A fixed-wing unmanned aerial vehicle simulation model training method based on a cascade neural network

Abstract

A kind of fixed-wing unmanned aerial vehicle simulation model training method based on cascaded neural network, fixed-wing unmanned aerial vehicle simulation model includes four cascaded neural network sub-models;The method comprises the following steps: obtaining fixed-wing unmanned aerial vehicle simulation dataset, each data in fixed-wing unmanned aerial vehicle simulation dataset includes multiple control parameters of fixed-wing unmanned aerial vehicle, multiple operating states;Wherein, multiple motion states include multiple intermediate motion states multiple final motion states;The fixed-wing unmanned aerial vehicle simulation dataset is split to obtain the corresponding training set of four neural network sub-models, four neural network sub-models include two dynamics sub-models and two kinematics sub-models;The corresponding training set of four neural network sub-models is used to train four neural network sub-models in cascade respectively, to obtain the model after training.The method can carry out adaptive error compensation, reduce fitting error, improve the precision of fitting, effectively reduce the error accumulation in network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a fixed-wing unmanned aerial vehicle simulation model training method based on cascaded neural networks. BACKGROUND

[0002] An unmanned aerial vehicle is an unmanned aerial vehicle that can fly in the air. The types and designs of unmanned aerial vehicles are various, which can be roughly divided into two categories of rotary-wing unmanned aerial vehicles and fixed-wing unmanned aerial vehicles. The fixed-wing unmanned aerial vehicle is a kind of unmanned robot that uses fixed wings to generate lift to fly. It is usually composed of a fuselage, wings, tail, electric motor, remote control equipment and other electronic equipment. The flight principle of the fixed-wing unmanned aerial vehicle is to drive the propeller by the electric motor to generate thrust, so that the fuselage moves forward, and at the same time, the wings generate lift to make the fuselage leave the ground and keep moving in the air. By adjusting the flight angle, thrust, rudder angle and other parameters, the flight direction and height can be changed. The fixed-wing unmanned aerial vehicle is widely used in agriculture, surveying and mapping, environmental monitoring, search and rescue and other fields, and has the advantages of high efficiency, low cost, flexibility, repeatability and wide development prospects in future applications. More and more researchers carry out related experimental research on fixed-wing unmanned aerial vehicles based on unmanned system simulation platform, so it is very important to build an accurate and stable simulation model for the fixed-wing unmanned aerial vehicle to ensure the effectiveness, stability and reliability of the simulation experiment.

[0003] The neural network model has good compatibility and can effectively run on different platforms, which can effectively solve the problem of model use difficulty. However, most current researchers construct the simulation model of the fixed-wing unmanned aerial vehicle by using an integrated neural network, which has certain limitations. SUMMARY

[0004] Therefore, it is necessary to provide a fixed-wing unmanned aerial vehicle simulation model training method based on cascaded neural networks in view of the above technical problems.

[0005] According to a first aspect of the present application, a fixed-wing unmanned aerial vehicle simulation model training method based on cascaded neural networks is provided, the fixed-wing unmanned aerial vehicle simulation model comprising four cascaded neural network sub-models; the method comprises:

[0006] Obtaining a fixed-wing unmanned aerial vehicle simulation data set, each data in the fixed-wing unmanned aerial vehicle simulation data set comprising a plurality of control parameters and a plurality of running states of the fixed-wing unmanned aerial vehicle; wherein the plurality of running states comprise a plurality of intermediate running states and a plurality of final running states;

[0007] The fixed-wing unmanned aerial vehicle simulation dataset is split to obtain corresponding training sets of four neural network sub-models, the four neural network sub-models including two dynamic sub-models and two kinematic sub-models; wherein each training sample in the training set corresponding to each dynamic sub-model takes at least two control parameters and at least one motion state at a first time as characteristic values, and takes at least one intermediate motion state at a second time as a label value; each training sample in the training set corresponding to each kinematic sub-model takes at least one final motion state at the first time and at least one intermediate motion state at the second time output by at least one dynamic sub-model as characteristic values, and takes at least one final motion state at the second time as a label value; the second time is the next time of the first time;

[0008] The four neural network sub-models are cascade trained respectively by using the corresponding training sets of the four neural network sub-models, to obtain a trained fixed-wing unmanned aerial vehicle simulation model.

[0009] In an embodiment, the fixed-wing unmanned aerial vehicle simulation dataset is obtained by:

[0010] In an unmanned aerial vehicle simulation platform, a fixed-wing unmanned aerial vehicle is controlled based on all types of control parameter combinations to perform simulation flight, control parameters and operation states of the fixed-wing unmanned aerial vehicle during the simulation flight are obtained, and a first fixed-wing unmanned aerial vehicle simulation dataset is obtained.

[0011] The continuous space of the first fixed-wing unmanned aerial vehicle simulation dataset is equally divided into n intervals, the interval with the least number of data is taken as a reference interval, and data corresponding to the number of data in the reference interval are randomly selected as the final data of other intervals to ensure that the amount of data in each interval is consistent, thereby obtaining the fixed-wing unmanned aerial vehicle simulation dataset.

[0012] In an embodiment, the control parameters include throttle input instructions, flap input instructions, aileron input instructions, elevator input instructions, and rudder input instructions; the motion states include linear velocity, position, angular velocity, and attitude angle; wherein the intermediate motion states include linear velocity and angular velocity, and the final motion states include position and attitude angle.

[0013] In an embodiment, the two dynamic sub-models include a first dynamic sub-model and a second dynamic sub-model; and the two kinematic sub-models include a first kinematic sub-model and a second kinematic sub-model.

[0014] The input of the first dynamic sub-model is the throttle input instruction at the second time, the flap input instruction, and the linear velocity at the first time, and the output is the linear velocity at the second time;

[0015] The input of the second kinematics sub-model is the linear velocity at the second time output by the first kinematics sub-model, the angular velocity at the second time output by the second kinematics sub-model, and the position at the first time, and the output is the position at the second time.

[0016] The input of the first kinematics sub-model is the linear velocity at the second time output by the first kinematics sub-model, the angular velocity at the second time output by the second kinematics sub-model, and the position at the first time, and the output is the position at the second time.

[0017] The input of the second kinematics sub-model is the linear velocity at the second time output by the first kinematics sub-model, the angular velocity at the second time output by the second kinematics sub-model, and the position at the first time, and the output is the position at the second time.

[0018] In an embodiment, the method further comprises:

[0019] Before the four neural network sub-models are respectively trained in cascade using the corresponding training sets of the four sub-models, the hyperparameters of the fixed-wing unmanned aerial vehicle simulation model are optimized based on a hybrid enhanced particle swarm optimization algorithm; the hybrid enhanced particle swarm optimization algorithm comprises:

[0020] The hyperparameter set is taken as a particle; the hyperparameter set contained in each particle is divided into a fixed-dimension set and a variable-dimension set, wherein the elements in the fixed-dimension set include an initial learning rate and a network hidden layer number; the variable-dimension set is the number of nodes in each hidden layer; the position and velocity of each particle are initialized, and the fitness of each particle is calculated;

[0021] The position and velocity of each particle are updated according to the fitness, and the individual optimal solution and the global optimal solution are updated;

[0022] The above steps are iterated until a preset stopping condition is reached; and after the iteration is completed, the optimization result of the hyperparameters of the fixed-wing unmanned aerial vehicle simulation model is obtained;

[0023] The training of the four sub-models in cascade using the corresponding training sets of the four sub-models comprises:

[0024] The four sub-models are respectively trained in cascade based on the optimized hyperparameters.

[0025] According to a second aspect of the present application, a fixed-wing unmanned aerial vehicle simulation model training device based on a cascade neural network is provided, the fixed-wing unmanned aerial vehicle simulation model comprising four cascade neural network sub-models; the device comprises:

[0026] The data acquisition module is configured to acquire a fixed-wing unmanned aerial vehicle simulation dataset, each piece of data in the fixed-wing unmanned aerial vehicle simulation dataset comprising a plurality of control parameters and a plurality of operating states of the fixed-wing unmanned aerial vehicle; wherein the plurality of operating states comprise a plurality of intermediate operating states and a plurality of final operating states.

[0027] The training set splitting module is configured to split the fixed-wing unmanned aerial vehicle simulation dataset to obtain corresponding training sets of four neural network submodels, the four neural network submodels comprising two dynamic submodels and two kinematic submodels; wherein each training sample in the training set corresponding to each dynamic submodel takes at least two control parameters and at least one operating state at a first time point as characteristic values, and takes at least one intermediate operating state at a second time point as a label value; each training sample in the training set corresponding to each kinematic submodel takes at least one final operating state at the first time point and at least one intermediate operating state at the second time point output by at least one dynamic submodel as characteristic values, and takes at least one final operating state at the second time point as a label value; the second time point is a next time point of the first time point.

[0028] The training module is configured to perform cascaded training on the four neural network submodels respectively by using the corresponding training sets of the four neural network submodels, to obtain a trained fixed-wing unmanned aerial vehicle simulation model.

[0029] In an implementation, the data acquisition module is specifically configured to control a fixed-wing unmanned aerial vehicle to perform simulation flight based on all types of control parameter combinations in an unmanned aerial vehicle simulation platform, acquire control parameters and operating states of the fixed-wing unmanned aerial vehicle during the simulation flight, and obtain a first fixed-wing unmanned aerial vehicle simulation dataset.

[0030] The continuous space of the first fixed-wing unmanned aerial vehicle simulation dataset is equally divided into n intervals, the interval with the least number of data is taken as a reference interval, and data of the reference interval corresponding to the number of data in other intervals are randomly selected as the final data of the other intervals, so as to ensure that the amount of data in each interval is consistent, thereby obtaining the fixed-wing unmanned aerial vehicle simulation dataset.

[0031] In an implementation, the control parameters comprise throttle input instructions, flap input instructions, aileron input instructions, elevator input instructions, and rudder input instructions; and the operating states comprise linear velocities, positions, angular velocities, and attitude angles; wherein the intermediate operating states comprise linear velocities and angular velocities, and the final operating states comprise positions and attitude angles.

[0032] According to a third aspect of the present application, a computer device is provided, comprising a memory and a processor, the memory storing a computer program, and the processor implements the steps of the method according to any one of the first aspect when executing the computer program.

[0033] According to a fourth aspect of the present application, a computer readable storage medium is provided, having stored thereon a computer program, wherein the computer program, when executed by a processor, implements the steps of the method according to any one of the first aspect.

[0034] Based on the above-mentioned fixed-wing unmanned aerial vehicle simulation model training method based on a cascade neural network, a cascade framework is proposed according to the physical process of the fixed-wing unmanned aerial vehicle maneuvering, and an intelligent generation framework of the fixed-wing unmanned aerial vehicle maneuvering simulation model is proposed. The fixed-wing unmanned aerial vehicle maneuvering model is decomposed into four deep neural network sub-models, each network represents a corresponding physical model and has a clear physical meaning, and can be used alone. According to different requirements, each sub-model can be combined to obtain a maneuvering model with different input and output. The cascade training method is adopted, that is, the prediction output of the trained previous network is used as the training sample feature of the next network. Compared with the traditional integrated modeling method, this method can adaptively compensate errors, reduce fitting errors, improve fitting accuracy, and effectively reduce error accumulation in the network. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 A flowchart of a fixed-wing unmanned aerial vehicle simulation model training method based on a cascade neural network in an embodiment is shown.

[0036] Figure 2 A schematic diagram of the fixed-wing unmanned aerial vehicle maneuvering model modeling process in an embodiment is shown.

[0037] Figure 3 A cascade deep neural network framework for the fixed-wing unmanned aerial vehicle maneuvering process in an embodiment is shown.

[0038] Figure 4 A schematic diagram of the particle swarm optimization neural network hyperparameter optimization in an embodiment is shown.

[0039] Figure 5 A schematic diagram of the cascade fixed-wing unmanned aerial vehicle maneuvering deep neural network model research key technology in an embodiment is shown.

[0040] Figure 6 A comparison result of the x-direction velocity in the experimental results of the neural network model evaluation in an embodiment is shown.

[0041] Figure 7 A comparison result of the y-direction velocity in the experimental results of the neural network model evaluation in an embodiment is shown.

[0042] Figure 8 A comparison result of the z-direction velocity in the experimental results of the neural network model evaluation in an embodiment is shown.

[0043] Figure 9 Comparison results of x-direction angular velocity in experimental results of evaluating neural network models in one embodiment;

[0044] Figure 10 Comparison results of y-direction angular velocity in experimental results of evaluating neural network models in one embodiment;

[0045] Figure 11 Comparison results of z-direction angular velocity in experimental results of evaluating neural network models in one embodiment;

[0046] Figure 12 Comparison results of x-direction position in experimental results of evaluating neural network models in one embodiment;

[0047] Figure 13 Comparison results of y-direction position in experimental results of evaluating neural network models in one embodiment;

[0048] Figure 14 Comparison results of z-direction position in experimental results of evaluating neural network models in one embodiment;

[0049] Figure 15 Comparison results of roll angle in experimental results of evaluating neural network models in one embodiment;

[0050] Figure 16 Comparison results of pitch angle in experimental results of evaluating neural network models in one embodiment;

[0051] Figure 17 Comparison results of yaw angle in experimental results of evaluating neural network models in one embodiment;

[0052] Figure 18 A structure schematic diagram of a fixed-wing unmanned aerial vehicle simulation model training device based on a cascade neural network in one embodiment. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0054] The technical scheme of the present application will be further described below with reference to the drawings.

[0055] An unmanned aerial vehicle (UAV) is an unmanned aerial vehicle that can fly in the air. The types and designs of UAVs are various, and can be roughly divided into two categories: rotary-wing UAVs and fixed-wing UAVs. The fixed-wing UAV is a kind of unmanned robot that flies by using fixed wings to generate lift. It is usually composed of a fuselage, wings, tail, electric motor, remote control equipment and other electronic equipment. The flight principle of the fixed-wing UAV is to generate thrust by driving the propeller with the electric motor, make the fuselage move forward, and at the same time generate lift with the wings to make the fuselage leave the ground and keep moving in the air. By adjusting the flight angle, thrust, rudder angle and other parameters, the flight direction and height can be changed. The fixed-wing UAV is widely used in agriculture, surveying and mapping, environmental monitoring, search and rescue and other fields, and has the advantages of high efficiency, low cost, flexibility, repeatability and other advantages, and has a wide development prospect in future applications. More and more researchers carry out related experimental research on fixed-wing UAVs based on unmanned system simulation platform, so it is very important to build an accurate and stable simulation model for the fixed-wing UAV to ensure the effectiveness, stability and reliability of the simulation experiment.

[0056] The maneuvering model is the basis and core of the fixed-wing simulation model. Through modeling and simulation of the maneuvering function of the fixed-wing unmanned aerial vehicle, it can effectively support the training of various onboard intelligent algorithms of unmanned aerial vehicles. Based on this, there are many open source and fine simulation platforms available, but in the actual development and application of fixed-wing unmanned aerial vehicle simulation models, there are still the following problems: The traditional fixed-wing unmanned aerial vehicle maneuvering model modeling method constructs a maneuvering mathematical model through mathematical and physical analysis, which requires rich professional knowledge and is difficult to generalize. The construction of an accurate simulation model requires a lot of time and effort, and the internal parameters of each fixed-wing unmanned aerial vehicle model are different, and the dynamic characteristics are different. Each type of fixed-wing unmanned aerial vehicle needs to be modeled separately, which is also very time-consuming. It is very difficult for non-professionals to build a fine and usable fixed-wing unmanned aerial vehicle maneuvering simulation model. With the development of artificial intelligence, more and more researchers construct fixed-wing unmanned aerial vehicle maneuvering models through deep learning methods, and using artificial intelligence technology and deep learning methods to construct simulation models has become a new method. On the one hand, intelligent modeling uses deep learning algorithms to automatically learn and build models from a large amount of data. This highly automated feature greatly reduces the burden on developers and greatly shortens the development cycle of simulation models, speeds up engineering progress, and improves production efficiency. On the other hand, intelligent modeling through deep learning can model and fit complex nonlinear relationships. Deep learning has excellent feature learning ability and strong fitting ability when dealing with nonlinear problems, and also has good generalization ability. Compared with traditional modeling methods, deep learning methods can more accurately capture the implicit patterns and rules between data, improving the prediction and generalization ability of the model. These advantages have made deep learning achieve breakthroughs in many fields and provided powerful tools and methods for solving complex problems in the real world.

[0057] Neural network models have good compatibility and can run effectively on different platforms, which can effectively solve the problem of model use difficulty. However, most current researchers construct fixed-wing unmanned aerial vehicle simulation models through integrated neural networks, with fixed input and output data types, which has certain limitations.

[0058] The inventors have found that the component-based modeling idea is a widely used method in simulation model development. It divides the system into independent components or modules, models and designs each component separately, and simulates and analyzes the behavior of the overall system. This method has key features such as modularity, interface definition, reusability, module interaction, and hierarchical structure. The component-based modeling idea can improve the maintainability, reusability, and scalability of simulation models, and effectively solve the challenges of complex system modeling. The component-based modeling idea promotes team cooperation and model sharing, providing opportunities for researchers in different fields to cooperate and exchange. It is of great significance to the design and analysis of system modeling and simulation.

[0059] In addition, in machine learning, a high-quality, large-scale, diverse, representative, and balanced dataset is crucial for the accuracy, generalization ability, and performance of the model. Good dataset quality directly affects the accuracy of the model, while the size and diversity of the dataset help improve the generalization ability of the model. In addition, the dataset should accurately describe the characteristics of the problem domain and ensure the representativeness of the data distribution, but there is currently no authoritative fixed-wing unmanned aerial vehicle flight dataset.

[0060] Based on the above problems, the application provides a fixed-wing unmanned aerial vehicle model training method based on a cascading neural network, which uses high-quality unmanned aerial vehicle data to provide data support for subsequent research. Based on the intelligent generation framework of process component-based modeling and cascading neural networks, an efficient and accurate method is provided for the development and training of simulation models for fixed-wing unmanned aerial vehicles. The component-based modeling idea for the fixed-wing unmanned aerial vehicle maneuver simulation model can provide multiple control interfaces to meet different needs. The fixed-wing unmanned aerial vehicle maneuver simulation model is modeled according to its maneuver physical process, with the dynamics model and kinematics model modeled separately, and then each sub-model is cascaded to form a cascading deep neural network framework for the fixed-wing unmanned aerial vehicle maneuver process. In addition, automatic adjustment of network hyperparameters can further improve prediction accuracy, making the simulation model more reliable.

[0061] Specifically, as shown in Figure 1 The application provides a fixed-wing unmanned aerial vehicle simulation model training method based on a cascading neural network. The fixed-wing unmanned aerial vehicle simulation model includes four cascading neural network sub-models. The method includes:

[0062] S101, obtaining a fixed-wing unmanned aerial vehicle simulation dataset, each data in the fixed-wing unmanned aerial vehicle simulation dataset including multiple control parameters and multiple operating states of the fixed-wing unmanned aerial vehicle; wherein the multiple operating states include multiple intermediate operating states and multiple final operating states;

[0063] S102, splitting the fixed-wing unmanned aerial vehicle simulation dataset to obtain corresponding training sets of four neural network sub-models, the four neural network sub-models including two dynamic sub-models and two kinematic sub-models; wherein each training sample in the training set corresponding to each dynamic sub-model takes at least two control parameters and at least one intermediate motion state at the first time as characteristic values, and takes at least one intermediate motion state at the second time as a label value; each training sample in the training set corresponding to each kinematic sub-model takes at least one final motion state at the first time and at least one intermediate motion state at the second time output by at least one dynamic sub-model as characteristic values, and takes at least one final motion state at the second time as a label value; the second time is the next time of the first time;

[0064] S103, using the training sets corresponding to the four neural network sub-models to respectively perform cascade training on the four neural network sub-models to obtain a trained fixed-wing unmanned aerial vehicle simulation model.

[0065] Most researchers currently use integrated modeling methods, such as inputting throttle and outputting position and attitude changes. However, the above method proposed by the present application uses a cascade training method, that is, the predicted output of the trained previous network is used as the training sample feature of the next network. Compared with the traditional integrated modeling method, this method can perform adaptive error compensation, reduce fitting error, improve fitting accuracy, and effectively reduce error accumulation in the network. In addition, according to the physical process of fixed-wing unmanned aerial vehicle maneuvering, a cascade framework is proposed, and an intelligent generation framework of a fixed-wing unmanned aerial vehicle maneuvering simulation model is proposed. The fixed-wing unmanned aerial vehicle maneuvering model is decomposed into four deep neural network sub-models, each network represents the corresponding physical model and has a clear physical meaning. Each network can be used alone or combined according to different requirements to obtain different input and output maneuvering models.

[0066] The structure of the fixed-wing unmanned aerial vehicle simulation model proposed in the present application will be described below.

[0067] As Figure 2As shown, by analyzing the maneuvering process of the fixed-wing UAV in detail, the entire fixed-wing UAV maneuvering model is decomposed into a fixed-wing UAV dynamics model and a fixed-wing UAV kinematics model, and five control parameters (throttle, flap, aileron, elevator, rudder) are used for modeling. The throttle input instruction is used to control the thrust output of the engine or motor to control the speed of the aircraft. The flap input instruction is used to increase the lift and drag of the aircraft during flight. The aileron input instruction is used for roll control, which controls the roll movement of the aircraft by changing the position of the aileron. The elevator input instruction is used for pitch control, which controls the pitch movement of the aircraft by changing the position of the elevator. The rudder input instruction is used for yaw control, which controls the yaw movement of the aircraft by changing the position of the rudder.

[0068] The input of the fixed-wing UAV dynamics model is the current time control quantity, i.e. the throttle, flap, aileron, elevator, rudder input instruction, the linear velocity of the UAV in the world coordinate system at the current time and the angular velocity in the body coordinate system, and the output is the linear velocity of the UAV in the world coordinate system at the next time and the angular velocity in the body coordinate system. The input-output relationship is represented as follows:

[0069] (1)

[0070] (2)

[0071] wherein is the linear velocity of the fixed-wing UAV in the world coordinate system at the current time, is the angular velocity of the fixed-wing UAV in the body coordinate system at the current time, is the throttle input instruction, is the throttle input instruction at the current time, is the flap input instruction, is the flap input instruction at the current time, is the aileron input instruction, is the aileron input instruction at the current time, is the elevator input instruction, is the elevator input instruction at the current time, is the rudder input instruction, is the rudder input instruction at the current time.

[0072] ​​​​​​​The input of the fixed-wing UAV kinematic model is the linear velocity in the world coordinate system at the next moment, the angular velocity in the body coordinate system at the next moment, the position in the world coordinate system at the current moment, and the attitude angle in the body coordinate system at the current moment, obtained from the UAV dynamics model. The output is the position and attitude angle at the next moment. The input and output relationship is expressed as follows:

[0073] (3)

[0074] (4)

[0075] in yes The position of the fixed-wing drone in the world coordinate system at any moment, yes The position of the fixed-wing drone in the world coordinate system at any moment, yes The attitude angle of the fixed-wing UAV in the world coordinate system at all times, yes The attitude angle of the fixed-wing UAV in the world coordinate system at all times, yes The angular velocity of the fixed-wing UAV in the body coordinate system at any moment, yes The linear velocity of the fixed-wing UAV in the world coordinate system at any moment.

[0076] According to the above fixed-wing UAV maneuvering principle and modeling process, a cascaded deep neural network framework for the fixed-wing UAV maneuvering process is proposed, such as Figure 3 In this framework, D is a delay block, representing the data at the current moment, and NN1 to NN4 represent the neural networks corresponding to formulas (1)-(4), respectively.

[0077] Through the above analysis of the physical process of drone maneuvers, we can determine the input and output data types of each component. We collect 60 pieces of drone model data per second, including control parameters (throttle, flaps, ailerons, elevator, and rudder commands) and motion states (speed, position, angular velocity, and attitude). After filtering and normalizing the raw data, we obtain four sub-model training sets, test sets, and validation sets corresponding to the four sub-models, which serve as training samples for the deep neural network. The input and output data types and corresponding physical formulas for each network are shown in Table 1.

[0078] ;

[0079] Table 1 Input and output data and physical formulas of each network

[0080] The control parameters include throttle input instructions, flap input instructions, aileron input instructions, elevator input instructions, and rudder input instructions. Figure 3 The intermediate motion state includes linear velocity and angular velocity, and the final motion state includes position and attitude angle.

[0081] As shown in Figure 3 The two dynamic sub-models include a first dynamic sub-model NN1 and a second dynamic sub-model NN2, and the two kinematic sub-models include a first kinematic sub-model NN3 and a second kinematic sub-model NN4.

[0082] The input of the first dynamic sub-model is the throttle input instructions and the flap input instructions at the second time and the linear velocity at the first time, and the output is the linear velocity at the second time, i.e., the input-output relationship is .

[0083] The input of the second dynamic sub-model is the aileron input instructions, the elevator input instructions, the rudder input instructions at the second time, and the angular velocity at the first time, and the output is the angular velocity at the second time, i.e., the input-output relationship is .

[0084] The input of the first kinematic sub-model is the linear velocity at the second time output by the first dynamic sub-model, the angular velocity at the second time output by the second dynamic sub-model, and the position at the first time, and the output is the position at the second time, i.e., the input-output relationship is .

[0085] The input of the second kinematic sub-model is the angular velocity at the second time output by the second dynamic sub-model and the attitude angle at the first time, and the output is the attitude angle at the second time, i.e., the input-output relationship is .

[0086] Since the input data of the first kinematic sub-model includes the data output by the first dynamic sub-model and the data output by the second dynamic sub-model, and the input data of the second kinematic sub-model includes the data output by the second dynamic sub-model, the cascade training specifically includes training the first dynamic sub-model and the second dynamic sub-model using the training set first, then taking the outputs of the trained first dynamic sub-model and the second dynamic sub-model as part of the features of the training samples of the first kinematic sub-model and the second kinematic sub-model, and then training the first kinematic sub-model and the second kinematic sub-model.

[0087] It can be understood that the application decomposes the maneuvering model of the fixed-wing unmanned aerial vehicle into four deep neural network sub-models, each network represents a corresponding physical model and has a clear physical meaning, and thus can be used alone. The above combination is only one of the combination modes, and those skilled in the art can also combine each sub-model according to different requirements to obtain a maneuvering model with different inputs and outputs.

[0088] In an embodiment, a fixed-wing unmanned aerial vehicle simulation data set is acquired, including:

[0089] In the unmanned aerial vehicle simulation platform, the fixed-wing unmanned aerial vehicle is controlled to perform simulation flight based on all types of control parameter combinations, control parameters and operation states of the fixed-wing unmanned aerial vehicle during simulation flight are acquired, and a first fixed-wing unmanned aerial vehicle simulation data set is obtained.

[0090] The continuous space of the first fixed-wing unmanned aerial vehicle simulation data set is equally divided into n intervals, the interval with the least number of data is taken as a reference interval, and the data corresponding to the number of data in the reference interval is randomly selected as the final data of the other intervals to ensure that the amount of data in each interval is consistent, so that the fixed-wing unmanned aerial vehicle simulation data set is obtained.

[0091] Specifically, the present application analyzes the characteristics of unmanned aerial vehicle flight data from six dimensions (integrity, timeliness, accuracy, standardization, unbiasedness, and ease of use) in combination with the indicators of the machine learning data set, and proposes the following fixed-wing unmanned aerial vehicle data standards.

[0092] ①Integrity: Integrity is a key quality dimension, reflecting the degree to which the subject data associated with the entity has all expected attributes and associated strength values in a given environment.

[0093] ②Timeliness: The degree to which data has reliable persistence in a given situation to ensure that data keeps pace with the times.

[0094] ③Accuracy: The degree to which the attributes of the data correctly represent the true value of the relevant attributes of a certain concept or event in a specific environment.

[0095] ④Standardization: The degree to which data in a given environment meets data standards, data models, metadata, or authoritative reference data.

[0096] ⑤Unbiasedness: The degree of distribution of data categories or features in a given environment.

[0097] ⑥Ease of use: The degree to which the data set can be quickly used in a specific environment.

[0098] On the basis of the above standards, a comprehensive data generation method is proposed.

[0099] Different flight data of the unmanned aerial vehicle has different characteristics, and more comprehensive data is collected as much as possible in a flight trajectory to improve the data collection efficiency. The characteristics of different flight data of the fixed-wing unmanned aerial vehicle are analyzed, different unmanned aerial vehicle flight control modes are used to generate flight data meeting the data standard, the collection platform RflySim is generated, and the programming language is Python.

[0100] The RflySim platform is a tool chain, and contains many software used for unmanned system development, such as a Python 3.8 environment, a MATLAB / PSP toolbox, FilghtGear, QGroundControl, and RflySim3D and CopterSim independently developed for unmanned systems. The process of using the platform to simulate the fixed-wing unmanned aerial vehicle and obtain flight data is as follows:

[0101] 1. Run SITLRun.bat, set the number of unmanned aerial vehicles, use Pixhawk as the control module, UE4 as the unmanned aerial vehicle simulation environment, QGroundControl as the unmanned aerial vehicle ground station, configure the unmanned aerial vehicle, set the unmanned aerial vehicle model to be simulated, the starting position and attitude in CopterSim, initialize a fixed-wing unmanned aerial vehicle, correctly connect the flight control, and start to control autonomous flight.

[0102] 2. Configure the corresponding Python 3.8 environment, and write appropriate control algorithms for comprehensive data collection.

[0103] 3. The data sampling frequency is 60 Hz, the control state and motion state data in the flight process are obtained, each data includes the throttle, aileron, elevator, rudder command data obtained by each motor, the linear velocity of the fixed-wing unmanned aerial vehicle, the angular velocity, the position and the attitude angle. A total of 100000 pieces of flight data of the fixed-wing unmanned aerial vehicle are collected.

[0104] The pseudo code for controlling the fixed-wing unmanned aerial vehicle through the Rflysim interface is as follows:

[0105]

[0106] On the basis of comprehensive data generation, the data are further screened in a non-biased manner.

[0107] The application proposes a data screening method based on non-bias to output a uniformly distributed data set. The continuous space is equally divided into n intervals, and the interval with the least data is taken as the reference, and the number of data in other intervals is randomly selected as the final data to ensure the consistency of the data in each interval. Uniform data distribution can help the model avoid over-reliance on certain specific situations, thereby improving the generalization ability of the model. If there are too many or too few samples of certain specific situations in the data set, the model may have higher or lower weights for these situations, leading to biased prediction results for other situations.

[0108] The algorithm pseudo code is as follows:

[0109]

[0110] Based on the above scheme, a set of fixed-wing unmanned aerial vehicle data generation based on comprehensiveness and data screening based on non-bias is proposed to obtain a large amount of high-quality unmanned aerial vehicle flight data, comprehensively covering various flight states of unmanned aerial vehicle maneuvering behavior, which is beneficial to build an unmanned aerial vehicle maneuvering simulation model and improve the authenticity and credibility of the simulation.

[0111] In order to make the training of the deep neural network model more effective and accurate, it is necessary to determine the hyperparameters of the deep neural network model before model training.

[0112] In one embodiment, the application proposes to optimize the hyperparameters of the fixed-wing unmanned aerial vehicle simulation model based on a hybrid enhanced particle swarm optimization algorithm before cascading training of the four neural network sub-models using the corresponding training sets of the four neural network sub-models; the hybrid enhanced particle swarm optimization algorithm includes:

[0113] The hyperparameter set is taken as a particle; the hyperparameter set contained in each particle is divided into a fixed dimension set and a variable dimension set, wherein the elements in the fixed dimension set include the initial learning rate and the network hidden layer number; the variable dimension set is the number of nodes in each hidden layer;

[0114] Initialize the position and speed of each particle, and calculate the fitness of each particle;

[0115] Update the position and speed of each particle according to the fitness, and update the individual optimal solution and the global optimal solution;

[0116] Loop the above steps until the preset stopping condition is reached;

[0117] After the iteration is completed, the optimization result of the hyperparameters of the fixed-wing unmanned aerial vehicle simulation model is obtained.

[0118] The corresponding training sets of the four sub-models are used to train the four sub-models in cascade, including:

[0119] Based on the optimized hyperparameters, cascade training is performed on the four sub-models, and the trained model is tested based on the test set. If the test passes, the model that passes the test is verified based on the validation set. If the test fails, training continues. Since the training samples of each sub-model include features and labels, supervised learning can be used for training.

[0120] The specific algorithm flow of the particle swarm algorithm can refer to related technologies, that is, the calculation of fitness, updating of optimal solution, and entire iteration process in the algorithm can refer to related technologies of the particle swarm algorithm, which will not be described here. The focus of this proposal is to propose the variability of dimensions, that is, to propose two sets of fixed dimension set and variable dimension set, that is, to use the particle swarm algorithm based on the two sets for calculation. The hybrid enhanced particle swarm optimization algorithm proposed in this application will be described in detail below.

[0121] In order to overcome the problem of unchangeable solution dimension in the standard particle swarm algorithm, a new method named hybrid enhanced particle swarm algorithm is proposed. This method realizes the variability of solution dimension by using the relationship between the number of hidden layers and the solution dimension. The particle solution is divided into two sets, one is the fixed dimension set , the other is the variable dimension set , which can be represented as:

[0122] (5)

[0123] The elements in the fixed dimension set have an initial learning rate and the number of network hidden layers . The fixed dimension set can be represented as:

[0124] (6)

[0125] The elements in the variable dimension set are the number of nodes in each hidden layer . The variable dimension set can be represented as:

[0126] (7)

[0127] where is the dimension of the variable set , and the number of network hidden layers in the fixed dimension set determines the dimension of the variable dimension set , that is:

[0128] (8) ​

[0129] The pseudo code of the hybrid enhanced particle swarm optimization algorithm is as follows:

[0130]

[0131] The calculation process of the particle swarm optimization algorithm for optimizing the hyperparameters of the neural network is as shown in the figure Figure 4 Since there is inconsistency between the state of the unmanned aerial vehicle and the control quantity in the collected flight data, that is, the current state and the control quantity of a piece of collected flight data may not be at the same time, a long short-term (LSTM) neural network is added to eliminate this error, that is, PSO-LSTM. The algorithm calculation steps are as follows: first step: data processing. After data normalization of the original data set, the data set is divided into a training set, a test set and a validation set. Second step: particle swarm optimization algorithm for optimizing hyperparameters. Third step: model evaluation. The optimal neural network model is evaluated through the validation set, and MSE is used as the evaluation index. Although the training time of a single neural network is relatively short, when the particle swarm optimization algorithm is used to optimize the neural network, multiple neural networks need to be trained according to the number of particles in each iteration, and in the framework proposed in this paper, a total of 4 neural network models are needed, so a lot of time is needed. In this paper, the particle swarm optimization algorithm is suitable for parallel distributed computing, so distributed computing of the network can be realized by using a computer cluster.

[0132] The pseudo code of the parallel particle swarm optimization algorithm for optimizing the hyperparameters of the neural network is as follows:

[0133]

[0134] Based on the hybrid enhanced particle swarm optimization algorithm proposed in the application. By simulating the search and movement process of particles in the solution space, the optimal solution of the neural network is found, and the optimal hyperparameters of the deep neural network are determined. In the generation of the maneuvering simulation model, the application has better performance and improves the performance of the neural network model.

[0135] In a specific implementation manner, the flight data used in the application is generated by an Rflysim simulation platform published by a reliable flight control group of Beihang University, a fixed-wing unmanned aerial vehicle is configured by using Pixhawk as a control module and QGroundControl as a ground station of the unmanned aerial vehicle, and Unreal Engine 4 is used as a simulation environment of the unmanned aerial vehicle. In the experiment, various flight data of the unmanned aerial vehicle is obtained from a data storage interface of the Rflysim, and the data sampling frequency is 60 Hz. Each piece of collected data includes throttle, flap, aileron, elevator, rudder command data obtained by each motor, linear velocity of the fixed-wing unmanned aerial vehicle, angular velocity, position and attitude angle. A total of 100,000 pieces of flight data of the fixed-wing unmanned aerial vehicle are collected. The specific implementation process is as follows:

[0136] 1. The fixed-wing UAV maneuver model is divided into four deep neural network sub-models. The key technologies of the cascaded fixed-wing UAV maneuver deep neural network model are shown in Figure 5 ;

[0137] 2. Follow Figure 3 The connection method shown in the network combines the various sub-models to determine the input and output of different maneuver models;

[0138] 3. Data processing: normalize the original data set used to train the simulation model and divide it into training set, test set and validation set;

[0139] 4. Use the improved particle swarm algorithm to optimize hyperparameters;

[0140] 5. Using the cascade training method, the prediction output of the previous trained network is used as part of the training sample features of the next network. Select appropriate evaluation indicators to evaluate the optimal neural network model. The experimental results are shown in Figures 6-17 , Figures 6-17 The comparison results between the actual value and the prediction are shown ( Figures 6-8 The speed comparison results are shown; Figures 9-11 The comparison results of angular velocity are shown; Figures 12-14 The comparison results of the positions are shown; Figures 15-17 shows the comparison results of attitude angles).

[0141] like Figure 18 As shown, corresponding to the above-mentioned fixed-wing UAV simulation model training method based on cascade neural network, the present application also proposes a fixed-wing UAV simulation model training device based on cascade neural network, wherein the fixed-wing UAV simulation model includes four cascaded neural network sub-models; the device includes:

[0142] The data acquisition module 710 is configured to acquire a fixed-wing UAV simulation data set, wherein each data item in the fixed-wing UAV simulation data set includes multiple control parameters and multiple operating states of the fixed-wing UAV; wherein the multiple motion states include multiple intermediate motion states and multiple final motion states;

[0143] The training set splitting module 720 is configured to split the fixed-wing unmanned aerial vehicle simulation dataset to obtain corresponding training sets of four neural network sub-models, the four neural network sub-models including two dynamic sub-models and two kinematic sub-models; wherein each training sample in the training set corresponding to each dynamic sub-model takes at least two control parameters and at least one motion state at a first time as characteristic values, and takes at least one intermediate motion state at a second time as a label value; each training sample in the training set corresponding to each kinematic sub-model takes at least one final motion state at the first time and at least one intermediate motion state at the second time output by the at least one dynamic sub-model as characteristic values, and takes at least one final motion state at the second time as a label value; the second time is a next time of the first time;

[0144] The training module 730 is configured to perform cascaded training on the four neural network sub-models respectively by using the corresponding training sets of the four neural network sub-models, to obtain a trained fixed-wing unmanned aerial vehicle simulation model.

[0145] In an embodiment, the data acquisition module 710 is specifically configured to control a fixed-wing unmanned aerial vehicle to perform simulation flight based on all types of control parameter combinations in an unmanned aerial vehicle simulation platform, acquire control parameters and operation states of the fixed-wing unmanned aerial vehicle during simulation flight, and obtain a first fixed-wing unmanned aerial vehicle simulation dataset.

[0146] The continuous space of the first fixed-wing unmanned aerial vehicle simulation dataset is equally divided into n intervals, the interval with the least number of data is taken as a reference interval, and data of the reference interval corresponding to the number of data in other intervals are randomly selected as final data of the other intervals, so as to ensure that the amount of data in each interval is consistent, thereby obtaining the fixed-wing unmanned aerial vehicle simulation dataset.

[0147] In an embodiment, the control parameters include throttle input instructions, flap input instructions, aileron input instructions, elevator input instructions, and rudder input instructions; the motion states include linear velocities, positions, angular velocities, and attitude angles; wherein the intermediate motion states include linear velocities and angular velocities, and the final motion states include positions and attitude angles.

[0148] In an embodiment, the two dynamic sub-models include a first dynamic sub-model and a second dynamic sub-model; and the two kinematic sub-models include a first kinematic sub-model and a second kinematic sub-model.

[0149] The input of the first dynamic sub-model is a throttle input instruction at the second time, a flap input instruction at the second time, and a linear velocity at the first time, and the output is a linear velocity at the second time.

[0150] The input of the second kinematics sub-model is the angular velocity at the second time moment output by the second dynamics sub-model and the attitude angle at the first time moment, and the output is the attitude angle at the second time moment.

[0151] The input of the first kinematics sub-model is the linear velocity at the second time moment output by the first dynamics sub-model, the angular velocity at the second time moment output by the second dynamics sub-model and the position at the first time moment, and the output is the position at the second time moment.

[0152] The input of the second kinematics sub-model is the angular velocity at the second time moment output by the second dynamics sub-model and the attitude angle at the first time moment, and the output is the attitude angle at the second time moment.

[0153] In an implementation, the training module 730 is further configured to, before performing the cascaded training on the four neural network sub-models respectively by using the corresponding training sets of the four neural network sub-models, optimize the hyperparameters of the fixed-wing UAV simulation model based on a hybrid enhanced particle swarm optimization algorithm; the hybrid enhanced particle swarm optimization algorithm comprises:

[0154] The hyperparameter set is taken as a particle; the hyperparameter set contained in each particle is divided into a fixed-dimension set and a variable-dimension set, wherein the elements in the fixed-dimension set include an initial learning rate and a network hidden layer number; the variable-dimension set is a number of hidden layer nodes;

[0155] The position and the speed of each particle are initialized, and the fitness of each particle is calculated;

[0156] The position and the speed of each particle are updated according to the fitness, and the individual optimal solution and the global optimal solution are updated;

[0157] The above steps are iterated until a preset stop condition is reached;

[0158] After the iteration is completed, the optimization result of the hyperparameters of the fixed-wing UAV simulation model is obtained;

[0159] The four sub-models are trained in a cascaded manner based on the optimized hyperparameters.

[0160] In one embodiment, a computer device is provided, which can be a server. The computer device comprises a processor, a memory, a network interface and a database connected by a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data. The network interface of the computer device is used to communicate with the external terminal through the network connection. The computer program is executed by the processor to implement a fixed-wing unmanned aerial vehicle simulation model training method based on a cascade neural network.

[0161] In one embodiment, a computer device is provided, which comprises a memory and a processor, and the memory stores a computer program, and the processor executes the computer program to implement the steps of the method in any of the above embodiments.

[0162] In one embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the steps of the method in any of the above embodiments.

[0163] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM) and memory bus dynamic RAM (RDRAM) and the like.

[0164] Any combination of the technical features in the above embodiments can be made, and for the sake of brevity, not all possible combinations are described above, however, as long as the combination of the technical features does not exist in contradiction, it shall be considered within the scope of the present disclosure.

[0165] The above embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it shall not be understood as a limitation on the patent scope of the present application. It shall be pointed out that, for ordinary skilled persons in the art, several modifications and improvements can be made without departing from the concept of the present application, and these shall be within the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A method for training a fixed-wing unmanned aerial vehicle simulation model based on a cascaded neural network, characterized in that the fixed-wing unmanned aerial vehicle simulation model comprises four cascaded neural network sub-models; the method comprises: obtaining a fixed-wing unmanned aerial vehicle simulation dataset, each piece of data in the fixed-wing unmanned aerial vehicle simulation dataset comprising a plurality of control parameters and a plurality of operating states of the fixed-wing unmanned aerial vehicle; wherein the plurality of operating states comprise a plurality of intermediate operating states and a plurality of final operating states; splitting the fixed-wing unmanned aerial vehicle simulation dataset to obtain corresponding training sets of the four neural network sub-models, the four neural network sub-models comprising two dynamics sub-models and two kinematics sub-models; wherein each training sample in the training set corresponding to each dynamics sub-model takes at least two control parameters and at least one intermediate operating state at a first time as characteristic values, and at least one intermediate operating state at a second time as a label value; each training sample in the training set corresponding to each kinematics sub-model takes at least one final operating state at the first time and at least one intermediate operating state at the second time output by at least one dynamics sub-model as characteristic values, and at least one final operating state at the second time as a label value; the second time is the next time of the first time; performing cascaded training on the four neural network sub-models respectively using the corresponding training sets of the four neural network sub-models to obtain a trained fixed-wing unmanned aerial vehicle simulation model. The obtaining of the fixed-wing unmanned aerial vehicle simulation dataset comprises:

2. The method of claim 1, wherein, controlling the fixed-wing unmanned aerial vehicle to perform simulation flight based on all types of control parameter combinations in an unmanned aerial vehicle simulation platform, obtaining the control parameters and operating states during the simulation flight of the fixed-wing unmanned aerial vehicle, and obtaining a first fixed-wing unmanned aerial vehicle simulation dataset; equally dividing the continuous space of the first fixed-wing unmanned aerial vehicle simulation dataset into n intervals, taking the interval with the least number of data as a reference interval, and randomly selecting data corresponding to the number of data in the reference interval as the final data of other intervals to obtain the fixed-wing unmanned aerial vehicle simulation dataset.

3. The method of claim 1 or 2, characterized in that: the control parameters comprise throttle input instructions, flap input instructions, aileron input instructions, elevator input instructions, and rudder input instructions; and the operating states comprise linear velocities, positions, angular velocities, and attitude angles; wherein the intermediate operating states comprise linear velocities and angular velocities, and the final operating states comprise positions and attitude angles. the two dynamics sub-models comprise a first dynamics sub-model and a second dynamics sub-model; and the two kinematics sub-models comprise a first kinematics sub-model and a second kinematics sub-model; 4. The method of claim 1, wherein, the input of the first dynamics sub-model is the throttle input instruction at the second time, the flap input instruction at the second time, and the linear velocity at the first time; and the output of the first dynamics sub-model is the linear velocity at the second time; the input of the second dynamics sub-model is the aileron input instruction at the second time, the elevator input instruction at the second time, the rudder input instruction at the second time, and the angular velocity at the first time; and the output of the second dynamics sub-model is the angular velocity at the second time. ​ The input of the first kinematics sub-model is the linear velocity at the second time point of the first dynamics sub-model output, the angular velocity at the second time point of the second dynamics sub-model output, and the position at the first time point, and the output is the position at the second time point; The input of the second kinematics sub-model is the angular velocity at the second time point of the second dynamics sub-model output and the attitude angle at the first time point, and the output is the attitude angle at the second time point.

5. The method of claim 1, wherein, The method further comprises: Before the four neural network sub-models are respectively trained in a cascaded manner using the corresponding training sets of the four neural network sub-models, the hyperparameters of the fixed-wing UAV simulation model are optimized based on a hybrid enhanced particle swarm optimization algorithm; the hybrid enhanced particle swarm optimization algorithm comprises: The hyperparameter set is taken as a particle; the hyperparameter set contained in each particle is divided into a fixed dimension set and a variable dimension set, wherein the elements in the fixed dimension set include an initial learning rate and a network hidden layer number; the variable dimension set is the number of nodes in each hidden layer; The position and speed of each particle are initialized, and the fitness of each particle is calculated; The position and speed of each particle are updated according to the fitness, and the individual optimal solution and the global optimal solution are updated; The above steps are iterated until a preset stop condition is reached; After the iteration is completed, the optimization result of the hyperparameters of the fixed-wing UAV simulation model is obtained; The four sub-models are respectively trained in a cascaded manner based on the optimized hyperparameters.

6. A fixed-wing UAV simulation model training device based on a cascaded neural network, characterized in that: The fixed-wing UAV simulation model comprises four cascaded neural network sub-models; the device comprises: A data acquisition module is configured to acquire a fixed-wing UAV simulation dataset, wherein each piece of data in the fixed-wing UAV simulation dataset comprises a plurality of control parameters and a plurality of operating states of the fixed-wing UAV; the plurality of operating states comprise a plurality of intermediate operating states and a plurality of final operating states; A training set splitting module is configured to split the fixed-wing UAV simulation dataset to obtain corresponding training sets of four neural network sub-models, wherein the four neural network sub-models comprise two dynamics sub-models and two kinematics sub-models; each training sample in the training set corresponding to each dynamics sub-model takes at least two control parameters and at least one operating state at a first time point as characteristic values, and takes at least one intermediate operating state at a second time point as a label value; each training sample in the training set corresponding to each kinematics sub-model takes at least one final operating state at the first time point and at least one intermediate operating state at the second time point output by at least one dynamics sub-model as characteristic values, and takes at least one final operating state at the second time point as a label value; the second time point is the next time point of the first time point; A training module is configured to train the four neural network sub-models in a cascaded manner using the corresponding training sets of the four neural network sub-models, and obtain a trained fixed-wing UAV simulation model.

7. The device according to claim 6, characterized in that, ​ The data acquisition module is specifically configured to control a fixed-wing unmanned aerial vehicle to perform simulation flight based on all types of control parameter combinations in an unmanned aerial vehicle simulation platform, acquire control parameters and operation states of the fixed-wing unmanned aerial vehicle during simulation flight, and obtain a first fixed-wing unmanned aerial vehicle simulation data set. The continuous space of the first fixed-wing unmanned aerial vehicle simulation data set is equally divided into n intervals, the interval with the least number of data is taken as a reference interval, and data of the reference interval corresponding to the number of data of other intervals are randomly selected as the final data of other intervals to ensure that the data quantity in each interval is consistent, thereby obtaining the fixed-wing unmanned aerial vehicle simulation data set.

8. The apparatus of claim 6 or 7, wherein The control parameters include throttle input instructions, flap input instructions, aileron input instructions, elevator input instructions, and rudder input instructions; and the motion states include linear velocities, positions, angular velocities, and attitude angles; wherein the intermediate motion states include linear velocities and angular velocities, and the final motion states include positions and attitude angles. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to implement the steps of the method of any one of claims 1 to 5.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 5.

Images