Virtual low-altitude air traffic controller AI model training system

Abstract

The invention, which relates to the technical field of aviation training and simulation, discloses a virtual low-altitude air traffic controller AI model training system comprising a data acquisition and preprocessing module, an AI core model module and a communication and interaction module. The data acquisition and preprocessing module is used for collecting data from a simulated low-altitude radar, low-altitude weather, low-altitude geographic information and a low-altitude flight plan data source, and performing cleaning, denoising, format conversion and integration processing on the data to generate preprocessed data used by the AI core model. According to the virtual low-altitude air traffic controller AI model training system provided by the invention, the effect and quality of low-altitude flight training of a pilot are remarkably improved, and the pilot can perform a large amount of practical operation and emergency treatment training in a safe environment by simulating various real low-altitude flight scenes and complex conditions; the coping ability of the system to low-altitude flight rules, control programs and various complex conditions is enhanced, the training period is shortened, and the training cost is reduced.

Description

Technical Field

[0001] The present invention relates to the field of aviation training and simulation technology, and specifically to a virtual low-altitude air traffic controller AI model training system. Background Art

[0002] With the deepening of the reform of low-altitude airspace management in China, the low-altitude economy industry has witnessed an explosive growth. The booming development of emerging business forms such as unmanned aerial vehicle (UAV) logistics distribution, aerial emergency rescue, commuter aviation, and low-altitude tourism has led to an exponential increase in the number of low-altitude aircraft. According to the statistical data of the Civil Aviation Administration, in 2024, the number of registered general aviation aircraft in the country exceeded 12,000, the number of UAV operating enterprises exceeded 8,000, the average daily flight operations in the low-altitude airspace exceeded 500,000 times, and the utilization efficiency of airspace resources increased by 300% compared with five years ago. This explosive growth has put forward higher requirements for the cultivation of low-altitude flight professionals and low-altitude traffic management technology.

[0003] The existing low-altitude flight training system mainly relies on physical simulation cabins and the mode of instructor-guided flight. The training cost per flight operation is relatively high, and the training coverage rate of key subjects such as complex meteorological conditions and special situation handling needs to be improved. The existing simulation equipment is limited by the scale of the preset scenario library, and there are certain challenges in dynamically generating complex three-dimensional environments such as urban canyons and high-voltage wire networks and simulating new scenarios such as swarm operations of group UAVs. In terms of the verification of traffic management systems, the simulation platform based on mathematical modeling has problems such as low scene restoration degree and lack of human-computer interaction, and there is a certain deviation between the generated test data and the real operating environment, resulting in a relatively long system verification cycle. In addition, the existing technical system lacks a closed-loop training mechanism for AI models, making it difficult to optimize control strategies in real time through deep learning and unable to provide intelligent virtual instructor support for pilots. In response to this, we have proposed a virtual low-altitude air traffic controller AI model training system. Summary of the Invention

[0004] To solve the above technical problems, a virtual low-altitude air traffic controller AI model training system is provided, and this technical solution solves the above problems.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A virtual low-altitude air traffic controller AI model training system, comprising: a data acquisition and preprocessing module, an AI core model module, and a communication and interaction module; The data acquisition and preprocessing module is used to collect data from simulated low-altitude radar, low-altitude meteorology, low-altitude geographic information, and low-altitude flight plan data sources, and perform cleaning, denoising, format conversion, and integration processing on the data to generate preprocessed data for use by the AI core model; The AI core model module is electrically connected to the data acquisition and preprocessing module. The AI core model module includes: a low-altitude control instruction generation model, a low-altitude traffic situation perception and prediction model, and a simulation scenario generation and management model. The low-altitude control instruction generation model is constructed based on the natural language processing technology of deep learning. After learning a large number of low-altitude flight control instruction samples and relevant regulations and specifications texts, it can generate control instructions that meet safety requirements according to the low-altitude flight situation. The low-altitude traffic situation perception and prediction model adopts a structure that combines a graph neural network and a recurrent neural network. By modeling the low-altitude traffic network and learning the characteristics of traffic flow time series, it can real-time perceive and predict the flight trajectories and potential conflict risks of low-altitude aircraft. The simulation scenario generation and management model generates various low-altitude flight simulation scenarios according to training requirements, and monitors and adjusts them in real time; The communication and interaction module is electrically connected to the AI core model module. The communication and interaction module is used to establish a two-way communication link with the pilot training simulation device, as well as a data interaction interface with the low-altitude traffic management system simulation platform. It also includes a user interface for instructors to operate. The user interface can display the low-altitude flight situation map and relevant data, and realize the control and intervention of the simulation training scenario and the behavior of virtual low-altitude air traffic controllers.

[0006] Preferably, the data collected by the data acquisition and preprocessing module includes: The collected simulated low-altitude radar data covers the position, speed, altitude, and heading information of low-altitude aircraft with different models, flight trajectories, and speed changes; Among them, the low-altitude meteorological simulation data includes but is not limited to: wind speed, wind direction, air temperature, air pressure, visibility, cloud height, and precipitation probability; The low-altitude geographic information data includes but is not limited to: terrain and landform, obstacle distribution, airport and landing site locations; The low-altitude flight plan data includes but is not limited to: detailed information on general aviation flight plans, unmanned aircraft mission plans, and low-altitude tourism route plans; The preprocessing of the collected data specifically includes: Perform outlier correction on the radar simulation data, and use a sliding window filtering algorithm to smooth the noise. The filtering formula is: , Among them, is the original data at the th moment within the window, is the window length, is the output after filtering; Perform coordinate system normalization on the geographic information data, and convert data in different coordinate systems to the WGS84 coordinate system. The conversion formula is: , Among them, Indicates the coordinates in the converted WGS84 coordinate system, 、 is the coordinate value in the original coordinate system, 、 is the translation parameter, is the rotation angle; The coordinate system conversion function is calculated through the following steps: Rotate the original coordinates counterclockwise around the origin by the angle , and the calculation formula is: , where, is the spatial position of the point to be interpolated, is the spatial position of the th known meteorological observation point, is the meteorological parameter value of the th observation point, is the weight coefficient, satisfying .

[0007] Preferably, the low-altitude control instruction generation model is specifically: A natural language processing model based on the Transformer architecture, with the input being the flight situation vector and the flight plan text , and the output being the structured control instruction ; The parameters of the model are optimized through the cross-entropy loss function: , where, is the true instruction label, is the predicted probability; The training dataset of the low-altitude control instruction generation model contains the following annotation information: The initial state vector of the aircraft ; The control instruction sequence , and each instruction is annotated with the corresponding trigger condition and the execution result ; The abnormal scenario annotation, including the engine fault code and the corresponding emergency instruction set ; During training, the curriculum learning strategy is adopted, and training is carried out in stages according to the scenario complexity. The initial stage only includes the normal instructions of a single aircraft, and the final stage covers the multi-aircraft mixed conflict scenario. The loss function is weighted as: , where, For the reinforcement learning reward function, , ; The generated control instructions include: aircraft identification before takeoff, departure airport and destination airport, route information, transponder code, altitude clearance, departure frequency and departure time; reasonable altitude, speed and heading clearances during the low-altitude cruise phase; landing clearances that accurately reflect runway usage, meteorological conditions, and the distribution of surrounding obstacles during the landing phase, as well as the accurate transmission and confirmation of flight system messages.

[0008] Preferably, the low-altitude traffic situation awareness and prediction model is specifically: Adopting a hybrid structure of a graph neural network (GNN) and a long short-term memory network (LSTM), with the input being the aircraft node feature matrix and the adjacency matrix , and the output being the future and the set of avoidance paths ; The node update formula of the GNN is: , where is the hidden state of the -th layer node , is the activation function, is the weight matrix, is the concatenation operation, is the -th layer node 's hidden state, is the node 's all neighbors 's sum of hidden states at the -th layer, is the neighbor set of the node ; The conflict detection algorithm of the low-altitude traffic situation awareness and prediction model is: Construct the aircraft spatio-temporal trajectory matrix , with dimensions including longitude, latitude, altitude, and speed; calculate the Euclidean distance between two aircraft and at the future time: , In the formula, is the Euclidean distance between aircraft and at the future time, , , are the aircraft At the three-dimensional coordinates at the moment, , , are the three-dimensional coordinates of the aircraft at the moment; If , it is marked as a potential conflict, where represents the safety threshold.

[0009] Preferably, the simulation scenario generation and management model is specifically: Generate diverse flight scenarios based on the Generative Adversarial Network (GAN). The adversarial loss between the generator and the discriminator is: , In the formula, is the total loss function of the GAN, is the expectation operator, represents sampling a sample from the true data distribution , is the output probability of the discriminator for the true sample ; represents sampling noise from the input noise distribution of the generator , is the simulation sample generated by the generator according to the noise ; is the output probability of the discriminator for the generated sample ; The scenario diversity control method of the simulation scenario generation and management model includes: Define the scenario parameter space ; Use Latin hypercube sampling to generate the initial scenario set ; Evaluate the difference between the generated scenario and the true data distribution through KL divergence: , In the formula, is the KL divergence, is the probability distribution of the true data, is the probability distribution of the generated data, is the probability that the sample appears in the true data, is the probability that the sample appears in the generated data; If , adjust the dimension of the generator's latent space Sampling distribution

[0010] Preferably, the training method of the AI core model includes: Collect historical simulated flight records, standard flight scenarios, instructor demonstration operations, and abnormal emergency situation simulation data to construct a training data set, and perform fine annotation and preprocessing on the data, covering flight scenarios, control instructions, and flight result information; Adopt a multi-level training strategy combining supervised learning, reinforcement learning, and generative adversarial network technology. In the supervised learning stage, minimize the error between the prediction and the real instruction. In the reinforcement learning stage, optimize the decision-making strategy according to the flight safety and efficiency indicators. The generative adversarial network technology improves the simulation ability and instruction flexibility; Regularly evaluate and verify the model performance with an independent test data set. The indicators include the accuracy of control instruction generation, the accuracy of traffic conflict prediction, and the system response time. Adjust and optimize the model structure parameters, training data, and algorithms according to the results, and establish a version management mechanism to backtrack and select the optimal version.

[0011] Preferably, the data interaction sub-function of the communication and interaction module and the low-altitude traffic management system simulation platform realizes two-way data sharing and collaborative work, outputs the decision-making instructions of the virtual low-altitude air traffic controller and receives the platform feedback information, and the function integration sub-function embeds some system function modules into the simulation platform.

[0012] Preferably, the implementation method of the voice communication subsystem of the communication and interaction module is: Adopt a pre-emphasis filter Eliminate low-frequency noise and preprocess the voice signal, where represents the transfer function of the pre-emphasis filter, represents the complex variable of the Z-transform, represents the unit delay operation; Based on the preprocessed voice signal, segment the voice signal into frames, where the frame length is 25 ms and the frame shift is 10 ms; Apply the Hamming window function to each frame of signal to reduce spectral leakage, where the expression of the Hamming window function is: , In the formula, represents the value of the Hamming window at the th sampling point, is the DC component coefficient of the Hamming window, is the cosine component coefficient of the Hamming window, represents the parameter of the cosine function; Perform a fast Fourier transform on each windowed frame of signal to calculate the Mel spectrum coefficients , with a dimension of 40, where It is the dimension index of MFCC coefficients to complete MFCC feature extraction; Based on the extracted MFCC features, through an end-to-end speech recognition model with CTC loss, input the MFCC feature sequence, and output the matching of text instructions and control instruction templates.

[0013] Preferably, the application method of the system includes: In pilot training, the instructor uses the system to set up various simulation scenarios according to the training syllabus and the situation of the trainees. The trainees interact with the virtual low-altitude air traffic controller to receive instruction operations. The instructor monitors and guides the trainees through the interface, and uses the system data analysis to formulate personalized training plans; In the simulation of the low-altitude traffic management system, researchers and developers use the system to build complex scenarios to test and verify management strategies, algorithms and technologies. Evaluation indicators such as flight efficiency, safety and airspace utilization rate are used to collect and analyze data for improvement and optimization. It is also used to simulate the mixed flight of different types of low-altitude aircraft, providing a basis for formulating rules and standards.

[0014] Preferably, the specific implementation manner of the system includes: Build each module of the system and the hardware infrastructure, initialize and load the basic data and knowledge, and pre-train the AI core model; Continuously collect and update simulation data to enrich the training set, use distributed training technology to accelerate training, and optimize the model according to the evaluation results; Integrate the system with simulation devices and platforms and conduct tests to ensure the normal operation of data interaction and function integration, and conduct joint debugging tests and compatibility tests; Promote the application in multiple fields, collect feedback to improve performance and experience, and update the system knowledge base and model according to technological development and regulatory changes.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: The virtual low-altitude air traffic controller AI model training system proposed by the present invention significantly improves the effect and quality of pilots' low-altitude flight training. By simulating various real low-altitude flight scenarios and complex situations, pilots can conduct a large number of practical operations and emergency response trainings in a safe environment, enhancing their ability to handle low-altitude flight rules, control procedures, and various complex situations, shortening the training cycle, reducing training costs, improving the comprehensive quality and flight safety of pilots, and providing an efficient, flexible, and realistic simulation platform for the research, development, and optimization of the low-altitude traffic management system. It can quickly verify the feasibility and effectiveness of different low-altitude traffic management strategies, algorithms, and technologies, accelerate the innovation and development of low-altitude traffic management technologies, improve the utilization rate and management efficiency of low-altitude airspace, and promote the healthy development of the low-altitude flight industry. Since the system is only used for training and simulation purposes, it avoids the safety risks that may be brought by directly applying unproven AI technologies in the actual low-altitude traffic management system, and provides strong support and guarantee for technological progress and talent cultivation in the low-altitude flight field on the premise of ensuring low-altitude flight safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 FIG. is the overall architecture diagram of the virtual low-altitude air traffic controller AI model training and application system; Figure 2 FIG. is the internal structure and data flow diagram of the AI core model module; Figure 3 FIG. is the schematic diagram of data association and synchronization with the pilot training simulation device; Figure 4 FIG. is the integrated architecture diagram with the low-altitude traffic management system simulation platform; Figure 5 FIG. is the schematic diagram of the model training process. DETAILED DESCRIPTION OF THE INVENTION

[0017] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art can think of other obvious variations.

[0018] REFERENCE Figure 1As shown in the figure, the data layer includes low-altitude radar simulation data sources, low-altitude meteorological simulation data sources, low-altitude geographic information data sources, and low-altitude flight plan data sources, which respectively provide information such as the position, meteorology, geography, and flight plan of the aircraft. These data are collected into the data acquisition and preprocessing module. The AI core model module in the model layer includes a low-altitude control instruction generation model, a low-altitude traffic situation awareness and prediction model, and a simulation scenario generation and management model. They conduct data interaction with each other and receive the input of preprocessed data. The application layer has a pilot training simulation application and a low-altitude traffic management system simulation application, which receive the data output by the AI core model module and conduct data interaction with the corresponding devices or platforms. The instructor operation interface in the user layer can perform operations such as parameter setting on the model and monitor the operation of the application layer.

[0019] Reference Figure 2 As shown in the figure, the low-altitude control instruction generation model receives the input data such as the preprocessed flight situation, flight plan, and text features of regulations and norms, and generates instructions such as takeoff, cruise, and landing permits, which are output to the relevant parts of the application. The low-altitude traffic situation awareness and prediction model takes the real-time aircraft dynamics, geography, and historical traffic flow data (optional) as input, outputs the real-time traffic situation analysis results to the control instruction generation model, and at the same time outputs the predicted traffic situation data for display. The simulation scenario generation and management model generates a complete simulation scenario based on the training requirement parameters and basic geography and meteorological data, and provides it to the control instruction generation model, the situation awareness and prediction model, and the driving simulation application respectively.

[0020] Reference Figure 3 As shown in the figure, the virtual low-altitude air traffic controller system sends simulation data such as the aircraft status and control instruction data to the pilot training simulation equipment, and the pilot training simulation equipment then feeds back the pilot operation information and equipment operation status data to it, realizing the association and synchronization of data to support the pilot training simulation application.

[0021] Reference Figure 4 As shown in the figure, the virtual low-altitude air traffic controller system sends traffic situation awareness data and control instruction data to the simulation platform, and conducts docking and interaction through the interface-related data; after receiving the data, the simulation platform displays, executes the instructions, and feeds back the execution situation, and at the same time shares the simulation scenario configuration data to realize the collaborative work of both parties to support the low-altitude traffic management system simulation application.

[0022] Reference Figure 5As shown, starting from the starting point, the data collection module collects various types of data, such as historical simulated flight records, etc., and then passes the data to the data annotation and preprocessing module to form a dataset with detailed information. If the data meets the training requirements, it enters the supervised learning training module to obtain model parameters and structural status, and then successively passes through the reinforcement learning training module (to further optimize the model status) and the generative adversarial network training module (optional step), and finally evaluates the results of multiple indicators in the model evaluation and verification module. If the performance target is reached, the model saving and version management module saves the model and related information; if not, it returns to the data collection module to restart the training process; if the data does not meet the training requirements, it also needs to return to the data collection module to re-obtain appropriate data. This process demonstrates the complete process of model training and the logical relationship between each link.

[0023] The specific implementation method process of this system is as follows: (I) System construction and initialization According to the above system architecture, respectively construct the data acquisition and preprocessing module, the AI core model module, the communication and interaction module, and related hardware infrastructure. In the data acquisition and preprocessing module, configure the connection interfaces and data acquisition software with each simulation data source to ensure that relevant data for low-altitude flight simulation can be obtained stably and efficiently, and perform preprocessing and storage. For the AI core model module, select a suitable deep learning framework (such as TensorFlow or PyTorch), construct a model network structure based on technologies such as natural language processing, graph neural network, and recurrent neural network, and perform initial parameter settings. In the communication and interaction module, install and configure communication devices and software, establish communication links with pilot training simulation devices and low-altitude traffic management system simulation platforms, and develop instructor operation interfaces and user interaction interfaces to ensure the operability and visual display effect of the system.

[0024] Perform initialization operations on the system, including loading basic data and knowledge such as initial low-altitude geographic information data, low-altitude flight rules and regulations libraries, standard control instruction templates, and preset simulated flight scenario libraries to provide initial support and basis for the operation of the system. At the same time, pre-train the AI core model, use a small amount of initial training data to perform preliminary parameter adjustment and optimization on the model, so that it has basic low-altitude air traffic control instruction generation and traffic situation perception capabilities, and make preparations for subsequent formal training and applications.

[0025] (II) Model training and optimization Continuously collect and update low-altitude flight simulation data to continuously enrich the diversity and scale of the training dataset. Using the data collection and preprocessing module, regularly obtain the latest low-altitude flight simulation data from data sources such as simulation radar systems, meteorological simulation systems, and flight plan simulation systems, and integrate and annotate it with historical data to form new training samples. At the same time, generate various complex low-altitude flight scenarios and emergency situations through simulation flight software to further expand the coverage of the training data and improve the model's response ability to different situations. For example, simulate low-altitude flights under different weather conditions, such as heavy rain, strong wind, low visibility, etc., as well as various aircraft failures and anomalies, such as engine flameout, communication failure, instrument malfunction, etc., so that the model can learn the control strategies and response methods in these extreme situations.

[0026] Adopt distributed training technology and use multiple computing devices (such as GPU clusters) to perform parallel training on the AI core model to accelerate the training process and improve training efficiency. During the training process, use an adaptive learning rate adjustment strategy to dynamically adjust the learning rate according to the training progress and performance of the model, ensuring that the model can update parameters at an appropriate step size in different training stages, avoid falling into local optima, and improve the convergence speed and performance of the model. At the same time, introduce model regularization techniques, such as L1 and L2 regularization, Dropout, etc., to prevent model overfitting, enhance the generalization ability of the model, and enable the model to accurately generate control instructions and predict traffic situations in unseen low-altitude flight scenarios.

[0027] Regularly evaluate and optimize the model. Use an independent test dataset to perform performance tests on the trained model. The evaluation metrics include the accuracy, recall rate, F1 value of control instruction generation, the accuracy, false alarm rate, and miss rate of traffic conflict prediction, and the system response time, etc. According to the evaluation results, make targeted adjustments and optimizations to the model. For example, if it is found that the accuracy of control instruction generation by the model is low in certain specific scenarios, the reason may be insufficient training data or an unreasonable model structure. To address this issue, the amount of training data for relevant scenarios can be increased, or the model structure can be adjusted, such as increasing the number of network layers, adjusting the neuron connection method, etc., and then retrain and evaluate until the performance of the model reaches the expected goal.

[0028] (III) System Integration and Testing Integrate the virtual low-altitude air traffic controller system with the pilot training simulation equipment and conduct joint debugging tests to ensure accurate data association and synchronization between the system and the simulation equipment, and clear and stable audio communication. During the integration process, strictly test and verify the data interfaces, check the accuracy, integrity, and real-time nature of data transmission, ensure that information such as the position, speed, and status of the simulated aircraft can be transmitted to the virtual low-altitude air traffic controller system in a timely and accurate manner, and at the same time, the control instructions generated by the system can also be conveyed to the simulation equipment quickly and accurately, enabling the pilot to receive accurate instructions and make corresponding operations during the simulated flight. Test the audio communication system, check the accuracy of speech recognition and text-to-speech conversion, as well as the stability of two-way communication, to avoid problems such as voice signal interruption, distortion, or incorrect instruction transmission, and ensure the smooth progress of simulated flight training.

[0029] Integrate the virtual low-altitude air traffic controller system with the low-altitude traffic management system simulation platform and conduct functional tests and compatibility tests to ensure the normal operation of data interaction and functional integration between the system and the simulation platform and their ability to work together. In the functional tests, verify the implementation of functions such as control instruction generation, traffic situation awareness, and simulation scenario management of the virtual low-altitude air traffic controller system in the simulation platform, and check whether it can meet the various requirements of the simulation platform for low-altitude traffic management, such as whether it can accurately generate control instructions that conform to the simulation scenario, whether it can perceive and predict the low-altitude traffic situation in real time and display it visually on the platform, and whether it can flexibly adjust the simulation scenario according to the instructions and data of the simulation platform. In the compatibility tests, check whether the system is compatible with the software environment, hardware devices, communication protocols, etc. of the simulation platform, ensure that the two can work together stably under different configurations and operating conditions, avoid compatibility problems such as system crashes, data loss, and communication anomalies, and ensure the accuracy and reliability of the low-altitude traffic management system simulation.

[0030] (IV) Application Promotion and Continuous Improvement Carry out application promotion in fields such as pilot training schools, aviation training institutions, and scientific research institutions, showcase the functions and advantages of the system to potential users, provide trial use and training services, collect user feedback, and continuously improve the system's performance and user experience. According to the needs and suggestions of users, expand and optimize the functions of the system, such as adding more simulated flight scenarios and emergency situations, improving the usability and functionality of the instructor operation interface, and enhancing the stability and reliability of the system, etc., to meet the diverse needs of different users and improve the market competitiveness and application value of the system.

[0031] Continuously monitor the technological developments and regulatory policy changes in the low-altitude flight field, and update the system's knowledge base and models in a timely manner to ensure that the system always complies with the latest low-altitude flight standards and requirements. For example, when new low-altitude flight regulations are introduced, incorporate the relevant regulatory content into the system's knowledge system in a timely manner so that the control instructions generated by the virtual low-altitude air traffic controller comply with the regulatory requirements; when new low-altitude flight technologies emerge (such as the emergence of new types of aircraft, the application of new navigation technologies, etc.), upgrade and improve the system accordingly to enable it to adapt to the new flight environment and technical conditions, providing continuous support and guarantee for pilot training and the simulation of the low-altitude air traffic management system.

[0032] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification is only the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection required by the present invention is defined by the appended claims.

Claims

1. Virtual low-altitude air traffic controller AI model training system, characterized by: include: Data acquisition and preprocessing module, AI core model module and communication and interaction module; The data acquisition and preprocessing module is used to collect data from simulated low-altitude radar, low-altitude meteorology, low-altitude geographic information, and low-altitude flight plan data sources, and clean, denoise, format convert, and integrate the data to generate preprocessed data for use by the AI ​​core model; The AI ​​core model module is electrically connected to the data acquisition and preprocessing module, and the AI ​​core model module includes: a low-altitude control instruction generation model, a low-altitude traffic situation perception and prediction model, and a simulation scenario generation and management model. The low-altitude control instruction generation model is constructed based on the natural language processing technology of deep learning. After learning a large number of low-altitude flight control instruction samples and relevant laws and regulations, it can generate control instructions that meet safety requirements according to the low-altitude flight situation. The low-altitude traffic situation perception and prediction model adopts a structure that combines a graph neural network and a recurrent neural network. By modeling the low-altitude traffic network and learning the time series characteristics of the traffic flow, it can perceive and predict the flight trajectory and potential conflict risks of low-altitude aircraft in real time. The simulation scenario generation and management model generates various low-altitude flight simulation scenarios according to training requirements, and monitors and adjusts them in real time; The communication and interaction module is electrically connected to the AI ​​core model module. The communication and interaction module is used to establish a two-way communication link with the pilot training simulation equipment and a data interaction interface with the low-altitude traffic management system simulation platform. It also includes a user interface for instructor operation. The user interface can display low-altitude flight situation diagrams and related data, and realize control and intervention of simulated training scenarios and virtual low-altitude air traffic controller behaviors.

2. The virtual low-altitude air traffic controller AI model training system according to claim 1 is characterized in that: The data collected by the data collection and preprocessing module includes: The collected simulated low-altitude radar data covers the position, speed, altitude, and heading information of low-altitude aircraft of different models, flight trajectories, and speed changes; Among them, low-altitude meteorological simulation data include but are not limited to: wind speed, wind direction, temperature, air pressure, visibility, cloud height and precipitation probability; Low-altitude geographic information data includes but is not limited to: topography, obstacle distribution, airport and take-off and landing point locations; Low-altitude flight plan data includes, but is not limited to: detailed information on general aviation flight plans, drone mission plans, and low-altitude tourism route plans; The preprocessing of the collected data specifically includes: The radar simulation data is corrected for outliers and the sliding window filtering algorithm is used to smooth the noise. The filtering formula is: , in, For the window The original data at that moment, is the window length, is the output after filtering; The coordinate system of geographic information data is unified, and data in different coordinate systems are converted to the WGS84 coordinate system. The conversion formula is: ， in, Represents the converted WGS84 coordinate system coordinates, , is the coordinate value in the original coordinate system, , is the translation parameter, is the rotation angle; The coordinate system conversion function Calculate by following the steps below: The original coordinates Angle of counterclockwise rotation around the origin , the calculation formula is: , In the formula, is the middle coordinate; The rotated coordinates along Axis Translation , Axis Translation , get the WGS84 coordinates: , The meteorological data is interpolated in time and space, and the missing values ​​are filled by Kriging interpolation method. The interpolation formula is: , in, is the spatial position of the point to be interpolated, For the The spatial location of the known meteorological observation points, For the Meteorological parameter values ​​at each observation point, is the weight coefficient, satisfying .

3. The virtual low-altitude air traffic controller AI model training system according to claim 1 is characterized in that: The specific model for generating low-altitude control instructions is as follows: Natural language processing model based on Transformer architecture, with flight situation vector as input and flight plan text , the output is structured control instructions ; The model optimizes parameters through the cross entropy loss function: ， in, is the real instruction label, is the predicted probability; The training data set of the low-altitude control instruction generation model contains the following annotation information: The initial state vector of the aircraft ; Control instruction sequence , each instruction Mark the corresponding trigger conditions And execution results ; Abnormal scene annotation, including engine fault codes And corresponding emergency command set ; The course learning strategy is adopted during training, and the training is divided into stages according to the complexity of the scene. The initial stage only contains normal instructions of a single aircraft, and the final stage covers multi-aircraft mixed conflict scenarios. The loss function is weighted as: ， in, is the reinforcement learning reward function, , ; The generated control instructions include: aircraft identification before takeoff, departure airport and destination airport, route information, transponder code, altitude permission, departure frequency and departure time; reasonable altitude, speed and heading permission during the low-altitude cruising phase; landing permission during the landing phase that accurately reflects the runway usage, weather conditions, and distribution of surrounding obstacles, as well as accurate sending and confirmation of flight system messages.

4. The virtual low-altitude air traffic controller AI model training system according to claim 1 is characterized in that: The low-altitude traffic situation awareness and prediction model is as follows: A hybrid structure of graph neural network (GNN) and long short-term memory network (LSTM) is used, and the input is the aircraft node feature matrix and the adjacency matrix , output for the future Traffic conflict probability at time and avoidance path set ; The node update formula of the GNN is: ， in, For the Layer Node The hidden state of is the activation function, is the weight matrix, For splicing operation, For the Layer Node The hidden state of For Node All neighbors In the The sum of the hidden states of the layer, For Node The set of neighbors of ; The conflict detection algorithm of the low-altitude traffic situation awareness and prediction model is: Constructing the space-time trajectory matrix of the aircraft , including longitude, latitude, altitude, and speed; calculate the future Time two aircraft and Euclidean distance : ， In the formula, For the future Time Flying Machine and The Euclidean distance between , , For aircraft exist The three-dimensional coordinates of the time, , , For aircraft exist The three-dimensional coordinates of the moment; like , it is marked as a potential conflict, where Indicates the safety threshold.

5. The virtual low-altitude air traffic controller AI model training system according to claim 1 is characterized in that: The simulation scenario generation and management model is as follows: Based on the generative adversarial network (GAN), a variety of flight scenes are generated. The adversarial loss between the generator and the discriminator is: ， In the formula, is the total loss function of GAN, is the expectation operator, Represents the distribution from real data Samples sampled from , For the discriminator For real samples The output probability of represents the noise distribution input from the generator The noise sampled , For the generator According to the noise The generated simulation samples, For the discriminator To generate samples The output probability of The scenario diversity control method of the simulation scenario generation and management model includes: Defining the scene parameter space ; Generate the initial scene set using Latin hypercube sampling ; The difference between the generated scene and the real data distribution is evaluated by KL divergence: ； In the formula, is the KL divergence, is the probability distribution of the real data, is the probability distribution of generated data, For real data samples The probability of occurrence is the sample in the generated data Probability of occurrence; like , then adjust the generator hidden space dimension The sampling distribution of .

6. The virtual low-altitude air traffic controller AI model training system according to claim 1 is characterized in that: The training method of the AI ​​core model includes: Collect historical simulated flight records, standard flight scenarios, instructor demonstration operations, and abnormal emergency simulation data to build a training data set, and pre-process the data with detailed annotations, covering flight scenarios, control instructions, and flight result information; A multi-level training strategy combining supervised learning, reinforcement learning and generative adversarial network technology is adopted. The supervised learning stage minimizes the error between prediction and actual command, the reinforcement learning stage optimizes the decision-making strategy according to flight safety and efficiency indicators, and the generative adversarial network technology improves simulation capabilities and command flexibility. Regularly evaluate and verify model performance using independent test data sets. Indicators include control instruction generation accuracy, traffic conflict prediction accuracy, and system response time. Adjust and optimize model structure parameters, training data, and algorithms based on the results, and establish a version management mechanism to retrospectively select the optimal version.

7. The virtual low-altitude air traffic controller AI model training system according to claim 1 is characterized in that: The communication and interaction module and the data interaction sub-function of the low-altitude traffic management system simulation platform realize two-way data sharing and collaborative work, output the virtual low-altitude air traffic controller's decision instructions and receive platform feedback information, and the function integration sub-function embeds some system function modules into the simulation platform.

8. The virtual low-altitude air traffic controller AI model training system according to claim 1 is characterized in that: The voice communication subsystem implementation method of the communication and interaction module is as follows: Using pre-emphasis filter Eliminate low-frequency noise and pre-process the speech signal, where: represents the transfer function of the pre-emphasis filter, represents the complex variable of the Z transform, Indicates a unit delay operation; Based on the preprocessed speech signal, the speech signal is divided into frames, wherein the frame length is 25ms and the frame shift is 10ms; Apply the Hamming window function to each frame signal to reduce spectrum leakage, where the Hamming window function expression is: ， In the formula, The Hamming window is The value of the sampling point, is the DC component coefficient of the Hamming window, is the cosine component coefficient of the Hamming window, Represents the parameters of the cosine function; Represents the parameters of the cosine function; Perform fast Fourier transform on each frame of the windowed signal to calculate the Mel spectrum coefficients , with a dimension of 40, where It is the dimension index of MFCC coefficients and completes MFCC feature extraction; Based on the extracted MFCC features, the end-to-end speech recognition model with CTC loss is used to input the MFCC feature sequence and output the text instructions to match the control instruction template.

9. The virtual low-altitude air traffic controller AI model training system according to claim 1 is characterized in that: The application method of the system includes: In pilot training, instructors use the system to set up a variety of simulation scenarios based on the training syllabus and the trainees' situations. Trainees interact with virtual low-altitude air traffic controllers to receive instructions and operate. Instructors monitor and guide trainees through the interface and use system data analysis to develop personalized training plans. In the simulation of the low-altitude traffic management system, researchers and developers use the system to build complex scenarios to test and verify management strategies, algorithms and technologies, evaluate indicators such as flight efficiency, safety and airspace utilization, collect and analyze data for improvement and optimization, and also simulate mixed flight conditions of different types of low-altitude aircraft to provide a basis for the formulation of rules and standards.

10. The virtual low-altitude air traffic controller AI model training system according to claim 1 is characterized in that: Specific implementations of the system include: Build system modules and hardware infrastructure, initialize and load basic data and knowledge, and pre-train AI core models; Continuously collect and update simulation data to enrich the training set, use distributed training technology to accelerate training, and optimize the model based on evaluation results; Integrate the system and simulate the equipment and platform and test them to ensure the normal operation of data interaction and functional integration, and conduct joint debugging and compatibility testing; Promote applications in multiple fields, collect feedback to improve performance experience, and update system knowledge base and models based on technological development and regulatory changes.

Images