Method and system for generating mass flight simulation data of unmanned aerial vehicle

Abstract

The invention provides a method and a system for generating mass unmanned aerial vehicle flight simulation data. The method comprises the following steps: S101, collecting data; s102, recording and storing data; s103, processing the data; s104, carrying out data modeling; s105, generating data; s106, outputting the data; and S107, adjusting the data in real time. According to the invention, various virtual weather conditions, terrains and obstacles can be created, and different flight scenes and coping strategies can be tested; detailed flight data including position, speed and the like of the unmanned aerial vehicle are provided to help comprehensive understanding of flight behaviors of the unmanned aerial vehicle; data modeling is carried out through the combination of three loss functions, the performance and training speed of a data model are improved, and sample features and convergence gradients are rapidly extracted.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles, and in particular to a method and system for generating massive unmanned aerial vehicle flight simulation data. Background Art

[0002] The rapid development of UAV technology has made the application of UAVs increasingly popular in various fields, and the demand for UAV flight simulation data has gradually increased. Traditional flight simulation methods mainly rely on field flight experiments. However, field flight experiments are not only costly, but also affected by factors such as weather and airspace restrictions, making it difficult for data collection to meet the needs. On the other hand, the number of UAV-related control systems is increasing, and system acceptance requires a large amount of data to verify system functions and performance. Traditional acceptance requires actual field flights for verification, which is inefficient and has poor acceptance results. It also needs to consider factors such as weather, no-fly airspace, and plan applications, which makes it difficult to reflect the diversity and complexity of real flight scenarios. Summary of the invention

[0003] In view of this, an object of the present invention is to provide a method and system for generating massive UAV flight simulation data, so as to solve or at least partially solve the above-mentioned problems existing in the prior art.

[0004] To achieve the above-mentioned object of the invention, the present invention provides a method and system for generating massive UAV flight simulation data, the method comprising the following steps:

[0005] S101, data collection: obtain real drone flight data;

[0006] S102, data recording and storage: storing the acquired real UAV flight data in the cloud or in a local database, and recording the attribute information of the real UAV flight data;

[0007] S103, data processing: processing the real UAV flight data;

[0008] S104, data modeling: performing data modeling on the real UAV flight data that has been processed by traditional data modeling and machine learning data modeling to generate a data model;

[0009] S105, data generation: generating UAV flight simulation data based on the generated data model, wherein the data generation sets the relevant parameters required for the UAV flight simulation data in two ways: flight plan setting and custom setting;

[0010] S106, data output: outputting the UAV flight simulation data generated by the data;

[0011] S107, real-time data adjustment: dynamically adjust the UAV flight simulation data after data output.

[0012] Furthermore, the steps of data processing are as follows:

[0013] S31. Data cleaning: Process missing values, delete duplicate data, and monitor and process outliers on real UAV flight data;

[0014] S32, data conversion: format conversion, data normalization / standardization, feature engineering, feature extraction and feature selection of real UAV flight data after data cleaning;

[0015] S33, data integration: align the real UAV flight data after data conversion, and then merge the data to obtain a unified data set;

[0016] S34, data segmentation: split the unified data set into training set, test set and validation set;

[0017] S35, Data balance: Dealing with unbalanced data in training, test and validation sets;

[0018] S36, Data desensitization: Pseudonymize and encrypt the data after data balancing;

[0019] S37. Data processing record: record each step of data processing.

[0020] Furthermore, the traditional data modeling includes:

[0021] Physical modeling: Dynamic model and aerodynamic model are used for physical modeling;

[0022] Environmental modeling: Describes the external conditions and environmental factors of drone flight based on meteorological conditions, terrain and obstacles, and sets wind speed and terrain data for the flight area.

[0023] Furthermore, the machine learning data modeling includes:

[0024] The variational autoencoder (VAEs) model is used for machine learning data modeling. The variational autoencoder model consists of an encoder, a latent space sampling layer, a convolutional layer, a pooling layer, and a decoder. The variational autoencoder model is defined.

[0025] Furthermore, the model definition includes:

[0026] Encoder: It is configured to encode the input real UAV flight data into the distribution parameters of the latent space, including the mean and variance of the position and velocity;

[0027] Latent space sampling layer: This is configured to sample using the distribution parameters of the mean and variance of the velocity, sampling a latent variable from the distribution parameters generated by the encoder;

[0028] Convolutional layer: It is configured to extract neighboring data features of real drone flight data;

[0029] Pooling layer: It is configured to reduce the number of features extracted by the convolutional layer;

[0030] Decoder: This is configured to decode the latent variables into reconstructed samples in the data space.

[0031] Furthermore, the machine learning data modeling includes:

[0032] Define the loss function of the variational autoencoder and train the variational autoencoder model. The loss function of the variational autoencoder consists of the contrast loss function, the triple loss function and the mean square error loss function.

[0033] Furthermore, the steps of defining the loss function of the variational autoencoder are as follows:

[0034] S41. Define contrast loss function, triplet loss function and mean square error loss function;

[0035] The contrast loss function is defined as follows:

[0036]

[0037] Among them, L contrastive is the contrast loss function, y is the label classification, f(x1) is the spatial position of sample 1, f(x2) is the spatial position of sample 2, and m is the predefined margin;

[0038] The triplet loss function is defined as follows:

[0039] L triplet =max(0,∥f(A)-f(P)∥ 2 -∥f(A)-f(N)∥ 2 +α)

[0040] Among them, L triplet is the triplet loss function, f(A) is the spatial position of the anchor sample, f(P) is the spatial position of the positive sample, f(N) is the spatial position of the negative sample, and α is a hyperparameter;

[0041] The mean square error loss function is defined as follows:

[0042]

[0043] Among them, L MSEis the mean square error loss function, n is the number of samples, i is the sample number, y i is the true value of the i-th sample, is the predicted value of the i-th sample;

[0044] S42. Based on the contrast loss function, the triplet loss function and the mean square error loss function, three dynamic valve variables are introduced. The dynamic valve variables are dynamically adjusted through a one-dimensional variation function and a smoothing function.

[0045] The one-dimensional variation function is defined as follows:

[0046] g contrastive (t)=min(1,a contrastive ·t+b contrastive )

[0047] g triplet (t)=min(1,a triplet ·t+b triplet )

[0048] g MSE (t)=min(1,a MSE ·t+b MSE )

[0049] Among them, gcontrastive(t) is the contrast loss dynamic valve variable, g triplet (t) is the triplet loss dynamic valve variable, g MSE (t) is the mean square error loss dynamic valve variable, a contrastive is the contrast loss slope, a triplet is the triplet loss slope, a MsE is the mean square error loss slope, b contrastive is the contrast loss intercept, b triplet is the triplet loss intercept, b MSE is the mean square error loss intercept, t is the training gradient;

[0050] The smoothing function is defined as follows:

[0051]

[0052]

[0053]

[0054] in, is the slope of the curve after the comparison loss converges, is the slope of the curve after the triplet loss converges, is the slope of the curve after the mean square error loss converges, (t-t0) is the distance between the control curve and the center, t0 is the center position of the control curve, t is the current position of the control curve, and k contrastive To compare the steepness of the loss control curve, k triplet is the steepness of the triplet loss control curve, k MSE is the steepness of the mean square error loss control curve;

[0055] Combine and calculate the one-dimensional variation function and the smoothing function as follows:

[0056]

[0057]

[0058]

[0059] S43. According to the defined valve variable, the contrast loss function, the triplet loss function and the mean square error loss function are combined into a total loss function, which is defined as follows:

[0060] L total (t) = g triplet (t)·L triplet +g contrastive (t)·L contrastive +g MSE (t)·L MSE

[0061] Among them, L total (t) is the total loss function.

[0062] Furthermore, the flight plan settings and custom settings include:

[0063] The flight plan setting imports the flight plan through a visual interface, and the data model generates the UAV flight simulation data according to the flight plan content;

[0064] The custom settings include scene settings, drone model settings, flight behavior mode settings, data volume settings, and flight area / route settings. After completing the scene settings, select a mode for generating data.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] A method and system for generating massive UAV flight simulation data proposed in the present invention can create various virtual weather conditions, terrains and obstacles to test different flight scenarios and response strategies; can provide detailed data of UAV flight, including position, speed, etc., to help fully understand the flight behavior of UAV; through the combination of three loss functions to carry out data modeling, improve the performance and training speed of the data model, quickly extract sample features and convergence gradients, and can also control the training cost of the model, with practical economy. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only preferred embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0068] Figure 1 A flowchart of a method and system for generating massive UAV flight simulation data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The enumerated embodiments are only used to explain the present invention and are not used to limit the scope of the present invention.

[0070] This embodiment provides a method and system for generating massive UAV flight simulation data, the method comprising the following steps:

[0071] like Figure 1 As shown, a flow chart of a method and system for generating massive UAV flight simulation data is provided, including:

[0072] S101. Data collection: Obtain real drone flight data, including:

[0073] Acquire real drone flight data through internal system databases, open source data sets, protocol data interfaces, file transfers, etc., including but not limited to flight trajectory, attitude, speed, altitude, environmental conditions and other data.

[0074] S102, data recording and storage: The acquired real UAV flight data is stored in the cloud or in a database, and the attribute information of the real UAV flight data is recorded, including:

[0075] The acquired real UAV flight data is stored in a secure location in an appropriate format, such as a database, data warehouse or cloud storage, to ensure the security and privacy of the real UAV flight data, while recording attribute information such as the source of the real UAV flight data, flight environment, and UAV type.

[0076] S103, data processing: processing the real UAV flight data, specifically including:

[0077] S31. Data cleaning: Process missing values, delete duplicate data, and monitor and process outliers on real UAV flight data;

[0078] Missing value processing: fill in missing values ​​(such as using mean, median, most frequent value, etc.) or delete records with a large number of missing values;

[0079] Deduplication: Remove duplicate records to ensure the uniqueness of real drone flight data;

[0080] Outlier detection and processing: Identify and process outliers (e.g., replace or delete them with other values).

[0081] S32, data conversion: format conversion, data normalization / standardization, feature engineering, feature extraction and feature selection of real UAV flight data after data cleaning;

[0082] Format conversion: convert real drone flight data into the required format;

[0083] Data normalization / standardization: Scale the real drone flight data to a specific range (such as 0-1) or make it have a standard normal distribution to improve the convergence speed and effect of the model;

[0084] Feature engineering: creating new features or transforming existing features to improve data model performance;

[0085] Feature extraction: Extract useful features from real drone flight data;

[0086] Feature selection: Select features that have a significant impact on the data model and remove irrelevant or redundant features.

[0087] S33, data integration: align the real UAV flight data after data conversion, and then merge the data to obtain a unified data set;

[0088] Data alignment: ensuring that records between different real drone flight data are aligned, such as based on timestamp alignment;

[0089] Data merging: Merge different real drone flight data into a unified dataset.

[0090] S34, data segmentation: split the unified data set into training set, test set and validation set;

[0091] Training set and test set: Split the unified data set into training set and test set to evaluate the performance of the data model;

[0092] Validation set: The validation set is further divided during the training process for parameter adjustment and validation of the data model.

[0093] S35, Data balance: Dealing with unbalanced data in training, test and validation sets;

[0094] Handling unbalanced data: To address the problem of class imbalance, methods such as oversampling, undersampling, and SMOTE can be used.

[0095] S36, Data desensitization: Pseudonymize and encrypt the data after data balancing;

[0096] When processing data, real drone flight data is desensitized by pseudonymization, encryption and other operations to ensure drone-related data privacy and security regulations.

[0097] S37, data processing record: record each step of data processing;

[0098] Record every step of data processing to ensure traceability and reproducibility.

[0099] S104, data modeling: data modeling is performed on the real UAV flight data after data processing through traditional data modeling and machine learning data modeling to generate a data model, specifically including:

[0100] (1) Traditional data modeling: It is divided into physical modeling and environmental modeling. The physical modeling adopts dynamic model and aerodynamic model for physical modeling. The dynamic model is a six-degree-of-freedom dynamic model, which is used to simulate the posture and position changes of the UAV. It takes into account all possible degrees of freedom (translation along the X, Y, and Z axes and rotation around these axes) and combines the actual physical forces (such as gravity, air resistance, thrust, etc.) for simulation, thereby providing highly realistic and accurate data.

[0101] The aerodynamic model is to establish the dynamic equation of the UAV based on the data of different UAV fuselages, such as the size, shape and mass distribution of the fuselage, wings, tail, etc., and obtain the aerodynamic characteristics of the relevant UAV, such as lift, drag, thrust, etc. (these characteristics will be obtained through data collection, experimental data or simulation software);

[0102] The environmental modeling is mainly carried out by meteorological conditions, terrain and obstacles, which is used to describe the external conditions and environmental factors of UAV flight and set the wind speed and terrain data of the flight area. The meteorological conditions include meteorological parameters such as wind speed, wind direction, temperature, humidity, etc. These parameters can be generated through meteorological data or empirical formulas involved in data collection; the terrain and obstacles are used to describe the terrain characteristics and obstacle distribution of the flight area, which can be generated using geographic information system (GIS) data or three-dimensional modeling software.

[0103] (2) Machine learning data modeling: A variational autoencoder (VAE) model is used for machine learning data modeling. The variational autoencoder model can generate drone simulation data with high diversity. The variational autoencoder model can generate new samples and maintain the consistency and diversity of the data by learning the potential representation of the data. The variational autoencoder model consists of an encoder, a latent space sampling layer, a convolutional layer, a pooling layer, and a decoder. The loss function of the variational autoencoder model consists of a contrast loss function, a triplet loss function, and a mean square error loss function. Machine learning data modeling is divided into model definition and definition of the loss function of the variational autoencoder. The variational autoencoder model is trained by defining the loss function of the variational autoencoder. The model definition is expressed as follows:

[0104] Encoder: It is configured to encode the input real UAV flight data into the distribution parameters of the latent space, which include the mean and variance of the position and speed. The input data is normalized so that the data model can recognize and process the data.

[0105] Latent space sampling layer: It is configured to sample using the distribution parameters of the mean and variance of the speed, thereby sampling from a Gaussian distribution. The main purpose is to sample a latent variable from the distribution parameters generated by the encoder, which can automatically extract factors that affect flight from real drone flight data, generate data with the decoder, and optimize the final generated simulation data;

[0106] Convolutional layer: It is configured to extract the neighboring data features of real drone flight data, which is more effective, especially for continuous data. Flight is a continuous behavior, and the data at each moment is associated with the previous and next data. The convolutional layer can extract the neighboring data features to make each segment of data smoother.

[0107] Pooling layer: It is configured to reduce the number of features extracted by the convolutional layer, prevent overfitting, and reduce the amount of computation;

[0108] Decoder: This is configured to decode the latent variables into reconstructed samples in the data space.

[0109] The contrast loss function can reduce the distance between similar sample pairs in the feature space and expand the distance between different sample pairs, so that the features of UAV simulation data (such as position, speed, acceleration, etc.) can be distinguished and extracted; the triple loss function trains the model by reducing the difference between the anchor sample (anchor) and the positive sample (positive) and expanding the gap between the anchor sample and the negative sample (negative), which can make the gradient of the data model converge quickly and reduce the dependence on training data and computing resources.

[0110] The steps of defining the loss function of the variational autoencoder are as follows:

[0111] S41. Define contrast loss function, triplet loss function and mean square error loss function;

[0112] The contrast loss function is defined as follows:

[0113]

[0114] Among them, L contrastive is the contrast loss function, y is the label classification, y=1 means that the two samples belong to the same class, y=0 means that the two samples belong to different classes, f(x1) is the spatial position of sample 1, f(x2) is the spatial position of sample 2, m is the predefined margin, and m represents the minimum distance between samples of different classes;

[0115] The triplet loss function is defined as follows:

[0116] L triplet =max(0,∥f(A)-f(P)∥ 2 -∥f(A)-f(N)∥ 2 +α)

[0117] Among them, L triplet is the triplet loss function, f(A) is the spatial position of the anchor sample, f(P) is the spatial position of the positive sample, f(N) is the spatial position of the negative sample, α is a hyperparameter, and α represents the minimum distance difference between positive and negative samples;

[0118] The mean square error loss function is defined as follows:

[0119]

[0120] Among them, L MSE is the mean square error loss function, n is the number of samples, i is the sample number, y i is the true value of the i-th sample, is the predicted value of the i-th sample;

[0121] S42. Based on the contrast loss function, the triplet loss function and the mean square error loss function, three dynamic valve variables are introduced. The dynamic valve variable is a function of the number of training iterations t. The dynamic valve variable determines the weight of each loss function in the total loss. The dynamic valve variable is dynamically adjusted through a one-dimensional variation function and a smoothing function.

[0122] The one-dimensional variation function is defined as follows:

[0123] g contrastive (t)=min(1,a contrastive ·t+b contrastive )

[0124] g triplet (t)=min(1,a triplet ·t+b triplet )

[0125] g MSE (t)=min(1,a MSE ·t+b MSE )

[0126] Among them, g contrastive (t) is the dynamic valve variable of the comparative loss, g triplet (t) is the triplet loss dynamic valve variable, g MSE (t) is the mean square error loss dynamic valve variable, a contrastive is the contrast loss slope, a triplet is the triplet loss slope, a MSE is the mean square error loss slope, b contrastive is the contrast loss intercept, b triplet is the triplet loss intercept, b MSE is the mean square error loss intercept, and t is the training gradient;

[0127] The smoothing function is defined as follows:

[0128]

[0129]

[0130]

[0131] in, is the slope of the curve after the comparison loss converges, is the slope of the curve after the triplet loss converges, is the slope of the curve after the mean square error loss converges, (t-t0) is the distance between the control curve and the center, t0 is the center position of the control curve, t is the current position of the control curve, and k contrastive To compare the steepness of the loss control curve, k tripletis the steepness of the triplet loss control curve, k MSE To control the steepness of the mean square error loss curve, the smoothing function can control the rate of change of the weight parameter, which converges slowly in the early stage of training and faster in the later stage;

[0132] Combine and calculate the one-dimensional variation function and the smoothing function as follows:

[0133]

[0134]

[0135]

[0136] S43. According to the defined dynamic valve variable, the contrast loss function, the triplet loss function and the mean square error loss function are combined into a total loss function, which is defined as follows:

[0137] L total (t) = g triplet (t)·L triplet +g contrastive (t)·L contrastive +g MSE (t)·L MSE

[0138] Among them, L total (t) is the total loss function.

[0139] As the training progresses (i.e., as t increases), the weights of each loss function will be dynamically adjusted according to the changes in the valve variable, so that the variational autoencoder model can focus on optimizing different loss functions at different training stages. Through the dynamic valve variable mechanism, the loss function L contrastive , to enhance the distinguishing ability of the embedded representation; as the training progresses, the triple loss function L can be gradually increased triplet weights to further optimize the relative relationship between samples; finally, by gradually increasing the mean square error loss function L MSE The weights of , make the model fit the training data more finely during the convergence stage.

[0140] S105, data generation: generating UAV flight simulation data based on the generated data model. The data generation sets the relevant parameters required for the UAV flight simulation data in two ways: flight plan setting and custom setting, specifically including:

[0141] Flight plan settings: The flight plan can be imported through the visual interface, and the data model will generate UAV flight simulation data according to the flight plan content.

[0142] Custom settings: (1) Scenario settings: select the flight scenario of the drone simulation data through the visual interface. The flight scenario settings include:

[0143] Geographical terrain: plateaus, cities, plains, rivers, forests, rivers, oceans, highways, ports, water conservancy facilities, etc.;

[0144] Flight purposes: logistics, plant protection, aerial photography, emergency rescue, environmental testing, surveying and mapping, law enforcement, power inspection, security inspection, traffic inspection, large-scale activities, test flights, etc.

[0145] Environmental conditions: sunny, cloudy, windy, rainy, snowy, foggy, etc.

[0146] Time period: day, night, custom time period.

[0147] (2) UAV model setting: Select the UAV model through the visual interface. The UAV model setting includes:

[0148] Drone size types: micro, light, small, medium, large.

[0149] Drone types: multi-rotor, fixed-wing, helicopter, vertical take-off and landing fixed-wing, autogyro, airship, and others.

[0150] Drone brands and types: Drone brands include DJI, XAG, Docom, Ehang, Zongheng, Fuya Intelligent, Kobite, and Keweitai.

[0151] (3) Flight behavior mode setting: Select the behavior mode of the UAV flight process through the visual interface. The flight behavior mode setting includes:

[0152] Flight altitude: custom altitude, random.

[0153] Flight speed: custom speed, random.

[0154] Flight modes: beyond visual range flight, within visual range flight.

[0155] Flight anomalies: deviation from the planned trajectory, entering a no-fly zone, flying out of the planned flight area, speeding, falling, drone collision, etc.

[0156] Action settings after arriving at the destination: automatic return, automatic hover, automatic landing.

[0157] (4) Data volume setting: Set the amount of UAV flight simulation data through the visual interface.

[0158] (5) Flight area / route setting: Set the flight area or route of the drone through the visual interface.

[0159] After setting the scene, you need to select the mode for generating data. Selecting the corresponding mode will generate a series of simulated flight data that meets the requirements of these scene settings. The specific data generation modes include:

[0160] Rule data: This mode will generate simulated flight data that meets the requirements of scenario setting through the data model modeled by the traditional data;

[0161] Machine learning data: This mode will generate simulated flight data that meets the scenario setting requirements through the data model modeled by the machine learning data;

[0162] Data cloning: This mode searches for raw data that meets the requirements of the scene settings, and based on the statistical characteristics of the raw data, uses the mean and variance of the raw data to generate synthetic data that is as similar as possible to the distribution of the original data.

[0163] S106, data output: outputting the UAV flight simulation data generated by the data, specifically including:

[0164] The output of UAV flight simulation data can be selected in a variety of formats and methods according to application requirements and environment, including:

[0165] (1) File output: CSV file, JSON file;

[0166] (2) Real-time data stream output: such as WebSocket and MQTT;

[0167] (3) Database output: SQL database, NoSQL database;

[0168] (4) API interface output: Send data through the API interface to facilitate integration with other systems;

[0169] (5) Visual output: Visualize and dynamically display UAV flight simulation data on a three-dimensional or two-dimensional map.

[0170] S107, real-time data adjustment: dynamically adjust the UAV flight simulation data after data output, specifically including:

[0171] During the data output and simulation of UAV flight, the parameters of the UAV flight simulation data can be adjusted in real time through the visual interface, and the flight path, status and behavior of the UAV can be dynamically adjusted. This real-time adjustment helps to simulate a more complex and realistic flight environment. The specific adjustable contents include: flight altitude, flight speed, flight anomalies, environmental conditions, flight area, flight destination, etc. It supports individual or batch setting of UAV flight simulation data parameters, and can set immediate execution or scheduled execution of adjustment instructions, so as to simulate the UAV flight environment more realistically and flexibly.

[0172] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method and system for generating massive UAV flight simulation data, characterized in that: include: S101, data collection: obtain real drone flight data; S102, data recording and storage: storing the acquired real UAV flight data in the cloud or in a local database, and recording the attribute information of the real UAV flight data; S103, data processing: processing the real UAV flight data; S104, data modeling: performing data modeling on the real UAV flight data that has been processed by traditional data modeling and machine learning data modeling to generate a data model; S105, data generation: generating UAV flight simulation data based on the generated data model, wherein the data generation sets the relevant parameters required for the UAV flight simulation data in two ways: flight plan setting and custom setting; S106, data output: outputting the UAV flight simulation data generated by the data; S107, real-time data adjustment: dynamically adjust the UAV flight simulation data after data output.

2. The method and system for generating massive UAV flight simulation data according to claim 1, characterized in that: The steps of data processing are as follows: S31. Data cleaning: Process missing values, delete duplicate data, and monitor and process outliers on real UAV flight data; S32, data conversion: format conversion, data normalization / standardization, feature engineering, feature extraction and feature selection of real UAV flight data after data cleaning; S33, data integration: align the real UAV flight data after data conversion, and then merge the data to obtain a unified data set; S34, data segmentation: split the unified data set into training set, test set and validation set; S35, Data balance: Dealing with unbalanced data in training, test and validation sets; S36, Data desensitization: Pseudonymize and encrypt the data after data balancing; S37. Data processing record: record each step of data processing.

3. The method and system for generating massive UAV flight simulation data according to claim 1, characterized in that: The traditional data modeling includes: Physical modeling: Dynamic model and aerodynamic model are used for physical modeling; Environmental modeling: Describes the external conditions and environmental factors of drone flight based on meteorological conditions, terrain and obstacles, and sets wind speed and terrain data for the flight area.

4. The method and system for generating massive UAV flight simulation data according to claim 1, characterized in that: The machine learning data modeling includes: The variational autoencoder (VAEs) model is used for machine learning data modeling. The variational autoencoder model consists of an encoder, a latent space sampling layer, a convolutional layer, a pooling layer, and a decoder. The variational autoencoder model is defined.

5. The method and system for generating massive UAV flight simulation data according to claim 4, characterized in that: The model definition includes: Encoder: It is configured to encode the input real UAV flight data into the distribution parameters of the latent space, including the mean and variance of the position and velocity; Latent space sampling layer: This is configured to sample using the distribution parameters of the mean and variance of the velocity, sampling a latent variable from the distribution parameters generated by the encoder; Convolutional layer: It is configured to extract neighboring data features of real drone flight data; Pooling layer: It is configured to reduce the number of features extracted by the convolutional layer; Decoder: This is configured to decode the latent variables into reconstructed samples in the data space.

6. The method and system for generating massive UAV flight simulation data according to claim 1, characterized in that: The machine learning data modeling includes: Define the loss function of the variational autoencoder and train the variational autoencoder model. The loss function of the variational autoencoder consists of the contrast loss function, the triple loss function and the mean square error loss function.

7. The method and system for generating massive UAV flight simulation data according to claim 6, characterized in that: The steps of defining the loss function of the variational autoencoder are as follows: S41. Define contrast loss function, triplet loss function and mean square error loss function; The contrast loss function is defined as follows: Among them, L contrastive is the contrast loss function, y is the label classification, f(x1) is the spatial position of sample 1, f(x2) is the spatial position of sample 2, and m is the predefined margin; The triplet loss function is defined as follows: L triplet =max(0,∥f(A)-f(P)∥ 2 -∥f(A)-f(N)∥ 2 +a) Among them, L triplet is the triplet loss function, f(A) is the spatial position of the anchor sample, f(P) is the spatial position of the positive sample, f(N) is the spatial position of the negative sample, and α is a hyperparameter; The mean square error loss function is defined as follows: Among them, L MSE is the mean square error loss function, n is the number of samples, i is the sample number, y i is the true value of the i-th sample, is the predicted value of the i-th sample; S42. Based on the contrast loss function, the triplet loss function and the mean square error loss function, three dynamic valve variables are introduced. The dynamic valve variables are dynamically adjusted through a one-dimensional variation function and a smoothing function. The one-dimensional variation function is defined as follows: g contrastive (t)=min(1,a contrastive ·t+b contrastive ) g triplet (t)=min(1,a triplet ·t+b triplet ) g MSE (t)=min(1,a MSE ·t+b MSE ) Among them, g contrastive (t) is the dynamic valve variable of the comparative loss, g triplet (t) is the triplet loss dynamic valve variable, g MSE (t) is the mean square error loss dynamic valve variable, a contrastive is the contrast loss slope, a triplet is the triplet loss slope, a ,SE is the mean square error loss slope, b contrastive is the contrast loss intercept, b triplet is the triplet loss intercept, b MSE is the mean square error loss intercept, and t is the training gradient; The smoothing function is defined as follows: in, is the slope of the curve after the comparison loss converges, is the slope of the curve after the triplet loss converges, is the slope of the curve after the mean square error loss converges, (t-t0) is the distance between the control curve and the center, t0 is the center position of the control curve, t is the current position of the control curve, and k contrastive To compare the steepness of the loss control curve, k triplet is the steepness of the triplet loss control curve, k MSE is the steepness of the mean square error loss control curve; Combine and calculate the one-dimensional variation function and the smoothing function as follows: S43. According to the defined valve variable, the contrast loss function, the triplet loss function and the mean square error loss function are combined into a total loss function, which is defined as follows: L total (t)=g triplet (t)·L triplet +g contrastive (t)·L contrastive +g MSE (t)·L MSE Among them, L total (t) is the total loss function.

8. The method and system for generating massive UAV flight simulation data according to claim 1, characterized in that: In step S105, the flight plan settings and custom settings include: The flight plan setting imports the flight plan through a visual interface, and the data model generates the UAV flight simulation data according to the flight plan content; The custom settings include scene settings, drone model settings, flight behavior mode settings, data volume settings, and flight area / route settings. After completing the scene settings, select a mode for generating data.

Images