Air traffic autonomous operation track generation method and device based on generative AI

Abstract

The method and device for generating autonomous operation flight path of air traffic based on generative AI can quickly generate the operation flight path in the autonomous operation scene of air traffic, and provide sufficient data resources for the research and evaluation of autonomous operation of air traffic. The method comprises the following steps: (1) defining a typical autonomous operation scene, defining six autonomous operation scenes according to the autonomous operation demand analysis, and describing and defining the characteristics of each scene; (2) constructing a generative AI model, training the model according to the historical data provided by the first-level data center, and saving the trained model data; training the model according to the historical data provided by the first-level data center, and saving the trained model data; (3) generating the flight path data of each operation scene by using the generative AI model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of civil aviation general aviation flight, and particularly relates to an air traffic autonomous operation flight path generation method based on generative AI, and also relates to an air traffic autonomous operation flight path generation device based on generative AI. BACKGROUND

[0002] Air traffic autonomous operation refers to that in the future air traffic system, an aircraft can autonomously perform a navigation task without relying on a centralized ground control system. The air traffic autonomous operation mode introduces and utilizes new technologies such as big data, cloud computing, the Internet of Things, artificial intelligence and blockchains, etc., to realize intelligent upgrading and transformation of the aviation industry, and focuses on promoting information sharing and data exchange to support a more collaborative and integrated aviation ecosystem.

[0003] However, the construction of ground-air integrated autonomous operation is still in the period of conceptual research and key technology research, and therefore lacks actual operation data in various autonomous operation scenarios, resulting in a lack of effective evaluation methods for the efficiency of ground-air integrated autonomous operation, and it is difficult to effectively evaluate the correctness of the direction of further research and efficiency improvement after autonomous operation. SUMMARY

[0004] To overcome the defects of the prior art, the technical problem to be solved by the present application is to provide an air traffic autonomous operation flight path generation method based on generative AI, which can quickly generate operation flight paths in air traffic autonomous operation scenarios and provide sufficient data resources for the research and evaluation of air traffic autonomous operation.

[0005] The technical solution of the present application is that the air traffic autonomous operation flight path generation method based on generative AI includes the following steps:

[0006] (1) defining typical autonomous operation scenarios, defining six autonomous operation scenarios according to autonomous operation demand analysis, and describing and defining the characteristics of each scenario;

[0007] (2) constructing a generative AI model, training according to the historical data provided by the first-level data center, and saving the trained model data; training according to the historical data provided by the first-level data center, and saving the trained model data;

[0008] (3) generating flight path data for each operation scenario using the generative AI model;

[0009] The six autonomous operation scenarios described and defined in step (1) are as follows:

[0010] (1.1) Approach procedure autonomous decision operation scene, under the autonomous operation concept, the aircraft realizes the autonomous judgment of the continuous descent condition of the on-board end, flight profile planning and continuous descent operation after air-ground negotiation in the approach procedure;

[0011] (1.2) Straighten running scene under airspace flexible use, under the autonomous operation scene, the active selection of the temporary flight line of the crew and the straighten flight are realized;

[0012] (1.3) Autonomous diversion operation scene under severe weather, under the autonomous diversion operation scene under severe weather, the aircraft determines the best route to bypass the weather according to its business target and risk tolerance, avoids the congested route, and implements after automatic negotiation and confirmation with the ground automation system;

[0013] (1.4) Single-aircraft TBO required arrival time operation scene under safe conditions;

[0014] (1.5) Aircraft emergency disposal scene under the influence of en route strong convection, in the single-aircraft TBO required arrival time operation under safe conditions, the aircraft receives the critical point control arrival time (CTA) requirement of the ground digital control, combines the surrounding situation service provided by the ground, and autonomously controls the aircraft to accelerate and decelerate under the premise of ensuring safety with the previous aircraft and the surrounding aircraft, realizes the required arrival time (RTA) function of TBO, and flies according to the predetermined flight path;

[0015] (1.6) Autonomous wake separation operation scene under high-density scene, if the aircraft encounters severe strong convection weather and cannot avoid or bypass, under the autonomous operation scene, the aircraft makes independent decisions, decides whether to make an emergency descent according to the situation, and through air-ground information sharing and collaborative decision-making, the pilot perceives the surrounding airspace operation situation, the flight intention of other flights, the use of each height layer, the traffic situation, and the potential conflict situation, and the pilot autonomously adjusts the change scheme.

[0016] The present application defines typical autonomous operation scenes, analyzes the autonomous operation requirements, defines six autonomous operation scenes, and describes and defines the characteristics of each scene; a generative AI model is constructed, historical data provided by the first-level data center is used for training, and the trained model data is saved; historical data provided by the first-level data center is used for training, and the trained model data is saved; the generative AI model is used to generate flight path data for each operation scene; therefore, the operation flight path in the autonomous operation scene of air traffic can be quickly generated, and sufficient data resources are provided for the research and evaluation of autonomous operation of air traffic.

[0017] An air traffic autonomous operation flight path generation device based on generative AI is also provided, which comprises:

[0018] A definition scenario module is configured to define typical autonomous operation scenarios, six autonomous operation scenarios are defined according to autonomous operation demand analysis, and the characteristics of each scenario are described and defined;

[0019] A model construction module is configured to construct a generative AI model, train the generative AI model according to historical data provided by the first data center, and save the trained model data; train the generative AI model according to historical data provided by the first data center, and save the trained model data;

[0020] A data generation module is configured to use the generative AI model to generate flight path data for each operation scenario;

[0021] The six autonomous operation scenarios of the definition scenario module are described and defined as follows:

[0022] (1.1) Approach procedure autonomous decision operation scenario, under the concept of autonomous operation, the aircraft realizes continuous descent condition autonomous judgment, flight profile planning, and continuous descent operation after air-ground negotiation in the approach procedure;

[0023] (1.2) Straightening operation scenario under flexible use of airspace, under the autonomous operation scenario, the crew actively selects and straightens the flight route;

[0024] (1.3) Autonomous diversion operation scenario in severe weather, under the autonomous diversion operation scenario in severe weather, the aircraft determines the best route to bypass the weather according to its business target and risk tolerance, avoids congested routes, and implements after automatic negotiation and confirmation with the ground automation system;

[0025] (1.4) Single-aircraft TBO required arrival time operation scenario under safe conditions;

[0026] (1.5) Aircraft emergency disposal scenario under the influence of en route strong convection, in the single-aircraft TBO required arrival time operation under safe conditions, the aircraft receives the critical point control arrival time (CTA) requirement of the ground digital control, combines the surrounding situation service provided by the ground, and under the premise of ensuring safety with the previous aircraft and surrounding aircraft, autonomously controls the aircraft to accelerate and decelerate, realizes the required arrival time (RTA) function of TBO, and flies according to the predetermined flight path;

[0027] (1.6) Autonomous wake separation operation scenario under high-density scenarios, if the aircraft encounters severe strong convection weather and cannot avoid or bypass, under the autonomous operation scenario, the aircraft independently makes decisions, decides whether to make an emergency descent according to the situation, and through air-ground information sharing and collaborative decision-making, the pilot perceives the surrounding airspace operation situation, other flight intentions, height layer usage, traffic conditions, and potential conflict situations, and the pilot autonomously adjusts and changes the plan. Attached Figure Description

[0028] Figure 1 The diagram shown is an overall flowchart of the autonomous air traffic trajectory generation method based on generative AI according to the present invention.

[0029] Figure 2 The diagram shows a flowchart of the steps (2.1.2.1) of the autonomous air traffic trajectory generation method based on generative AI according to the present invention.

[0030] Figure 3 The diagram shows a flowchart of steps (2.1.2.2) of the autonomous air traffic trajectory generation method based on generative AI according to the present invention.

[0031] Figure 4 The diagram shows a flowchart of step (3) of the autonomous air traffic trajectory generation method based on generative AI according to the present invention.

[0032] Figure 5 The following is a flowchart of steps (2.1.2.2.2.3.2) and (2.1.2.2.2.1.2) of the autonomous air traffic operation trajectory generation method based on generative AI according to the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0034] To make the description of this disclosure more detailed and complete, illustrative descriptions of embodiments and specific examples of the present invention are provided below; however, these are not the only forms of implementing or utilizing the specific examples of the present invention. The embodiments cover features of multiple specific examples and methods and steps for constructing and operating these specific examples, and their order. However, other specific examples may also be used to achieve the same or equivalent functions and order of steps.

[0035] like Figure 1 As shown, this method for generating autonomous air traffic routes based on generative AI includes the following steps:

[0036] (1) Define typical autonomous operation scenarios. Based on the analysis of autonomous operation requirements, define six autonomous operation scenarios and describe and define the characteristics of each scenario.

[0037] (2) Construct a generative AI model, train it based on the historical data provided by the first-level data center, and save the trained model data; train it based on the historical data provided by the first-level data center, and save the trained model data;

[0038] (3) Use the generative AI model to generate track data for each operating scenario;

[0039] Among them, the description and definition of the six autonomous operating scenarios in step (1) are as follows:

[0040] (1.1) Approach procedure autonomous decision-making operating scenario. Under the concept of autonomous operation, the aircraft realizes continuous descent condition autonomous judgment, flight profile planning, and continuous descent operation after air-ground negotiation in the approach procedure;

[0041] (1.2) Straightening operation scenario under flexible use of airspace. Under the autonomous operation scenario, the crew actively selects the temporary flight route and performs the straightening flight;

[0042] (1.3) Autonomous rerouting operating scenario under adverse weather. Under the autonomous rerouting operating scenario under adverse weather, the aircraft determines the best route to bypass the weather according to its business objectives and risk tolerance, avoids congested routes, and implements after automatic negotiation and confirmation with the ground automation system;

[0043] (1.4) Single-aircraft TBO required arrival time operating scenario under safe conditions;

[0044] (1.5) Aircraft emergency disposal scenario under the influence of en route strong convection. In the single-aircraft TBO required arrival time operating scenario under safe conditions, the aircraft receives the critical point control arrival time (CTA) requirement from the ground digital control, combines the surrounding situation service provided by the ground, and independently controls the aircraft acceleration and deceleration under the premise of ensuring safety with the previous aircraft and surrounding aircraft, realizes the required arrival time (RTA) function of TBO, and flies according to the predetermined track;

[0045] (1.6) Autonomous wake separation operating scenario under high-density scenarios. If the aircraft encounters severe strong convection weather and cannot avoid or bypass, the aircraft makes independent decisions under the autonomous operating scenario, decides whether to perform an emergency descent according to the situation, and through air-ground information sharing and collaborative decision-making, the pilot perceives the surrounding airspace operating situation, other flight intentions, height layer usage, traffic conditions, and potential conflict situations, and the pilot autonomously adjusts and changes the plan.

[0046] The application defines typical autonomous operation scenarios, defines six autonomous operation scenarios according to autonomous operation demand analysis, and describes and defines the characteristics of each scenario; a generative AI model is constructed, which is trained according to the historical data provided by the first-level data center, and the trained model data is saved; the historical data provided by the first-level data center is trained, and the trained model data is saved; the generative AI model is used to generate the flight path data of each operation scenario; therefore, the operation flight path in the autonomous operation scenario of air traffic can be quickly generated, and sufficient data resources are provided for the research and evaluation of autonomous operation of air traffic.

[0047] Preferably, the step (2) comprises the following sub-steps:

[0048] (2.1) training the model according to historical data, updating the iterative model parameters;

[0049] (2.2) saving the best model for subsequent use;

[0050] (2.3) preserving the scalability of the model, and iteratively upgrading the model as the historical data becomes richer and accumulates.

[0051] Preferably, the step (2.1) comprises the following sub-steps:

[0052] (2.1.1) aircraft flight trajectory establishment;

[0053] (2.1.2) flight trajectory diffusion model construction;

[0054] (2.1.3) operation scene feature extraction, using VAE model to map in latent variable space to learnable low-dimensional variables;

[0055] (2.1.4) training according to the historical data provided by the first-level data center, and saving the trained model data;

[0056] The step (2.1.1) comprises the following sub-steps:

[0057] (2.1.1.1) the historical data used in this patent are all ADS-B data, before use, the cat21 original data needs to be parsed into decimal data for subsequent use;

[0058] (2.1.1.2) according to date, 24-bit address code and flight tail number to determine a unique aircraft flight path;

[0059] (2.1.1.3) using TRP time to sort ADS-B messages;

[0060] (2.1.1.4) Select the time, longitude, latitude, speed, height, and north heading angle in the ADS-B message as the six characteristic information related to the track;

[0061] (2.1.1.5) Use filtering method to remove abnormal tracks;

[0062] The step (2.1.2) comprises the following sub-steps:

[0063] (2.1.2.1) Define the forward process of the diffusion model to obtain the noise level at each step;

[0064] (2.1.2.2) Define the AirTraj-UNet network structure to learn the track distribution;

[0065] (2.1.2.3) Train the AirTraj-UNet network structure to output the noise level at each step, and optimize the parameters by techniques such as back propagation and gradient descent according to the difference between the predicted noise level and the actual noise level, and adjust the model parameters;

[0066] The step (2.1.3) comprises the following sub-steps:

[0067] (2.1.3.1) Construct a VAE model, including an encoding module and a decoding module;

[0068] (2.1.3.2) Train and optimize the VAE model parameters using historical data;

[0069] (2.1.3.3) Save the best model parameters;

[0070] (2.1.3.4) Use the encoding part of the VAE model to encode the scene features, and the encoded embedding information is used in each attention module in the AirTraj-UNet structure to learn the correlation with the track information through cross-attention layers.

[0071] Preferably, as shown in Figure 2 The step (2.1.2.1) comprises the following sub-steps:

[0072] (2.1.2.1.1) Obtain the original data x0 from the training set;

[0073] (2.1.2.1.2) Gradually add noise to the data in a series of time steps t, for

[0074] At each time step t, Gaussian noise is added by the following formula:

[0075]

[0076] where αt is a predetermined decay coefficient of time step t, z t is a random variable sampled from a standard normal distribution

[0077] machine noise vector;

[0078] (2.1.2.1.3) repeat step (2.1.2.1.2) until reaching the final time step T, at which time the data x t is close to pure noise.

[0079] Preferably, as Figure 3 shown, the step (2.1.2.2) comprises the following sub-steps:

[0080] (2.1.2.2.1) inverse process starts from a data x t close to pure noise, gradually reduces the influence of noise in a series of time steps t, each time step uses a parameterized model θ to predict the noise component in the earlier state x t-1 or directly predict the noise-free state, each step involves calculating a denoised version

[0081] of the data x is achieved by the following formula:

[0082]

[0083] wherein, is the estimate of noise at time step t, α t is a noise adjustment parameter;

[0084] (2.1.2.2.2) construct the AirTraj-UNet network structure, learn the model parameters θ;

[0085] (2.1.2.2.3) after completing the last step t = 1, output which is the estimated value of the original data, as the generated trajectory.

[0086] Preferably, the step (2.1.2.2.2) comprises the following sub-steps:

[0087] (2.1.2.2.2.1) construct a down-sampling module, the length of the trajectory becomes 1 / 2 of the original length in the process of passing through the down-sampling module each time, so that the model learns the distribution of the trajectory in multiple scales;

[0088] (2.1.2.2.2.2) construct a correlation learning module containing a self-correlation attention mechanism and a cross-attention mechanism;

[0089] (2.1.2.2.2.3) constructing an up-sampling module, the length of the track becomes twice the original length in each process of passing through the up-sampling module, so that the track is finally restored to the original length.

[0090] Preferably, the step (2.1.2.2.2.3) comprises the following sub-steps:

[0091] (2.1.2.2.2.3.1) constructing a time encoding module comprising a sinusoidal mapping layer, a linear mapping layer, a nonlinear activation layer, and finally connecting a linear mapping layer again;

[0092] (2.1.2.2.2.3.2) constructing a correlation learning module comprising a batch normalization, a nonlinear activation layer, an expanded convolution

[0093] layer (as shown in FIG. Figure 5 ) of a residual connection network module;

[0094] (2.1.2.2.2.3.3) constructing a correlation learning module comprising a self-correlation attention mechanism and a cross-attention mechanism, using the self-correlation attention mechanism to learn the correlation between the data of the track itself, and using the cross-attention mechanism to learn the correlation between the track distribution and the scene features;

[0095] (2.1.2.2.2.3.4) adding the results of the time encoding to the residual connection network module after linear mapping, and then stacking multiple times in the order of the residual connection module and the correlation learning module.

[0096] Preferably, the step (2.1.2.2.2.1) comprises the following sub-steps:

[0097] (2.1.2.2.2.1.1) constructing a time encoding module comprising a sinusoidal mapping layer, a linear mapping layer, a nonlinear activation layer, and finally connecting a linear mapping layer again;

[0098] (2.1.2.2.2.1.2) constructing a correlation learning module comprising a batch normalization, a nonlinear activation layer, an expanded convolution

[0099] layer (as shown in FIG. Figure 5 ) of a residual connection network module;

[0100] (2.1.2.2.2.1.3) constructing a correlation learning module comprising a self-correlation attention mechanism and a cross-attention mechanism, using the self-correlation attention mechanism to learn the correlation between the data of the track itself, and using the cross-attention mechanism to learn the correlation between the track distribution and the scene features;

[0101] (2.1.2.2.2.1.4) The result of time coding is linearly mapped and added to the residual connection network module, and then multiple superpositions are performed in the order of residual connection module, correlation learning module.

[0102] Preferably, as Figure 4 shown, the step (3) comprises the following sub-steps:

[0103] (3.1) extracting the running scene features using the VAE encoding part;

[0104] (3.2) matching the model with the scene;

[0105] (3.3) generating the autonomous running track of air traffic according to the matching result.

[0106] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer readable storage medium. When the program is executed, it includes the steps of the above-mentioned embodiment method, and the storage medium can be ROM / RAM, a magnetic disc, an optical disc, a memory card, etc. Therefore, corresponding to the method of the present application, the present application also simultaneously includes an autonomous running track generation device for air traffic based on generative AI. The device is usually represented in the form of a functional module corresponding to each step of the method. The device comprises:

[0107] The scene defining module is configured to define typical autonomous running scenes. Six autonomous running scenes are defined according to autonomous running demand analysis, and the features of each scene are described and defined;

[0108] The model building module is configured to build a generative AI model. The model is trained according to the historical data provided by the first data center, and the trained model data is saved. The model is trained according to the historical data provided by the first data center, and the trained model data is saved;

[0109] The data generation module is configured to use the generative AI model to generate track data for each running scene;

[0110] The six autonomous running scene descriptions and definitions of the scene defining module are as follows:

[0111] (1.1) Approach procedure autonomous decision running scene. Under the concept of autonomous running, the aircraft realizes autonomous judgment of continuous descent conditions, flight profile planning, and continuous descent operation after air-ground negotiation in the approach procedure at the on-board end.

[0112] (1.2) Under the scenario of flexible use of airspace, under the scenario of autonomous operation, the temporary route of the unit is actively selected and the flight is taken by cutting off the straight line;

[0113] (1.3) Under the scenario of autonomous rerouting in bad weather, under the scenario of autonomous rerouting in bad weather, the aircraft determines the best route to bypass the weather according to its business objectives and risk tolerance, and avoids congested routes, and automatically negotiates and confirms with the ground automation system before implementation;

[0114] (1.4) Single-aircraft TBO required arrival time operation scenario under safe conditions;

[0115] (1.5) Aircraft emergency disposal scenario under the influence of strong convection on the air route, in the single-aircraft TBO required arrival time operation under safe conditions, the aircraft receives the critical point control arrival time (CTA) requirement of the ground digital control, combined with the surrounding situation service provided by the ground, under the premise of ensuring the safety of the aircraft and the surrounding aircraft, the aircraft is independently controlled to accelerate and decelerate, to realize the required arrival time (RTA) function of TBO, and fly according to the predetermined flight path;

[0116] (1.6) Autonomous wake separation operation scenario under high-density scenario, if the aircraft encounters severe strong convection weather and cannot avoid or bypass, under the scenario of autonomous operation, the aircraft makes independent decisions, decides whether to make an emergency descent according to the situation, through air-ground information sharing and collaborative decision-making, the pilot perceives the surrounding airspace operation situation, the flight intention of other flights, the use of each height layer, the flow situation, and the potential conflict situation, and the pilot adjusts and changes the scheme independently.

[0117] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Any simple modification, equivalent change and modification of the above embodiment according to the technical essence of the present application is still within the protection scope of the technical solution of the present application.

Claims

1. An autonomous flight path generation method for air traffic based on generative AI, characterized in that it comprises the following steps: (1) defining typical autonomous operation scenarios, analyzing the autonomous operation requirements, defining six autonomous operation scenarios, and describing and defining the characteristics of each scenario; (2) constructing a generative AI model, training the model according to the historical data provided by the first-level data center, and saving the trained model data; training the model according to the historical data provided by the first-level data center, and saving the trained model data; (3) generating flight path data for each operation scenario using the generative AI model; wherein step (1) six autonomous operation scenario descriptions and definitions are as follows: (1.1) approach procedure autonomous decision operation scenario, under the concept of autonomous operation, the aircraft realizes continuous descent condition autonomous judgment, flight profile planning, and continuous descent operation after air-ground negotiation in the approach procedure; (1.2) straightening operation scenario under flexible use of airspace, under the autonomous operation scenario, the flight crew actively selects and straightens the route; (1.3) autonomous diversion operation scenario in adverse weather, in the autonomous diversion operation scenario in adverse weather, the aircraft determines the best route to bypass the weather according to its business objectives and risk tolerance, and avoids congested routes, and automatically negotiates and confirms with the ground automation system before implementation; (1.4) single-aircraft TBO required arrival time operation scenario under safe conditions; (1.5) aircraft emergency disposal scenario under the influence of en route strong convection, in the single-aircraft TBO required arrival time operation scenario under safe conditions, the aircraft receives the key point control arrival time CTA requirement of the ground digital control, combines the surrounding situation service provided by the ground, and under the premise of ensuring safety with the previous aircraft and surrounding aircraft, autonomously controls the aircraft to accelerate and decelerate to realize the required arrival time RTA function of TBO, and flies according to the predetermined flight path; (1.6) autonomous wake separation operation scenario in high-density scenarios, if the aircraft encounters severe strong convection weather and cannot avoid or bypass, the aircraft makes independent decisions under the autonomous operation scenario, decides whether to make an emergency descent according to the situation, and through air-ground information sharing and collaborative decision-making, the pilot perceives the surrounding airspace operation situation, the flight intention of other aircraft, the use of each altitude layer, the traffic situation, and the potential conflict situation, and the pilot autonomously adjusts and changes the scheme; The step (2) comprises the following sub-steps: (2.1) training the model according to the historical data, updating and iterating the model parameters; (2.2) saving the best model for subsequent use; (2.3) preserving the scalability of the model, and iteratively upgrading the model as the historical data becomes richer and accumulates; The step (2.1) comprises the following sub-steps: (2.1.1) aircraft flight trajectory establishment; (2.1.2) flight trajectory diffusion model construction; (2.1.3) operation scenario feature extraction, using the VAE model to map the latent variable space to a learnable low-dimensional variable; ​ (2.1.4) According to the historical data provided by the primary data center, training is carried out, and the trained model data is saved; The step (2.1.1) comprises the following sub-steps: (2.1.1.1) The historical data used in the patent are all ADS-B data, and before use, the original data of cat21 needs to be parsed into decimal data for subsequent use; (2.1.1.2) The unique aircraft track is determined according to the date, 24-bit address code and flight tail number; (2.1.1.3) The ADS-B message is sorted using the TRP time; (2.1.1.4) Select the time, longitude, latitude, speed, height and north heading angle in the ADS-B message, which are six characteristic information related to the track; (2.1.1.5) Use filtering method to remove abnormal tracks; The step (2.1.2) comprises the following sub-steps: (2.1.2.1) Define the forward process of the diffusion model to obtain the noise level of each step; (2.1.2.2) Define the AirTraj-UNet network structure to learn the track distribution; (2.1.2.3) Train the AirTraj-UNet network structure to output the noise level of each step, and optimize the parameters by techniques such as back propagation and gradient descent according to the difference between the predicted noise level and the actual noise level, and adjust the model parameters; The step (2.1.3) comprises the following sub-steps: (2.1.3.1) Construct a VAE model, including an encoding module and a decoding module; (2.1.3.2) Train and optimize the VAE model parameters using historical data; (2.1.3.3) Save the best model parameters; (2.1.3.4) Use the encoding part of the VAE model to encode the scene features, and the embedded information after encoding is in each attention module in the AirTraj-UNet structure, which learns the correlation with the track information through cross-attention layers.

2. The method of claim 1, wherein the method is based on a generative AI autonomous operation trajectory generation method for air traffic. The step (2.1.2.1) comprises the following sub-steps: (2.1.2.1.1) Obtain the original data x0 from the training set; (2.1.2.1.2) Add noise to the data gradually in a series of time steps t, and for each time step t, add Gaussian noise through the following formula: where a t is a predetermined decay coefficient for time step t, z t is a random noise vector sampled from a standard normal distribution; (2.1.2.1.3) repeat step (2.1.2.1.2) until the final time step T is reached, at which time the data x t Approaching pure noise.

3. The method for generating autonomous air traffic operation tracks based on generative AI according to claim 2, characterized in that: The step (2.1.2.2) comprises the following sub-steps: (2.1.2.2.1) The inverse process predicts the noise component in x t Initially, the impact of noise is reduced step by step over a sequence of time steps t, each using a parametric model θ to predict an earlier state x t-1 either the noise component in x or directly the noise-free state, each step involving the computation of a denoised version This is achieved by the following equation: wherein, is an estimate of the noise at time step t, a t is a noise adjustment parameter; (2.1.2.2.2) Construct the AirTraj-UNet network structure and learn the model parameters θ; (2.1.2.2.3) After the last step t = 1 is completed, output This is the estimate of the original data as a generated track.

4. The method of claim 3, wherein the method is based on a generative AI autonomous operation trajectory generation for air traffic. The step (2.1.2.2.2) comprises the following sub-steps: (2.1.2.2.2.1) Build a downsampling module, which makes the length of the track become 1 / 2 of the original length after each downsampling module, so that the model learns the track distribution in multiple scales; (2.1.2.2.2.2) Build a correlation learning module containing a self-correlation attention mechanism and a cross-attention mechanism; (2.1.2.2.2.3) Build an upsampling module, which makes the length of the track become twice the original length after each upsampling module, so that the track finally restores to the original length.

5. The method for generating autonomous air traffic operation tracks based on generative AI according to claim 4, characterized in that: The step (2.1.2.2.2.3) comprises the following sub-steps: (2.1.2.2.2.3.1) constructing a time encoding module comprising a sinusoidal mapping layer, a linear mapping layer, a nonlinear activation layer, and finally connecting a linear mapping layer again; (2.1.2.2.2.3.2) constructing a residual connection network module comprising batch normalization, a nonlinear activation layer, and an expanded convolution layer; (2.1.2.2.2.3.3) constructing a correlation learning module comprising a self-correlation attention mechanism and a cross-attention mechanism, using the self-correlation attention mechanism to learn the correlation between the data of the track itself, and using the cross-attention mechanism to learn the correlation between the track distribution and the scene features; (2.1.2.2.2.3.4) adding the results of time encoding to the residual connection network module after linear mapping, and then stacking multiple times in the order of residual connection module and correlation learning module.

6. The method of claim 3, wherein the method further comprises: The step (2.1.2.2.2.1) comprises the following sub-steps: (2.1.2.2.2.1.1) constructing a time encoding module comprising a sinusoidal mapping layer, a linear mapping layer, a nonlinear activation layer, and finally connecting a linear mapping layer again; (2.1.2.2.2.1.2) constructing a residual connection network module comprising batch normalization, a nonlinear activation layer, and an expanded convolution layer; (2.1.2.2.2.1.3) constructing a correlation learning module comprising a self-correlation attention mechanism and a cross-attention mechanism, using the self-correlation attention mechanism to learn the correlation between the data of the track itself, and using the cross-attention mechanism to learn the correlation between the track distribution and the scene features; (2.1.2.2.2.1.4) adding the results of time encoding to the residual connection network module after linear mapping, and then stacking multiple times in the order of residual connection module and correlation learning module.

7. The method of claim 3, wherein the method further comprises: determining a flight path for the aircraft based on the generated AI model. The step (3) comprises the following sub-steps: (3.1) extracting the running scene features using the VAE encoding part; (3.2) matching the model with the scene; (3.3) generating the autonomous running track of air traffic according to the matching result.

Images