An air traffic management decision generation method based on a size model collaborative architecture

Abstract

The application belongs to the technical field of air traffic management, and discloses a civil aviation air traffic management decision generation method based on a large-small model collaborative architecture, which comprises the following steps: processing meteorological data by using a large model, converting meteorological radar image sequences into macro semantic feature vectors through a visual language model and a text embedding model, and searching for the most matched historical meteorological scene and the corresponding pre-tactical height layer distribution strategy based on vector similarity; taking the pre-tactical height layer distribution strategy as a constraint condition, constructing a mixed integer linear programming model with the target of minimizing flight conflicts and maximizing operation economy; and solving the small model by using an optimization solver to generate a tactical height layer distribution scheme for each aircraft in the sector that meets the constraints in real time. Through the large-small model collaboration, the advantages of the large model in macro semantic understanding and retrieval and the high efficiency of the small model in accurate constraint optimization are combined, and the intelligent level of the air traffic management scheme generation is improved.

Description

Technical Field

[0001] This invention belongs to the field of air traffic management technology, specifically relating to a method for generating civil aviation air traffic management decisions based on a collaborative architecture of large and small models. Background Technology

[0002] This section provides only background information relevant to this disclosure and is not necessarily prior art.

[0003] With the continued growth of global air transport, airspace resources are becoming increasingly scarce, posing a severe challenge to air traffic management. Efficiently and safely allocating flight altitudes is one of the core tasks of air traffic flow management. Altitude allocation schemes must simultaneously consider both macro-level strategic strategies and micro-level tactical execution.

[0004] Traditional air traffic management strategy development relies heavily on controller experience, resulting in high subjectivity and low efficiency. In recent years, data-driven intelligent methods have become a research hotspot. Existing technologies can be broadly categorized into two types:

[0005] One type of method is based on historical experience analogy. These methods compare the similarity between the current scenario and historical scenarios in macro-level indicators such as weather and airflow, retrieving and recommending historically adopted management strategies. Their advantage lies in leveraging accumulated practical experience to provide highly interpretable pre-tactical decision support. However, the strategies output by this type of method are typically macro-level and qualitative, lacking precise quantitative guidance for individual flights in specific spatiotemporal dimensions, making it difficult to directly generate actionable flight-level altitude profiles. Furthermore, the processing of input information by this type of method is often limited to structured features (such as images and numerical indicators), making it difficult to directly understand free text describing natural language and transform it into quantifiable constraints.

[0006] Another category is methods based on mathematical optimization models. These methods model the altitude layer allocation problem as a mathematical programming problem containing an objective function and multiple constraints. With the optimization objectives of minimizing conflicts and maximizing efficiency, and under the premise of satisfying constraints such as safety intervals and flight performance, they generate accurate four-dimensional trajectories for the flight set. This type of method has significant advantages in solving precisely constrained optimization problems, capable of handling complex numerical constraints and generating refined tactical solutions. However, its performance is highly dependent on the setting of model boundary conditions. Without reasonable macro-constraint guidance, when facing large-scale and complex airspaces, the solution space of the optimization model becomes too large, potentially leading to excessively long computation times or getting trapped in local suboptimal solutions. Furthermore, it is difficult to effectively integrate macro-level, semantic-based operational rules into the model.

[0007] Therefore, the existing technology has the following prominent drawbacks:

[0008] 1) Disconnect between macro and micro decision-making: Pre-tactical strategy recommendations and tactical trajectory planning are separated. Macro-level solutions based on historical experience and analogy cannot directly guide micro-level optimization models, while optimization models also struggle to effectively utilize qualitative knowledge from macro-level strategies, making it difficult to form a globally optimal solution.

[0009] 2) Limited application scope of large models: In existing technologies, large models are mainly used to process unstructured data such as images and text to support scene retrieval. However, their input format is limited to fixed feature representations, and they cannot directly accept a natural language description (such as a runtime context description) and generate structured high-level constraints based on it. The powerful semantic understanding capabilities of large models have failed to effectively coordinate with the subsequent precise operational optimization decision-making process, and have failed to play the bridging role "from description to constraint".

[0010] 3) Bottlenecks of Small Models: The solution efficiency and effectiveness of tactical-level optimization models (small models) are severely limited by the setting of their constraints. Without effective macroscopic prior knowledge as constraints, the number of variables and constraints in the model is enormous, resulting in long solution times and difficulty in meeting real-time requirements. If the constraints are too broad, the generated solutions may not conform to macroscopic operating rules, leading to poor practicality. The ideal solution should be for the large model to extract constraint boundaries from historical experience, and then for the small model to perform refined solutions. However, current technologies have not yet achieved this synergy. Summary of the Invention

[0011] To address the shortcomings of existing technologies, this invention aims to provide a civil aviation air traffic management decision generation method based on a collaborative architecture of large and small models. By combining the advantages of large models in macroscopic semantic understanding and scene retrieval with the advantages of small models in precise numerical optimization, a complete decision-making closed loop is formed from macroscopic pre-tactical strategies to microscopic tactical solutions.

[0012] A method for generating civil aviation air traffic management decisions based on a size-model collaborative architecture, characterized by the following steps:

[0013] Step 1: Obtain historical scene datasets, which include at least historical weather radar echo image sequences and historical altitude layer allocation strategies within corresponding time segments; extract semantic features from historical weather radar echo image sequences using a general large model, and store the resulting historical semantic feature vectors in association with the corresponding historical altitude layer allocation strategies; the large model is a pre-trained deep neural network with a transformer architecture that takes images and text as input and text as output, and here a publicly available general multimodal large model is used;

[0014] Step 2: Obtain the meteorological information to be decided, process the meteorological information to be decided through a large model, and generate the current semantic feature vector;

[0015] Step 3: Perform similarity matching between the current semantic feature vector obtained in Step 2 and the historical semantic feature vector, and use the historical altitude layer allocation strategy associated with the matching result as the pre-tactical altitude layer allocation scheme; the scheme includes restrictions on the use of altitude layers in specified airspaces, routes or waypoints;

[0016] Step 4: Obtain flight information within the sector to be planned, including flight number, aircraft type, planned route, estimated entry time into the sector, etc.; airspace structure information (including sector boundaries, waypoint coordinates, flight segments, etc.); and the pre-tactical altitude layer allocation scheme obtained in Step 3. Transform the pre-tactical altitude layer allocation scheme into altitude constraints on flights at sector boundary points.

[0017] Step 5: Construct a mathematical optimization model. The mathematical optimization model takes minimizing flight conflicts within the sector or minimizing flight conflicts while maximizing flight operation economy as the objective function, and takes flight performance, safety interval and the altitude constraint mentioned in Step 4 as the constraint conditions to model the altitude profile of the flight set within the sector.

[0018] Step 6: Solve the mathematical optimization model constructed in Step 5 using an optimization solver to generate a tactical-level altitude layer allocation scheme. The tactical-level altitude layer allocation scheme includes the specific flight altitude layers assigned to each flight in the sector at different time periods.

[0019] In some embodiments, step 1 specifically includes:

[0020] Step 1-1: For each frame of the historical weather radar echo image sequence, use the open-source general visual language model (VLM), such as Qwen3-VL 32B, to generate the corresponding natural language description text based on preset prompt words.

[0021] Step 1-2: Extract all natural language description text of a scene consisting of consecutive time frames. The keywords are concatenated chronologically to construct prompt words, which are then input into a general large language model (LLM) to generate a summary text describing the macroscopic meteorological evolution of the scene. ;

[0022] Steps 1-3: Compile the summary text Input to text embedding model Generate fixed-dimensional historical semantic feature vectors :

[0023]

[0024] Steps 1-4: Convert the historical semantic feature vector Historical height layer allocation strategy within this time segment Establish mapping relationships and store them in the database. The historical height layer allocation strategy... This includes, but is not limited to: restrictions on the set of available altitude layers for specific waypoints, flight segments, or sector entrances and exits, and restrictions on the direction of altitude layer usage.

[0025] In some embodiments, step 2 specifically includes:

[0026] Step 2-1: Obtain the meteorological information to be decided, which includes natural language text description and / or summary text generated after the meteorological radar echo image sequence to be decided is processed by the visual language big model and the big language model;

[0027] Step 2-2: Using a text embedding model The natural language text description or summary text is vectorized to generate the current semantic feature vector. .

[0028] In some embodiments, step 3 specifically includes:

[0029] Step 3-1: Calculate the current semantic feature vector semantic feature vectors of historical scenes The similarity is calculated using cosine similarity, and the formula is:

[0030]

[0031] in, This is the current semantic feature vector. For the first Semantic feature vectors of historical scenes, Represents the vector dot product. The L2 norm of a vector;

[0032] Step 3-2: Extract similarity The historical height layer allocation strategy associated with the highest historical scene As an output of the pre-tactical altitude layer allocation scheme, the pre-tactical altitude layer allocation scheme includes restrictions on the use of altitude layers in designated airspaces, routes, or waypoints, when there is similarity among multiple historical scenarios. When the altitude is similarly high, the most suitable historical scenario is selected manually. This scheme imposes macro-level constraints; for example, flights passing through the "XXX" waypoint and heading towards "YYY" can only use altitude-level data. 9200 meters.

[0033] In some embodiments, step 4, specifically constraining the flight's altitude at the sector boundary point, involves setting the entry sector altitude of the flight to be assigned an altitude layer. and departure sector height The height layer usage is restricted in the historical height layer allocation strategy obtained in step 3.

[0034] In some embodiments, step 5: tactical phase, constructing a mathematical optimization model.

[0035] During the tactical phase, a small model based on mixed-integer linear programming is constructed to accurately generate flight altitude profiles that satisfy pre-tactical constraints for each aircraft in the planning pool. The modeling process of the mathematical optimization model specifically includes the following sub-steps:

[0036] Step 5-1: Define the model set and parameters

[0037] First, define the basic set involved in the model to clarify the scope of the planning objects.

[0038] The planning pool is a collection of aircraft. In this embodiment, it refers to all flights that will enter the target sector and be assigned an altitude layer within a specific future time window.

[0039] The frozen pool is a collection of aircraft. It refers to flights that are already within the target sector at the start of planning or for which explicit altitude instructions have been issued and cannot be changed. The position and altitude information of the frozen pool flights are used as model input for conflict identification.

[0040] The planning pool is a collection of flight pairs that violate horizontal safety separations between aircraft. .

[0041] The group of flight pairs that violate horizontal safety separation between aircraft in the planning pool and the frozen pool. .

[0042] : The set of all flight pairs with potential level conflicts.

[0043] Planning pool aircraft The planning time domain refers to the set of time steps included in the aircraft's planning time domain. The continuous time period within the target sector.

[0044] Frozen pool aircraft The set of time steps within the planning time domain.

[0045] For any flight pair This is the set of specific time steps for the flight that violated the horizontal safety separation (e.g., horizontal distance less than 13.704 km). This set was obtained by extrapolating the flight's horizontal trajectory using an air traffic simulation engine.

[0046] The set of planned altitude adjustments an aircraft will make within a sector, for example... This indicates that a maximum of two increases are allowed.

[0047] : A set of altitude profiles, where 1 represents the first altitude profile and 2 represents the second altitude profile. By defining two profiles, an "altitude profile interval" is formed to accommodate the uncertainty of the horizontal track.

[0048] Aircraft The available altitude layer set. This set is one of the core constraints of the pre-tactical phase altitude layer allocation scheme; for example, for a flight on a specific heading, This altitude may be constrained by pre-planned tactical strategies.

[0049] Planning pool aircraft The precise time step corresponding to leaving the sector.

[0050] Secondly, define the key parameters required for the model, which are determined before the model is solved.

[0051] Frozen pool aircraft At time step The known height (unit: meters).

[0052] Planning pool aircraft Initial altitude upon entering the sector (in meters). This altitude may be subject to pre-tactical constraints.

[0053] Planning pool aircraft Protocol altitude (in meters) when leaving the sector. This altitude is also subject to pre-tactical plan constraints.

[0054] The maximum allowable altitude adjustment range for a single aircraft flight (in meters), for example, set to 600 meters.

[0055] The minimum allowable altitude adjustment for an aircraft in a single jump (unit: meters), for example, set to 300 meters.

[0056] The time interval (in seconds) during which the model discretizes the continuous time domain, for example, set to 30 seconds.

[0057] The minimum time interval (in seconds) that must be met between two consecutive altitude adjustment actions of the same aircraft, for example, set to 60 seconds.

[0058] Aircraft The rate of ascent (in meters per 30 seconds). To simplify the model, a uniform ascent is assumed, and approximations are made based on nominal performance, such as 100 meters per 30 seconds or 200 meters per 30 seconds.

[0059] Aircraft The rate of descent (unit: meters / 30 seconds) is approximated using the same method as above.

[0060] The minimum vertical safety separation (unit: meters) is 300 meters in airspace where RVSM is implemented.

[0061] A sufficiently large positive number to handle logical constraints in mixed-integer linear programming (Big M method).

[0062] Step 5-2: Define decision variables

[0063] To characterize the dynamic altitude profile of an aircraft, the following decision variables are defined:

[0064] : Status indicator variable; when the value is 1, it indicates that the aircraft The On each height profile, at time step At the start time, the execution of the first... The next adjustment action.

[0065] : Status indicator variable; when the value is 1, it indicates that the aircraft The On each height profile, at time step The start time has ended the execution of the first step. The next adjustment action.

[0066] : Direction indicator variable; when the value is 1, it indicates the aircraft The The first height profile The next adjustment action is a climb; when the value is 0, it means that the next adjustment action is a descent.

[0067] : Altitude layer occupies a variable; when the value is 1, it indicates an aircraft The On each height profile, at time step It is in a level flight state, and its flight altitude is at the upper altitude level. .

[0068] : Conflict indicator variable, when the value is 1, it indicates a conflict between flight pairs. At time step Violation of vertical safety separation .

[0069] : Continuous decision variables, representing the aircraft in the planning pool In the At a height profile, at time step Precise altitude at the end (in meters).

[0070] Continuous auxiliary variable, representing aircraft In the The first height profile During the second adjustment process, at the time step The absolute value of the vertical velocity within. Its value is based on... The value is automatically determined as or .

[0071] : Increase the status indicator variable; when the value is 1, it indicates that the aircraft The On each height profile, at time step The inside is in the process of altitude adjustment (i.e., neither before nor after the start); when the value is 0, it indicates that it is in level flight.

[0072] Step 5-3: Constructing Constraints

[0073] The model's constraints are designed to ensure that the generated flight altitude profile is physically feasible, compliant with regulatory rules, and strictly adheres to the altitude layer usage restrictions issued during the pre-tactical phase.

[0074] (1) Rate construction constraint: associate the direction decision variable with the performance parameter.

[0075]

[0076] (2) Boundary height constraint: enforce pre-tactical plans and inbound / outbound fan protocols.

[0077] Sector height limit:

[0078]

[0079] in for The minimum time step in.

[0080] Exit sector height limit:

[0081]

[0082] in If the flight does not have an outgoing fan protocol altitude requirement, this constraint can be relaxed to a range constraint.

[0083] (3) Height profile size constraint: Define the upper and lower bounds of the height profile interval. The first height profile represents the case of "latest height adjustment", and its height value is always greater than or equal to the second height profile (representing the case of "earliest height adjustment").

[0084]

[0085] (4) Altitude dynamic recursive constraint: describes the change law of aircraft altitude over time.

[0086]

[0087]

[0088] This formula shows that only when... The height only changes during the adjustment process; the direction of the change is from... Control (1 for increase, -1 for decrease).

[0089] (5) Decision variable logical consistency constraint: ensure the time logic of the adjustment action.

[0090] Irreversible action:

[0091]

[0092] The end was later than the start:

[0093]

[0094] (6) Single adjustment range constraint: Limit the amount of height change for each adjustment.

[0095] Upper limit constraint:

[0096]

[0097] Lower bound constraint:

[0098]

[0099] (7) Altitude interval constraint: Ensure that there is sufficient level flight time between two adjacent altitude adjustments.

[0100]

[0101] (8) Consistency constraint of elevation direction: ensure that the direction remains unchanged during the same elevation process.

[0102]

[0103] (9) Profile interval consistency constraint: ensure that the height adjustment range and direction of the first and second profiles are the same to form an effective time window.

[0104] The same increase:

[0105]

[0106] The adjustment direction is the same:

[0107]

[0108] in .

[0109] (10) Altitude layer feasibility constraint: Ensure that the altitude during the level flight phase belongs to the set of available altitude layers.

[0110] First, define the adjustment state:

[0111]

[0112] Then, when in level flight At that time, altitude Must equal A specific value in :

[0113]

[0114]

[0115]

[0116] (11) Vertical Conflict Identification Constraint: Conflicts are identified using altitude profile intervals, providing a large safety margin. The criterion is whether the vertical distance between the second altitude profile (lower one) of one pair of aircraft and the first altitude profile (higher one) of the other is less than [missing information]. .

[0117] For the planned pool of flights :

[0118]

[0119]

[0120]

[0121]

[0122]

[0123]

[0124] For the flight pairs in the planning pool and the freezing pool Due to the altitude of the frozen pool It is certain that the above formula can be used to... and Replace all with Apply constraints.

[0125] Step 5-4: Construct the objective function

[0126] The model supports multi-objective optimization to meet the interests of different stakeholders.

[0127] (1) First objective (security): Minimize the duration of conflict between all aircraft in the sector.

[0128]

[0129] (2) Second objective (economic efficiency): Maximize the proportion of high-altitude flight time of planned pool aircraft within the sector to reduce fuel consumption.

[0130]

[0131] Step 5-5: Model Solving Strategy

[0132] The hierarchical sequence method is used to handle the above multi-objective problem.

[0133] Phase 1: Considering only the objective function (Minimize conflict), ignore objective function Solve this single-objective MILP model to obtain the optimal objective function value. .

[0134] Phase Two: Applying Constraints Add it to the original model, and then replace the objective function with... (Maximize the time at high altitude), and solve again.

[0135] This method seeks the most economically efficient flight altitude profile while strictly ensuring safety.

[0136] In some embodiments, step 6 specifically involves solving the mixed-integer linear programming model constructed in step 5 using a commercial optimization solver (such as Gurobi or CPLEX). The solver will output the decision variable values ​​for all aircraft in the planning pool at each time step, thus forming a complete tactical-level altitude layer allocation scheme. This scheme includes the number of altitude adjustments for each aircraft, the start time, end time, direction, magnitude, and precise flight altitude for different time periods.

[0137] In some embodiments, the method further includes:

[0138] Step 7: For potential conflicts that the tactical-level altitude layer allocation scheme cannot resolve, output alarm information and transfer the case to the controller for manual handling or call other conflict resolution algorithms for intervention.

[0139] The beneficial effects of this invention are as follows:

[0140] By creatively applying a large model to process unstructured meteorological information, macroscopic pre-tactical strategies are retrieved, providing reasonable and efficient constraint boundaries for the small model (optimization model). This not only reduces the solution space of the optimization model and improves its efficiency, but also makes the generated tactical solutions more consistent with macroscopic operational rules, achieving collaborative decision-making where "the large model makes strategies and the small model makes tactics."

[0141] The semantic understanding capability of the large model enables the system to understand the natural language meteorological descriptions input by controllers, making the interaction more intelligent. The introduction of pre-tactical constraints means that the optimization model's results are no longer purely theoretical mathematical optimal solutions, but rather highly practical solutions.

[0142] The decision-making process is clear and transparent, with pre-tactical plans traceable to specific historically similar scenarios, and tactical plans having explicit mathematical optimization logic. The hierarchical decision-making architecture also ensures that if one part of the system fails, other parts can still function independently, improving overall robustness. Attached Figure Description

[0143] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0144] Figure 1 This is an overall flowchart of the civil aviation air traffic management method based on big-small model collaboration provided by the present invention;

[0145] Figure 2 This is the interface for the pre-tactical phase air traffic management decision generation and recommendation system;

[0146] Figure 3 This is the interface for the tactical phase air traffic management decision generation and recommendation system;

[0147] Figure 4 This is a schematic diagram of the sector structure in the embodiment;

[0148] Figure 5 This is the result of the intra-sector flight altitude profile interval planning for flight CCA4592, considering only the primary objective (safety).

[0149] Figure 6 This is the result of the intra-sector flight altitude profile interval planning for flight CCD1311, considering only the primary objective (safety).

[0150] Figure 7 This is the result of the intra-sector flight altitude profile interval planning for flight CES238, considering only the primary objective (safety).

[0151] Figure 8 This is the result of the intra-sector flight altitude profile interval planning for flight CES9195, considering only the primary objective (safety).

[0152] Figure 9 This is the result of the intra-sector flight altitude profile interval planning for flight CUA5227, considering only the primary objective (safety).

[0153] Figure 10 This is the result of the intra-sector flight altitude profile interval planning for flight CUA5357, considering only the primary objective (safety).

[0154] Figure 11 This is the result of the intra-sector flight altitude profile interval planning for flight DKH1145, considering only the primary objective (safety).

[0155] Figure 12 This is the result of the intra-sector flight altitude profile interval planning for flight DKH1715, considering only the primary objective (safety).

[0156] Figure 13 This is the result of the intra-sector flight altitude profile interval planning for flight CCA4592, taking into account both the first objective (safety) and the second objective (economy).

[0157] Figure 14 This is the result of the intra-sector flight altitude profile interval planning for flight CCD1311, taking into account both the first objective (safety) and the second objective (economy).

[0158] Figure 15 This is the result of the intra-sector flight altitude profile interval planning for flight CES238, taking into account both the primary objective (safety) and the secondary objective (economy).

[0159] Figure 16 This is the result of the intra-sector flight altitude profile interval planning for flight CES9195, taking into account both the primary objective (safety) and the secondary objective (economy).

[0160] Figure 17 This is the result of the intra-sector flight altitude profile interval planning for flight CUA5227, taking into account both the first objective (safety) and the second objective (economy).

[0161] Figure 18 This is the result of the intra-sector flight altitude profile interval planning for flight CUA5357, taking into account both the first objective (safety) and the second objective (economy).

[0162] Figure 19 This is the result of the intra-sector flight altitude profile interval planning for flight DKH1145, taking into account both the first objective (safety) and the second objective (economy).

[0163] Figure 20 This is the result of the intra-sector flight altitude profile interval planning for flight DKH1715, taking into account both the first objective (safety) and the second objective (economy). Detailed Implementation

[0164] The technical solutions in the embodiments of the invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0165] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0166] Example

[0167] This embodiment takes an air traffic management case in the Guangzhou Area Control Sector (ZGGGAR31) during the summer of 2025 as an example.

[0168] Step 1: Historical Data Processing and Feature Library Construction

[0169] Historical operational data for the Guangzhou area in 2025 was collected. Radar echo data was obtained from image data saved hourly. Historical altitude layer allocation strategies were extracted from system logs.

[0170] Image semantic description: This embodiment uses the Qwen3-VL 32B Visual Language Model (VLM) to generate a description for each frame of historical radar image. The prompt is: "Please use a single sentence in Chinese to professionally describe the intensity, main affected area, and morphology of convective weather in this weather radar image." For example, the description of a certain image frame... The radar echo indicated that there was a convective cloud cluster in the northwest, distributed in a band, moving southeast.

[0171] Scenario summary: Using a 6-hour time window ( The six generated text descriptions in the window are arranged into a list in chronological order. In this example, the Qwen3-32B Large Language Model (LLM) is input for summarization. The prompt is: "The following is a sequence of weather radar descriptions every 60 minutes over a 6-hour period. Please generate a coherent summary describing the overall weather evolution over these 6 hours." The resulting summary text is then obtained. .

[0172] Feature vectorization: This embodiment uses the BGE-large-zh-v1.5 text embedding model (output dimension) (This will summarize the text) Vectorized into a 1024-dimensional historical semantic feature vector .

[0173] Associative storage: This 1024-dimensional feature vector The historical height layer allocation strategy recorded in the logs for the corresponding time segments Establish the mapping relationship and store it in the database.

[0174] Step 2: Current meteorological information processing

[0175] Based on the weather forecast, the controller inputs a natural language description into the system interface. Upon receiving this text, the system directly calls the BGE-large-zh-v1.5 text embedding model used in step 1, vectorizing it into a 1024-dimensional semantic feature vector. .

[0176] Step 3: Similar Scene Retrieval and Strategy Recommendation

[0177] System Calculation With all in the database Cosine similarity:

[0178]

[0179] After traversing the database, retrieve the historical scene with the highest similarity (e.g., similarity score). ), and the historical height layer allocation strategy associated with this scenario. After manual screening, the system will apply this strategy. As a pre-tactical altitude layer allocation scheme, it is pushed to the tactical planning module and serves as a constraint for subsequent optimization.

[0180] Step 4: Data Preparation and Constraint Transformation in the Tactical Phase

[0181] This embodiment uses the ZGGGAR31 sector in the Guangzhou area as the experimental scenario, and the eight aircraft scheduled to enter this sector between 11:00 and 11:05 on June 1, 2025, are used as the planning pool. The 11 aircraft currently flying within the sector serve as a freezing pool. .

[0182] Flight information is obtained from flight plans and airspace structure data. Pre-tactical plans are then developed. This translates into constraints on flights in the planning pool. For example, if a flight in the planning pool is scheduled to fly north after passing a certain waypoint, then according to... The flight altitude level within the sector must be selected from the specified altitude level.

[0183] Steps 5 and 6: Tactical Phase Model Construction and Solution (Taking ZGGGAR31 sector as an example)

[0184] Set and parameter initialization:

[0185] Discrete time interval Seconds. For the eight planned pool flights, based on their flight plans and airspace structure, the horizontal trajectory is simulated using an air traffic operation simulation engine to determine their planned time domain. The simulation simultaneously identifies all potential horizontal conflict pairs and conflict time steps. Based on the strategy output from the pre-tactical phase, the set of available height layers is defined. The maximum elevation adjustment is set. Meters, minimum adjustment range Meters. Set the climb / descent rate. Refer to the nominal performance of the A320 model; for simplified calculations, meters / 30 seconds or meters / 30 seconds. Set the minimum vertical interval. rice.

[0186] Model building and solution:

[0187] Based on the initialization data in step 5-1, the system automatically constructs all the constraints and objective functions described in steps 5-3 and 5-4, generating a mixed integer linear programming (MILP) model.

[0188] In the initial experiments, only the safety objective was considered. The Gurobi optimization solver is invoked to solve the model.

[0189] Result: The solution took approximately 0.78 seconds. The optimal solution was obtained, and the objective function value was... This indicates that, within the constraints of pre-tactical altitude layer allocation, all potential flight conflicts within a sector can be completely avoided by optimizing and adjusting the altitude profile of flights in the planning pool.

[0190] Output scheme: such as Figures 5 to 12 As shown, this represents the planned altitude profile range for each flight. The horizontal axis represents the time when a flight enters the sector, and the vertical axis represents the flight recommendation. The polygon marked with a triangle is the first altitude profile, which is the upper limit of the recommended altitude layer. The polygon without a triangle is the second altitude profile, which is the lower limit of the recommended altitude layer (it can overlap with the upper limit). Flights should operate between the two altitude profiles.

[0191] Subsequently, multi-objective optimization is performed. The objective function is switched to... .

[0192] Result: The Gurobi solution was called again, taking approximately 82.20 seconds. The solver, while ensuring zero collisions, allocated flights to higher altitudes whenever possible.

[0193] Final solution: Optimized altitude profiles for each flight are shown below. Figures 13 to 20 As shown in the figure, all flights, while meeting the agreement altitude for entering and exiting sectors and the "east odd, west even" rule, have their flight profiles concentrated in the upper atmosphere, significantly improving operational economy.

[0194] Step 7: Real-time Conflict Mediation

[0195] If, after optimization, it is found that for certain dense airspaces, even with an optimized altitude layer allocation scheme, unavoidable conflicts still exist (i.e., the optimal value of target 1 is greater than 0), the system will highlight these conflicting flight pairs and their conflict periods, and issue alarm information. Controllers can then, based on the alarm information and their own experience, manually adjust the speed or heading of the relevant flights, or activate other more complex conflict resolution algorithms to handle the situation.

[0196] This invention provides an approach and method for generating civil aviation air traffic management decisions based on a collaborative architecture of large and small models. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for generating civil aviation air traffic management decisions based on a collaborative architecture of large and small models, characterized in that, Includes the following steps: Step 1: Obtain historical scene dataset, which includes at least historical weather radar echo image sequences and historical altitude layer allocation strategies within the corresponding time segments; extract semantic features from historical weather radar echo image sequences using a general large model, and associate and store the resulting historical semantic feature vectors with the corresponding historical altitude layer allocation strategies. Step 2: Obtain the meteorological information to be decided, process the meteorological information to be decided through a large model, and generate the current semantic feature vector; Step 3: Perform similarity matching between the current semantic feature vector obtained in Step 2 and the historical semantic feature vector, and use the historical altitude layer allocation strategy associated with the matching result as the pre-tactical altitude layer allocation scheme; the scheme includes restrictions on the use of altitude layers in specified airspaces, routes or waypoints; Step 4: Obtain flight information, airspace structure information, and the pre-tactical altitude layer allocation scheme obtained in Step 3 within the sector to be planned, and convert the pre-tactical altitude layer allocation scheme into altitude constraints on flights at sector boundary points. Step 5: Construct a mathematical optimization model. The mathematical optimization model takes minimizing flight conflicts within the sector or minimizing flight conflicts while maximizing flight operation economy as the objective function, and takes flight performance, safety interval and the altitude constraint mentioned in Step 4 as the constraint conditions to model the altitude profile of the flight set within the sector. Step 6: Solve the mathematical optimization model constructed in Step 5 using an optimization solver to generate a tactical-level altitude layer allocation scheme. The tactical-level altitude layer allocation scheme includes the specific flight altitude layers assigned to each flight in the sector at different time periods.

2. The method according to claim 1, characterized in that, Step 1 specifically includes: Step 1-1: For each frame of the historical weather radar echo image sequence, use the Visual Language Model (VLM) to generate the corresponding natural language description text based on preset prompt words. Steps 1-2: Concatenate all natural language description texts of a scene consisting of consecutive time frames in chronological order, construct prompt words, and input them into a general large language model (LLM) to generate a summary text describing the macro-meteorological evolution process of the scene. Steps 1-3: Input the summarized text into the text embedding model to generate a fixed-dimensional historical semantic feature vector; Steps 1-4: Establish a mapping relationship between the historical semantic feature vector and the historical height layer allocation strategy within the time segment, and store it in the database.

3. The method according to claim 1, characterized in that, Step 2 specifically includes: Step 2-1: Obtain the meteorological information to be decided, which includes natural language text description and / or summary text generated after the meteorological radar echo image sequence to be decided is processed by the visual language big model and the big language model; Step 2-2: Vectorize the natural language text description or summary text using a text embedding model to generate the current semantic feature vector.

4. The method according to claim 1, characterized in that, Step 3 specifically includes: Step 3-1: Calculate the similarity between the current semantic feature vector and the semantic feature vector of the historical scene. The similarity calculation uses cosine similarity, and the formula is as follows: in, This is the current semantic feature vector. For the first Semantic feature vectors of historical scenes, Represents the vector dot product. The L2 norm of a vector; Step 3-2: Extract similarity The historical height layer allocation strategy associated with the highest historical scene As an output of the pre-tactical altitude layer allocation scheme, the pre-tactical altitude layer allocation scheme includes restrictions on the use of altitude layers for designated airspace, routes, or waypoints.

5. The method according to claim 1, characterized in that, In step 4, the height constraint on the flight at the sector boundary point specifically means: limiting the entry and exit heights of the flight to be assigned a height layer to the height layer usage restrictions obtained in step 3 from the historical height layer allocation strategy.

6. The method according to claim 1, characterized in that, In step 5, the mathematical optimization model is a mixed integer linear programming model; the flight set includes a planning pool of flights and a frozen pool of flights; wherein, the planning pool of flights are flights to be assigned altitude layers, and the frozen pool of flights are flights whose altitude layers have been determined or cannot be changed; The objective function of the model is to minimize the duration of flight conflicts between flights in the planning pool and between flights in the planning pool and flights in the frozen pool. in, For all flight pairs with potential level conflicts, For flights A set of time steps with potential horizontal conflicts. For 0-1 decision variables, a value of 1 indicates that the flight is within the specified time step. The safe distance was violated.

7. The method according to claim 6, characterized in that, The mathematical optimization model also includes an objective function that maximizes the proportion of flight time spent at high altitudes: in, To plan the pool of flights, For flights The set of time steps in the planning time domain. For height profile set, For flights At time step , No. The height value on each height profile.

8. The method according to claim 7, characterized in that, When there are multiple objective functions, a hierarchical sequence method is used to solve them: first, solve the objective of minimizing flight conflicts, obtain the optimal value and add it as a constraint to the model, and then solve the objective of maximizing the proportion of high-altitude flight time.

9. The method according to claim 8, characterized in that, The constraints further include: Flight performance constraints: Limit the magnitude, rate, and time interval of each altitude adjustment for a flight; Altitude layer feasibility constraint: Restricts flights to one of the specified available altitude layers during non-adjustment phases; Logical constraints on decision variables: ensuring the continuity and rationality of the generated height profile over time; And, the boundary height constraints derived from the pre-tactical height layer allocation scheme.

10. The method according to claim 1, characterized in that, The method further includes: Step 7: For potential conflicts that the tactical-level altitude layer allocation scheme cannot resolve, output alarm information and transfer the case to the controller for manual handling or call other conflict resolution algorithms for intervention.

Images