

[0279] For example, time series segmentation can be performed. For dynamic data (such as NOTAM and weather), folds can be divided by time to prevent future data leakage. For example, the first fold is tested on January 2025, trained on January 2024 to December 2024; the second fold is tested on February 2025, and so on.

[0280] For example, for multimodal consistency, each fold tests multimodal associations (such as the matching of NOTAM “thunderstorm” with meteorological data) to assess whether the model correctly integrates AIXM coordinates, meteorological values, and terminology.

[0281] For example, for dynamic update verification, when new NOTAM formats or meteorological technologies (such as satellite wind fields) emerge, the new data are preferentially assigned to the test fold. For example, after the AIXM 5.2 update, the test fold contains 5.2 format samples, while the training fold retains the 5.1 format.

[0282] For example, for hybrid validation combining LOO and K-Fold.

[0283] For example, 5-fold cross-validation can be used for regular data to evaluate overall performance (NOTAM analysis, route design). For example, LOO can be used for rare data to target high-risk samples (such as military exercise no-fly zones and typhoon weather).

[0284] For example, when new data such as new formats NOTAM or AIXM versions appear, they can be verified with LOO first and then incorporated into K-Fold after being confirmed to be stable.

[0285] In at least one embodiment of the present disclosure, the performance of the aviation generative large model can be evaluated by the leave-one-out method and / or cross-validation method. The leave-one-out method is used to verify each sample one by one for rare and high-risk scenarios such as extreme weather and temporary no-fly zones, which can improve the accuracy of emergency notice parsing. The cross-validation method can be used for stratified sampling to cover different airspaces, seasons, and notice types, and the generalization ability of the model in complex scenarios such as different airspaces and seasonal weather can be evaluated to ensure the compliance rate of route planning, further reduce the calibration error of risk analysis, and achieve comprehensive verification of the accuracy and robustness of the model, provide quantitative evaluation support for key aviation dispatch tasks such as release decisions and route design, effectively reduce the risk of human omissions, and promote the reliable application and implementation of intelligent assistance systems in actual aviation operations.

[0286] Figure 4 A schematic diagram of an application of an aviation dispatch system provided by at least one embodiment of the present disclosure is shown.

[0287] like Figure 4 As shown, the application method of the aviation dispatch system includes steps S400 to S470.

[0288] Step S400: Data collection: Acquire multi-source NOTAM data (such as airport facility changes, airspace restrictions, etc.) and en-route weather data (such as wind speed, thunderstorms, temperature, etc.) to provide raw data for subsequent processing.

[0289] Step S410: NOTAM processing: parse the collected NOTAM data, extract entity information and construct a NOTAM knowledge graph.

[0290] Step S411: Storing the NOTAM data, storing the processed data for subsequent analysis.

[0291] Step S420: Meteorological data processing: analyzing and processing the meteorological data to construct a meteorological data model.

[0292] Step S421: Store meteorological data and store processed data to provide support for intelligent decision-making assistance.

[0293] Step S430: Intelligent assisted decision-making: comprehensively utilizing the stored NOTAM data and meteorological data to make decisions, such as route planning and intelligent assisted release decisions.

[0294] Step S440: Route planning: Utilize algorithms (such as A* algorithm, etc.) combined with risk assessment models to generate flight paths that avoid dangerous areas and optimize flight paths.

[0295] Step S441: Generate a recommended route.

[0296] Step S450: Intelligent assisted release decision: Based on aviation business rules and machine learning, assist in determining the release decision.

[0297] Step S451: Generate release decision recommendations (such as whether the flight can be released, when to release it, etc.).

[0298] Step S460: User interaction interface: Display the generated auxiliary decision information (recommended routes, release decision suggestions, etc.) to the user (such as the dispatcher) to support interactive operations.

[0299] Step S470: Data feedback and system optimization: Collect user operation feedback to optimize the system's data processing, model building, and decision generation, thereby improving system performance and decision accuracy.

[0300] At least one embodiment of the present disclosure can significantly improve the intelligence, accuracy, and efficiency of aviation dispatch decisions, effectively ensure flight safety, optimize route resource utilization, enhance the system's adaptability and reliability to complex scenarios, and promote the intelligent upgrade of aviation operations.

[0301] It should be noted that the specific operations, functions or beneficial effects of each step in the application method of the aviation dispatch system provided in any embodiment of the present disclosure, for example, can also refer to the relevant description of the aviation dispatch system, aviation dispatch generation device and / or aviation generative large model training method provided in any embodiment of the present disclosure, and will not be repeated here.

[0302] At least one embodiment of the present disclosure provides a method for generating an aviation dispatch.

[0303] The method for generating an air dispatch includes steps S500 to S510. The method for generating an air dispatch can be applied to the air dispatch generating device provided in any embodiment of the present disclosure.

[0304] Step S500: Based on multi-source NOTAM data and multi-source route weather data, object entity information is generated through an aviation generative big model.

[0305] Step S510: Generate a knowledge graph of NOTAMs based on the object entity information.

[0306] For example, the object entity information includes at least one of geographic location and facility type, airspace structure type, time condition type, or event type entity information.

[0307] For example, in some embodiments, step S510 may further include step S511.

[0308] Step S511: Generate a knowledge graph of NOTAMs based on object entity information and object entity relationship information.

[0309] For example, the object entity relationship information includes at least one of spatial relationship information, temporal relationship information, or impact relationship information.

[0310] It should be noted that, for the relevant content of the functions or beneficial effects of each step in the air dispatch generation method provided in any embodiment of the present disclosure, for example, reference can be made to the relevant description of the air dispatch generation device provided in any embodiment of the present disclosure, and will not be repeated here.

[0311] Figure 5 A schematic block diagram of a training device for an aviation generative large model provided by at least one embodiment of the present disclosure is shown.

[0312] At least one embodiment of the present disclosure provides a training device for an aviation generative large model. Figure 5 As shown, the training device 600 of the aviation generative large model includes a determination module 610 and a training module 620.

[0313] The determination module 610 is configured to obtain a training data set determined based on multi-source NOTAM data, en-route weather data, or aircraft parameter data.

[0314] The training module 620 is configured to perform multi-task training on the large aviation generative model based on the training dataset.

[0315] For example, large aviation generative models are built and trained using deep learning frameworks (such as PyTorch, TensorFlow, etc.).

[0316] In some embodiments of the present disclosure, the determination module 610 is further configured to determine a training data set.

[0317] For example, the determination module 610 may be configured to parse multi-source NOTAM data based on a corpus of NOTAM documents that includes object entity information annotations.

[0318] For example, the aviation generative big model includes an encoding layer, which is used to convert object entity information into a vector representation corresponding to the aviation generative big model.

[0319] For example, a coding layer may include sub-coding layers for aeronautical terminology and / or geospatial information.

[0320] For example, the performance metrics of large aeronautical generative models are evaluated based on leave-one-out or cross-validation methods.

[0321] For example, performance indicators include: accuracy of aeronautical notification parsing, reasonableness of decision recommendations, and / or compliance rate of route code generation.

[0322] For example, the multi-task includes a NOTAM parsing task, a decision support task, a risk analysis task, and / or a route design task.

[0323] For example, the aviation generative big model is a big model based on the transformer architecture built using a deep learning framework.

[0324] For example, in the embodiments of the present disclosure, the determination module 610 and the training module 620 can be hardware, software, firmware, or any feasible combination thereof. For example, the determination module 610 and the training module 620 can be dedicated or general-purpose circuits, chips, or devices, or can be a combination of a processor and memory. The embodiments of the present disclosure do not limit the specific implementation of each of the above modules.

[0325] It should be noted that, in the embodiment of the present disclosure, each module of the training device 600 may correspond to each step of the training method of the aviation generative large model provided in the present disclosure.

[0326] Figure 5 The components and structures of the training device 600 shown are merely exemplary and non-limiting. The training device 600 may further include other components and structures as needed.

[0327] Figure 6 A schematic block diagram of an electronic device provided by at least one embodiment of the present disclosure is shown.

[0328] At least some embodiments of the present disclosure also provide an electronic device, such as Figure 6 As shown, the electronic device 700 can be used to implement the air dispatch system, air dispatch generation device, and / or air generative large model training device provided in any embodiment of the present disclosure, for example, it can be implemented as a server, terminal, etc. For example, the air dispatch system, air dispatch generation device, and / or air generative large model training device can be implemented as a sender and / or receiver of instructions / data.

[0329] For example, the processor 701 may be a central processing unit (CPU), a graphics processing unit (GPU), or other forms of processing units with data processing capabilities and / or program execution capabilities; for example, the central processing unit (CPU) may be RISC, X86, or ARM architecture, etc.

[0330] The electronic device 700 in at least one embodiment of the present disclosure may include mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., as well as any devices such as digital TVs, desktop computers, servers, etc., and may also be a combination of any data processing devices and hardware, which is not limited by the embodiments of the present disclosure.

[0331] Figure 6 The electronic device 700 shown is merely an example and should not limit the functions and scope of use of the embodiments of the present disclosure.

[0332] For example, Figure 6 As shown, in some examples, processor 701 can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 702 or programs loaded from storage device 708 into random access memory (RAM) 703. RAM 703 also stores various programs and data required for the operation of the computer system. Processor 701, ROM 702, and RAM 703 are connected to each other via a communication channel 704. An input / output (I / O) interface 705 is also connected to communication channel 704. For example, a storage controller for RAM 703 includes the aforementioned aviation dispatch system, aviation dispatch generation device, and / or aviation generative large model training device.

[0333] For example, the following components can be connected to the I / O interface 705: an input device 706 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 707 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 708 including, for example, a magnetic tape, a hard disk, etc., for example, the storage controller of the storage device 708 includes the above-mentioned data protection device and / or data decryption device; a communication device 709 including, for example, a network interface card such as a LAN card, a modem, etc. The communication device 709 can allow the electronic device 700 to communicate with other devices wirelessly or by wire to exchange data, and perform communication processing via a network such as the Internet. The drive 710 is also connected to the I / O interface 705 as needed. Removable media 711, such as magnetic disks, optical disks, magneto-optical disks, semiconductor memories, etc., are installed on the drive 710 as needed so that the computer program read therefrom can be installed into the storage device 708 as needed. Although Figure 6 The electronic device 700 is shown as including various devices, but it should be understood that it is not required to implement or include all of the devices shown. More or fewer devices may be implemented or included instead.

[0334] For example, the electronic device 700 may further include a peripheral interface (not shown in the figure), etc. The peripheral interface may be various types of interfaces, such as a USB interface, a lightning interface, etc. The communication device 709 may communicate with a network and other devices through wireless communication, such as the Internet, an intranet, and / or a wireless network such as a cellular telephone network, a wireless local area network (LAN), and / or a metropolitan area network (MAN). Wireless communications may use any of a variety of communication standards, protocols, and technologies, including, but not limited to, Global System for Mobile Communications (GSM), Enhanced Data GSM Environment (EDGE), Wideband Code Division Multiple Access (W-CDMA), Code Division Multiple Access (CDMA), Time Division Multiple Access (TDMA), Bluetooth, Wi-Fi (e.g., based on IEEE 802.11a, IEEE 802.11b, IEEE 802.11g, and / or IEEE 802.11n standards), Voice over Internet Protocol (VoIP), Wi-MAX, protocols for email, instant messaging, and / or Short Message Service (SMS), or any other suitable communication protocol.

[0335] It should be noted that, in the embodiments of the present disclosure, the specific functions and technical effects of the electronic device 700 can be referred to, for example, the relevant descriptions of the aviation dispatch system, the aviation dispatch generation device, and the training method of the aviation generative large model in the embodiments of the present disclosure, and will not be repeated here.

[0336] Figure 7A schematic block diagram of another electronic device provided by at least one embodiment of the present disclosure is shown.

[0337] At least one embodiment of the present disclosure further provides an electronic device, such as Figure 7 As shown, the electronic device 800 includes at least one processor 810 and at least one memory 820 .

[0338] For example, memory 820 can be used to non-transitorily store computer-executable instructions (e.g., one or more computer program modules). Processor 810 can be used to execute these computer-executable instructions. When executed by processor 810, these computer-executable instructions can perform one or more steps of the aforementioned method for generating an aviation dispatch and / or the method for training a large, generative aviation model. Memory 820 and processor 810 can be interconnected via a bus or link, using wired, wireless, and / or other forms of communication media, etc., although the present disclosure is not limited in this regard.

[0339] For example, the processor 810 may be a central processing unit (CPU), a graphics processing unit (GPU), or other processing unit with data processing capabilities and / or program execution capabilities. For example, the central processing unit (CPU) may be a RISC, X86, or ARM architecture. The processor 810 may be a general-purpose processor or a dedicated processor, and may control other components in the electronic device 800 to perform desired functions.

[0340] For example, the memory 820 may include any combination of one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-volatile memory may include, for example, read-only memory (ROM), a hard disk, an erasable programmable read-only memory (EPROM), a portable compact disk read-only memory (CD-ROM), a USB memory, a flash memory, and the like. One or more computer program modules may be stored on the computer-readable storage medium, and the processor 810 may execute one or more computer program modules to implement the various functions of the electronic device 800. The computer-readable storage medium may also store various applications and various data, as well as various data used and / or generated by the applications.

[0341] At least one embodiment of the present disclosure also provides a non-transitory computer-readable storage medium for non-temporarily storing computer-executable instructions. When the computer-executable instructions are executed by a computer, the above-mentioned aviation dispatch generation method and / or aviation generative large model training method can be implemented.

[0342] Figure 8 A schematic diagram of a computer-readable storage medium provided for some embodiments of the present disclosure.

[0343] like Figure 8 As shown, the computer-readable storage medium 900 is used to store computer-executable instructions 910. For example, when the computer-executable instructions 910 are executed by a computer, one or more steps in the aviation dispatch generation method and / or the aviation generative large model training method described above can be performed.

[0344] For example, the computer-readable storage medium 900 can be applied to the electronic device 700 or the electronic device 800. For example, the relevant description of the non-volatile computer-readable storage medium 900 can also be referred to Figure 6 The storage device 708 in the electronic device 700 is shown as well as Figure 7 The corresponding description of the memory 820 in the electronic device 800 is not repeated here.

[0345] It should be noted that, in the embodiments of the present disclosure, the specific functions and technical effects of the computer-readable storage medium 900 can be referred to, for example, the above descriptions of the aviation dispatch system, the aviation dispatch generation device and method, and the training method and device of the aviation generative large model, and will not be repeated here.

[0346] There are a few points to note:

[0347] (1) The drawings of the embodiments of the present disclosure only relate to the structures involved in the embodiments of the present disclosure. Other structures may refer to conventional designs.

[0348] (2) In the absence of conflict, the embodiments of the present disclosure and the features therein may be combined with each other to form new embodiments.

[0349] The above description is only a specific embodiment of the present disclosure, but the protection scope of the present disclosure is not limited thereto. The protection scope of the present disclosure shall be based on the protection scope of the claims.

Claims

1. An aviation dispatch system comprising: an acquisition module configured to acquire navigation notice information and en-route weather information; an auxiliary decision-making module configured to generate auxiliary decision-making information based on the NOTAM information and the en-route weather information by using an aviation generative large model and a NOTAM knowledge graph; as well as The user interaction module is configured to display the auxiliary decision information.

2. The aviation dispatch system according to claim 1, wherein: The auxiliary decision information includes first auxiliary decision information and second auxiliary decision information; The auxiliary decision module is further configured to generate the first auxiliary decision information and the second auxiliary decision information based on the NOTAM information and the en-route weather information through the aviation generative large model and the NOTAM knowledge graph, The first auxiliary decision information includes route plan information, and the second auxiliary decision information includes release decision recommendation information.

3. The aviation dispatch system according to claim 2, wherein: The aviation generative large model includes a risk assessment model; The auxiliary decision-making module is further configured to generate the first auxiliary decision-making information based on the NOTAM information and the en-route weather information through a search algorithm, a risk assessment model and / or the NOTAM knowledge graph.

4. The aviation dispatch system according to claim 2, wherein: The auxiliary decision module is further configured to generate the second auxiliary decision information based on the NOTAM information and the en-route weather information through aviation business rules, the aviation generative big model and / or the NOTAM knowledge graph.

5. The aviation dispatch system according to claim 1, wherein: The user interaction module is further configured to generate and display a route plan in response to user operations on the operation interface through the code interactive route design tool and the aviation generative large model.

6. The aviation dispatch system according to claim 5, wherein: The code-interactive route planning tool is developed using the aviation generative big model, geographic information system and / or flight performance database.

7. The aviation dispatch system according to claim 5, wherein: The user interaction module is further configured to provide the data adjusted by the user operation to the aviation generative large model for model training and optimization.

8. The aviation dispatch system according to any one of claims 1 to 7, wherein: The NOTAM knowledge graph is based on multi-source NOTAM data and multi-source en-route weather data, and is optimized through the aviation generative big model.

9. An aviation dispatch generating device, comprising: A data processing module is configured to generate object entity information based on multi-source NOTAM data and multi-source en-route weather data through an aviation generative big model; as well as The knowledge graph module is configured to generate a knowledge graph of navigation notices based on the object entity information.

10. The aviation dispatch generating device according to claim 9, wherein: The object entity information includes at least one of geographic location and facility type, airspace structure type, time condition type or event type entity information.

11. The aviation dispatch generating device according to claim 9, wherein: The knowledge graph module is further configured to generate the NOTAM knowledge graph based on the object entity information and the object entity relationship information. The object entity relationship information includes at least one of spatial relationship information, temporal relationship information or impact relationship information.

12. A training method for a large-scale aviation generative model, comprising: Obtaining a training data set based on multi-source NOTAM data, en-route weather data, or aircraft parameter data; as well as Based on the training data set, multi-task training is performed on the aviation generative large model, wherein the aviation generative large model adopts a deep learning framework.

13. The method for training a large aviation generative model according to claim 12, further comprising: determining the training data set, The multi-source NOTAM data is parsed based on an aviation notice corpus including annotations of the object entity information.

14. The training method for an aviation generative large model according to claim 13, wherein: The aeronautical generative model includes an encoding layer, The encoding layer is used to convert the object entity information into a vector representation corresponding to the aviation generative large model.

15. The training method of the aviation generative large model according to claim 14, wherein: The coding layer includes sub-coding layers for aeronautical terminology and / or geospatial information.

16. The method for training a large aviation generative model according to claim 12, wherein: The performance indicators of the aviation generative large model are evaluated based on the leave-one-out method or cross-validation method. The performance indicators include: aviation notice parsing accuracy, decision recommendation rationality and / or route code generation compliance rate.

17. The method for training a large aviation generative model according to claim 12, wherein: The multiple tasks include a NOTAM parsing task, a decision support task, a risk analysis task and / or a route design task.

18. The method for training a large aviation generative model according to any one of claims 12 to 17, wherein: The aviation generative large model is a large model based on a transformer architecture built using a deep learning framework.

19. An electronic device comprising the aviation dispatch system according to any one of claims 1 to 8 and / or the aviation dispatch generating device according to any one of claims 9 to 11.

20. An electronic device comprising: processor; as well as A memory, wherein the memory stores at least one computer program, and when the at least one computer program is executed by the processor, the training method for the aviation generative large model according to any one of claims 12 to 18 is implemented.

21. A non-transitory computer-readable storage medium for non-temporarily storing computer-executable instructions, which, when executed by a computer, implement the method for training an aviation generative large model according to any one of claims 12 to 18.

Images