Cascade neural network-based fixed-wing unmanned aerial vehicle simulation model training method

Abstract

The invention discloses a cascaded neural network-based fixed-wing unmanned aerial vehicle simulation model training method. A fixed-wing unmanned aerial vehicle simulation model comprises four cascaded neural network sub-models; the method comprises the following steps: acquiring a fixed-wing unmanned aerial vehicle simulation data set, wherein each piece of data in the fixed-wing unmanned aerial vehicle simulation data set comprises various control parameters and various operation states of the fixed-wing unmanned aerial vehicle; wherein the multiple motion states comprise multiple intermediate motion states and multiple final motion states; splitting the fixed-wing unmanned aerial vehicle simulation data set to obtain training sets corresponding to four neural network sub-models, wherein the four neural network sub-models comprise two dynamical sub-models and two kinematic sub-models; and performing cascade training on the four neural network sub-models by using the training sets corresponding to the four neural network sub-models to obtain a trained model. According to the method, adaptive error compensation can be carried out, the fitting error is reduced, the fitting precision is improved, and error accumulation in the network is effectively reduced.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular, to a method for training a simulation model of a fixed-wing unmanned aerial vehicle based on a cascaded neural network. Background Art

[0002] An unmanned aerial vehicle (UAV) is an aircraft that can fly in the air and is unmanned. There are various types and designs of UAVs, which can be roughly divided into two categories: rotary-wing UAVs and fixed-wing UAVs. Among them, a fixed-wing UAV is an unmanned robot that generates lift using fixed wings. It usually consists of a fuselage, wings, a tail, an electric motor, remote control equipment, and other electronic devices. The flight principle of a fixed-wing UAV is that the electric motor drives the propeller to generate thrust, making the fuselage move forward. At the same time, lift is generated using the wings, making the fuselage leave the ground and stay in motion in the air. By adjusting parameters such as the flight angle, thrust, and rudder surface angle, its flight direction and altitude can be changed. Fixed-wing UAVs are widely used in agriculture, surveying and mapping, environmental monitoring, search and rescue, etc. They have the advantages of high efficiency, low cost, flexibility, repeatability, etc., and have broad development prospects in future applications. More and more researchers conduct related experimental studies on fixed-wing UAVs based on the unmanned system simulation platform. Therefore, it is crucial to build an accurate and stable simulation model for fixed-wing UAVs to ensure the effectiveness, robustness, and credibility of simulation experiments.

[0003] Neural network models have good compatibility and can operate effectively on different platforms, which can effectively solve the problem of difficult model usage. However, currently, most researchers construct the simulation model of a fixed-wing UAV through an integrated neural network, and this method has certain limitations. Summary of the Invention

[0004] Based on this, it is necessary to provide a method for training a simulation model of a fixed-wing UAV based on a cascaded neural network for the above technical problems.

[0005] According to the first aspect of the present application, a method for training a simulation model of a fixed-wing UAV based on a cascaded neural network is proposed. The simulation model of the fixed-wing UAV includes four cascaded neural network sub-models; the method includes: Obtain a simulation data set of a fixed-wing UAV. Each piece of data in the simulation data set of the fixed-wing UAV includes various control parameters and various operating states of the fixed-wing UAV; among them, the various motion states include various intermediate motion states and various final motion states; The fixed-wing UAV simulation dataset is split to obtain training sets corresponding to four neural network sub-models, and the four neural network sub-models include two dynamics sub-models and two kinematics sub-models; wherein, each training sample in the training set corresponding to each dynamics sub-model uses at least two control parameters and at least one motion state at the first moment as eigenvalue, and uses at least one intermediate operation state at the second moment as label value; each training sample in the training set corresponding to each kinematics sub-model uses at least one final motion state at the first moment and at least one intermediate motion state at the second moment output by at least one dynamics sub-model as eigenvalue, and uses at least one final motion state at the second moment as label value; the second moment is the next moment of the first moment. The corresponding training sets of the four neural network sub-models are respectively used to perform cascaded training on the four neural network sub-models to obtain a trained fixed-wing UAV simulation model.

[0006] In one implementation, the obtaining of the fixed-wing UAV simulation dataset includes: In a UAV simulation platform, the fixed-wing UAV is controlled to perform simulation flight based on all types of control parameter combinations, and the control parameters and operation states during the simulation flight of the fixed-wing UAV are obtained to obtain a first fixed-wing UAV simulation dataset. The continuous space of the first fixed-wing UAV simulation dataset is equally divided into n intervals. Taking the interval with the least number of data as the reference interval, the data with the number of data corresponding to the reference interval is randomly selected from other intervals as the final data of other intervals to ensure that the amount of data in each interval is the same, thereby obtaining the fixed-wing UAV simulation dataset.

[0007] In one implementation, the control parameters include: throttle input command, flap input command, aileron input command, elevator input command, rudder input command; the motion states include: linear velocity, position, angular velocity, attitude angle; wherein, the intermediate motion states include linear velocity and angular velocity, and the final motion states include position and attitude angle.

[0008] In one implementation, the two dynamics sub-models include a first dynamics sub-model and a second dynamics sub-model; the two kinematics sub-models include a first kinematics sub-model and a second kinematics sub-model. The input of the first dynamics sub-model is the throttle input command, flap input command at the second moment, and the linear velocity at the first moment, and the output is the linear velocity at the second moment. The input of the second dynamics sub-model is the aileron input command, elevator input command, rudder input command at the second moment, and the angular velocity at the first moment, and the output is the angular velocity at the second moment. The input of the first kinematic sub-model is the linear velocity at the second moment output by the first dynamic sub-model, the angular velocity at the second moment output by the second dynamic sub-model, and the position at the first moment, and the output is the position at the second moment; The input of the second kinematic sub-model is the angular velocity at the second moment output by the second dynamic sub-model and the attitude angle at the first moment, and the output is the attitude angle at the second moment.

[0009] In one implementation, the method further includes: Before cascaded training of the four neural network sub-models using the corresponding training sets of the four neural network sub-models respectively, optimize the hyperparameters of the fixed-wing UAV simulation model based on the hybrid enhanced particle swarm optimization algorithm; the hybrid enhanced particle swarm optimization algorithm includes: Regarding the hyperparameter set as a particle; the hyperparameter set included in each particle is divided into two parts, a fixed dimension set and a variable dimension set. Among them, the elements in the fixed dimension set include the initial learning rate and the number of network hidden layers; the variable dimension set is the number of nodes in each hidden layer; Initialize the position and velocity of each particle, and calculate the fitness of each particle; Update the position and velocity of each particle according to the fitness, and update the individual optimal solution and the global optimal solution; Loop and iterate the above steps until a preset stop condition is reached; After the iteration is completed, obtain the optimization result of the hyperparameters of the fixed-wing UAV simulation model; The cascaded training of the four sub-models using the corresponding training sets of the four sub-models respectively includes: Perform cascaded training on the four sub-models respectively based on the optimized hyperparameters.

[0010] According to the second aspect of the present application, a training device for a fixed-wing UAV simulation model based on a cascaded neural network is proposed. The fixed-wing UAV simulation model includes four cascaded neural network sub-models; the device includes: A data acquisition module, configured to acquire a fixed-wing UAV simulation data set. Each piece of data in the fixed-wing UAV simulation data set includes various control parameters and various operating states of the fixed-wing UAV; among them, the various motion states include various intermediate motion states and various final motion states; A training set splitting module, which is used to split the fixed-wing UAV simulation data set to obtain training sets corresponding to four neural network sub-models. The four neural network sub-models include two dynamics sub-models and two kinematics sub-models. Among them, each training sample in the training set corresponding to each dynamics sub-model uses at least two control parameters and at least one motion state at the first moment as eigenvalue, and at least one intermediate operation state at the second moment as label value. Each training sample in the training set corresponding to each kinematics sub-model uses at least one final motion state at the first moment and at least one intermediate motion state at the second moment output by at least one dynamics sub-model as eigenvalue, and at least one final motion state at the second moment as label value. The second moment is the next moment of the first moment. A training module, which is used to perform cascaded training on the four neural network sub-models respectively by using the training sets corresponding to the four neural network sub-models to obtain a trained fixed-wing UAV simulation model.

[0011] In one implementation manner, the data acquisition module is specifically configured to, in a UAV simulation platform, control a fixed-wing UAV to perform simulation flight based on all types of control parameter combinations, acquire the control parameters and operation states during the simulation flight of the fixed-wing UAV, and obtain a first fixed-wing UAV simulation data set. The continuous space of the first fixed-wing UAV simulation data set is equally divided into n intervals. Taking the interval with the least number of data as the reference interval, randomly select the data with the same number of data as that of the reference interval in other intervals as the final data of other intervals to ensure that the amount of data in each interval is the same, thereby obtaining the fixed-wing UAV simulation data set.

[0012] In one implementation manner, the control parameters include: throttle input instruction, flap input instruction, aileron input instruction, elevator input instruction, rudder input instruction. The motion states include: linear velocity, position, angular velocity, attitude angle. Among them, the intermediate motion states include linear velocity and angular velocity, and the final motion states include position and attitude angle.

[0013] According to the third aspect of the present application, a computer device is proposed, which includes a memory and a processor. The memory stores a computer program. It is characterized in that when the processor executes the computer program, the steps of the method described in any item of the first aspect are implemented.

[0014] According to the fourth aspect of the present application, a computer-readable storage medium is proposed, on which a computer program is stored. It is characterized in that when the computer program is executed by a processor, the steps of the method described in any item of the first aspect are implemented.

[0015] Based on the above method for training a simulation model of a fixed-wing unmanned aerial vehicle (UAV) based on a cascaded neural network, a cascaded framework is proposed according to the physical process of the fixed-wing UAV's maneuvering. An intelligent generation framework for the maneuvering simulation model of the fixed-wing UAV is proposed. The maneuvering model of the fixed-wing UAV is decomposed into four deep neural network sub-models. Each network represents the corresponding physical model and has a clear physical meaning, and can be used independently. It is also possible to combine each sub-model according to different requirements to obtain a maneuvering model with different input and output. The present invention adopts a cascaded training method, that is, the predicted output of the previously trained 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 the fitting error, improve the fitting accuracy, and effectively reduce the error accumulation in the network. Description of the Drawings

[0016] Figure 1 It is a schematic flow chart of a method for training a simulation model of a fixed-wing UAV based on a cascaded neural network in an embodiment; Figure 2 It is a schematic diagram of the modeling process of the maneuvering model of a fixed-wing UAV in an embodiment; Figure 3 It is a cascaded deep neural network framework for the maneuvering process of a fixed-wing UAV in an embodiment; Figure 4 It is a schematic diagram of optimizing the hyperparameters of a neural network by a particle swarm optimization algorithm in an embodiment; Figure 5 It is a schematic diagram of the key technologies for studying the cascaded maneuvering deep neural network model of a fixed-wing UAV in an embodiment; Figure 6 It is the comparison result of the x-direction speed in the experimental results for evaluating the neural network model in an embodiment; Figure 7 It is the comparison result of the y-direction speed in the experimental results for evaluating the neural network model in an embodiment; Figure 8 It is the comparison result of the z-direction speed in the experimental results for evaluating the neural network model in an embodiment; Figure 9 It is the comparison result of the x-direction angular velocity in the experimental results for evaluating the neural network model in an embodiment; Figure 10 It is the comparison result of the y-direction angular velocity in the experimental results for evaluating the neural network model in an embodiment; Figure 11 It is the comparison result of the z-direction angular velocity in the experimental results for evaluating the neural network model in an embodiment; Figure 12The comparison result of the x-direction position in the experimental results for evaluating the neural network model in one embodiment; Figure 13 The comparison result of the y-direction position in the experimental results for evaluating the neural network model in one embodiment; Figure 14 The comparison result of the z-direction position in the experimental results for evaluating the neural network model in one embodiment; Figure 15 The comparison result of the roll angle in the experimental results for evaluating the neural network model in one embodiment; Figure 16 The comparison result of the pitch angle in the experimental results for evaluating the neural network model in one embodiment; Figure 17 The comparison result of the yaw angle in the experimental results for evaluating the neural network model in one embodiment; Figure 18 The structural schematic diagram of a fixed-wing UAV simulation model training device based on a cascaded neural network in one embodiment. Detailed implementation manners

[0017] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0018] The technical solutions of the present invention will be further described below with reference to the accompanying drawings.

[0019] A UAV is an unmanned aircraft that can fly in the air. There are various types and designs of UAVs, which can be roughly divided into two categories: rotary-wing UAVs and fixed-wing UAVs. Among them, a fixed-wing UAV is an unmanned robot that generates lift by using fixed wing blades. It usually consists of a fuselage, wings, a tail, a motor, a remote control device and other electronic devices. The flight principle of a fixed-wing UAV is that the motor drives the propeller to generate thrust, making the fuselage move forward. At the same time, lift is generated by using the wing blades, making the fuselage leave the ground and stay in the air. By adjusting parameters such as the flight angle, thrust, and rudder surface angle, its flight direction and altitude can be changed. Fixed-wing UAVs are widely used in agriculture, surveying and mapping, environmental monitoring, search and rescue and other fields, and have the advantages of high efficiency, low cost, flexibility, repeatability, etc., and have broad development prospects in future applications. More and more researchers conduct relevant experimental studies on fixed-wing UAVs based on the unmanned system simulation platform. Therefore, building an accurate and stable simulation model for fixed-wing UAVs is crucial to ensure the effectiveness, robustness and credibility of simulation experiments.

[0020] The maneuvering model is the basis and core of the fixed-wing simulation model. By modeling the maneuvering functions of fixed-wing UAVs and using them for simulation, it can effectively support the training of various UAV onboard intelligent algorithms. Based on this, there are already many open-source, refined, and easy-to-use simulation platforms. However, in the actual development and application of fixed-wing UAV simulation models, the following problems still exist: The traditional modeling method of fixed-wing UAV maneuvering models constructs maneuvering mathematical models through mathematical and physical analysis, which requires rich professional knowledge and is difficult to generalize. Building an accurate simulation model takes a lot of time and effort. Moreover, the internal parameters and dynamic characteristics of each fixed-wing UAV model are different, and each type of fixed-wing UAV needs to be modeled separately, which is also very time-consuming. It is very difficult for non-professionals to build a refined and usable fixed-wing UAV maneuvering simulation model. With the development of artificial intelligence, more and more researchers build fixed-wing UAV maneuvering models through deep learning methods. Using artificial intelligence technology and deep learning methods to build simulation models has become an emerging method. On the one hand, intelligent modeling uses deep learning algorithms to automatically learn from a large amount of data and build models. This highly automated feature greatly reduces the burden on developers, significantly shortens the development cycle of simulation models, speeds up the engineering progress, and improves production efficiency. On the other hand, the black-box model built by intelligent modeling through deep learning can model and fit complex non-linear relationships. Deep learning has excellent feature learning capabilities, shows strong fitting capabilities when dealing with non-linear problems, and also has good generalization capabilities. Compared with traditional modeling methods, deep learning methods can more accurately capture the implicit patterns and laws between data, improving the prediction and generalization capabilities of the model. These advantages have enabled deep learning to achieve breakthrough results in many fields and provide powerful tools and methods for solving complex problems in the real world.

[0021] Neural network models have good compatibility and can operate effectively on different platforms, effectively solving the problem of difficult model usage. However, currently, most researchers build fixed-wing UAV simulation models through integrated neural networks. The data types of network input and output are fixed, which has certain limitations.

[0022] Through the research of the inventors, it is found that the component-based modeling idea is a method widely used in the development of simulation models. By dividing the system into independent components or modules and separately modeling and designing each component, it realizes the simulation and analysis of the overall system behavior. This method has key characteristics 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 teamwork and model sharing, provides opportunities for researchers in different fields to cooperate and communicate, and is of great significance for the design and analysis of system modeling and simulation.

[0023] 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 scale 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. However, there is currently no relatively authoritative fixed-wing UAV flight dataset.

[0024] Based on the above problems, this application proposes a method for training a fixed-wing UAV model based on a cascaded neural network, using high-quality UAV data to provide data support for subsequent research. The intelligent generation framework based on process component-based modeling and cascaded neural network provides an efficient and accurate method for the development and training of the simulation model of fixed-wing UAVs. Modeling the fixed-wing UAV maneuver simulation model with the component-based idea can provide multiple control interfaces to meet different needs. The fixed-wing UAV maneuver simulation model is modeled separately for the dynamic model and the kinematic model according to its maneuver physical process, and then each sub-model is cascaded to form a cascaded deep neural network framework for the fixed-wing UAV maneuver process. In addition, the prediction accuracy can be further improved by automatically adjusting the network hyperparameters, making the simulation model more reliable.

[0025] Specifically, as Figure 1 shown, this application proposes a method for training a fixed-wing UAV simulation model based on a cascaded neural network. The fixed-wing UAV simulation model includes four cascaded neural network sub-models; the method includes: S101, obtaining a fixed-wing UAV simulation dataset, where each piece of data in the fixed-wing UAV simulation dataset includes various control parameters and various operating states of the fixed-wing UAV; among them, the various motion states include various intermediate motion states and various final motion states; S102. Split the fixed-wing UAV simulation dataset to obtain the training sets corresponding to four neural network sub-models. The four neural network sub-models include two kinematic sub-models and two dynamic sub-models. Among them, for each training sample in the training set corresponding to each dynamic sub-model, at least two control parameters and at least one intermediate motion state at the first moment are used as eigenvalue, and at least one intermediate operation state at the second moment is used as label value. For each training sample in the training set corresponding to each kinematic sub-model, at least one final motion state at the first moment and at least one intermediate motion state at the second moment output by at least one dynamic sub-model are used as eigenvalue, and at least one final motion state at the second moment is used as label value. The second moment is the next moment of the first moment. S103. Use the training sets corresponding to the four neural network sub-models to perform cascaded training on the four neural network sub-models respectively, to obtain the trained fixed-wing UAV simulation model.

[0026] Currently, most researchers use the method of integrated modeling, such as inputting throttle and outputting the change of position and attitude. However, by adopting the above method proposed by the present invention, that is, using the cascaded training method, the predicted output of the previously trained 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 the fitting error, improve the fitting accuracy, and effectively reduce the error accumulation in the network. In addition, according to the physical process of the fixed-wing UAV's movement, a cascaded framework is proposed, and an intelligent generation framework for the fixed-wing UAV's maneuver simulation model is proposed. The fixed-wing UAV's maneuver model is decomposed into components to obtain 4 deep neural network sub-models. Each network represents the corresponding physical model and has clear physical meanings. They can be used alone or combined according to different requirements to obtain maneuver models with different inputs and outputs.

[0027] Next, the structure of the fixed-wing UAV simulation model proposed in this application will be described.

[0028] Such as Figure 2As shown in the figure, by analyzing the maneuvering process of a 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 command is used to control the thrust output of the engine or motor to control the speed of the aircraft. The flap input command is used to increase the lift and drag during the flight of the aircraft. The aileron input command is used for roll control, and the roll motion of the aircraft is controlled by changing the position of the aileron. The elevator input command is used for pitch control, and the pitch motion of the aircraft is controlled by changing the position of the elevator. The rudder input command is used for yaw control, and the yaw motion of the aircraft is controlled by changing the position of the rudder.

[0029] The input of the fixed-wing UAV dynamics model is the control quantity at the current moment, that is, the throttle, flap, aileron, elevator, and rudder input commands, the linear velocity of the UAV in the world coordinate system at the current moment, and the angular velocity in the body coordinate system. The output is the linear velocity of the UAV in the world coordinate system and the angular velocity in the body coordinate system at the next moment. The input-output relationship is expressed as follows: (1) (2) where is the linear velocity of the fixed-wing UAV in the world coordinate system at time is the angular velocity of the fixed-wing UAV in the body coordinate system at time is the throttle input command, is the throttle input command at time is the flap input command, is the flap input command at time is the aileron input command, is the aileron input command at time is the elevator input command, is the elevator input command at time is the rudder input command, is the rudder input command at time

[0030] The input of the fixed-wing UAV kinematics model is the linear velocity in the world coordinate at the next moment obtained from the fixed-wing UAV dynamics model, the angular velocity in the body coordinate system at the next moment, the position in the world coordinate at the current moment, and the attitude angle in the body coordinate system at the current moment. The output is the position and attitude angle at the next moment. The input-output relationship is expressed as follows: (3) (4) Wherein is the position of the fixed-wing UAV in the world coordinate system at time is the position of the fixed-wing UAV in the world coordinate system at time is the attitude angle of the fixed-wing UAV in the world coordinate system at time is the attitude angle of the fixed-wing UAV in the world coordinate system at time is the angular velocity of the fixed-wing UAV in the body coordinate system at time is the linear velocity of the fixed-wing UAV in the world coordinate system at time

[0031] According to the above-mentioned maneuvering principle and modeling process of the fixed-wing UAV, a cascaded deep neural network framework for the maneuvering process of the fixed-wing UAV is proposed, as shown in Figure 3 shown. In this framework, D is a delay block, representing the data at the current moment, and NN1 to NN4 respectively represent the neural networks corresponding to formulas (1)-(4).

[0032] By analyzing the physical process of the UAV's maneuvering as described above, the data types of the input and output of each component can be obtained. 60 pieces of UAV model data are collected per second, including control parameters (throttle, flap, aileron, elevator, rudder commands) and motion states (speed, position, angular velocity, attitude). After screening and normalizing the original data, four sub-model training sets, test sets, and validation sets corresponding to the 4 sub-models are obtained as the training samples of the deep neural network, and the model is trained. The input, output data types, and corresponding physical formulas of each network are shown in Table 1.

[0033] ; Table 1 Input and output data and physical formulas of each network That is, the control parameters proposed in this application include: throttle input command, flap input command, aileron input command, elevator input command, rudder input command; the motion states include: linear velocity, position, angular velocity, attitude angle. As shown in Figure 3 shown, among them, the intermediate motion states include linear velocity and angular velocity, and the final motion states include position and attitude angle.

[0034] As shown in Figure 3As shown, the two kinetic sub - models proposed in this application include the first kinetic sub - model NN1 and the second kinetic sub - model NN2; the two kinematic sub - models include the first kinematic sub - model NN3 and the second kinematic sub - model NN4; The input of the first kinetic sub - model is the throttle input command, flap input command at the second moment, and the linear velocity at the first moment, and the output is the linear velocity at the second moment, that is, its input - output relationship is ; The input of the second kinetic sub - model is the aileron input command, elevator input command, rudder input command at the second moment, and the angular velocity at the first moment, and the output is the angular velocity at the second moment, that is, its input - output relationship is ; The input of the first kinematic sub - model is the linear velocity at the second moment output by the first kinetic sub - model, the angular velocity at the second moment output by the second kinetic sub - model, and the position at the first moment, and the output is the position at the second moment, that is, its input - output relationship is ; The input of the second kinematic sub - model is the angular velocity at the second moment output by the second kinetic sub - model and the attitude angle at the first moment, and the output is the attitude angle at the second moment, that is, its input - output relationship is 。

[0035] Since the input data of the first kinematic sub - model includes the data output by the first kinetic sub - model and the data output by the second kinetic sub - model, and the input data of the second kinematic sub - model includes the data output by the second kinetic sub - model. Therefore, the cascaded training is specifically to first train the first kinetic sub - model and the second kinetic sub - model using the training set, and then use the outputs of the trained first kinetic sub - model and the second kinetic 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 train the first kinematic sub - model and the second kinematic sub - model.

[0036] It can be understood that this application decomposes the fixed - wing UAV maneuver model into 4 deep - neural - network sub - models. Each network represents a corresponding physical model and has a clear physical meaning, so it can be used alone. The above combination method is just one of the combination methods. Those skilled in the art can also combine each sub - model according to different needs to obtain maneuver models with different inputs and outputs.

[0037] In one implementation, obtaining a fixed - wing UAV simulation data set includes: In the UAV simulation platform, based on all types of control - parameter combinations, control the fixed - wing UAV to perform simulation flights, obtain the control parameters and operation status during the fixed - wing UAV simulation flight, and obtain the first fixed - wing UAV simulation data set; The continuous space of the first fixed-wing UAV simulation dataset is equally divided into n intervals. Taking the interval with the least number of data entries as the reference interval, randomly select the corresponding number of data entries in the reference interval for other intervals as the final data for these intervals, so as to ensure that the amount of data in each interval is consistent, thereby obtaining the fixed-wing UAV simulation dataset.

[0038] Specifically, in combination with the metrics of the machine learning dataset, the present invention analyzes the characteristics of UAV flight data from six dimensions (completeness, timeliness, accuracy, standardization, unbiasedness, usability), and proposes the following fixed-wing UAV data standards.

[0039] ① Completeness: Completeness is a key quality dimension that reflects the degree to which the subject data associated with an entity has all the expected attributes and associated strength values in a given environment.

[0040] ② Timeliness: The degree to which data has reliable persistence in a given situation, to ensure that the data keeps up with the times and is not outdated.

[0041] ③ Accuracy: To what extent the attributes of the data can correctly represent the true values of the relevant attributes of a concept or event in a specific environment.

[0042] ④ Standardization: The degree to which data conforms to data standards, data models, metadata, or authoritative reference data in a given environment.

[0043] ⑤ Unbiasedness: The degree of difference in the distribution of data categories or features in a dataset in a given environment.

[0044] ⑥ Usability: The degree to which a dataset can be quickly used when applied in a specific environment.

[0045] Based on the above standards, a comprehensive data generation method is proposed.

[0046] Different flight data of UAVs have different characteristics. As much comprehensive data as possible should be collected in a flight trajectory to improve the data collection efficiency. By analyzing the characteristics of different flight data of fixed-wing UAVs, different control methods of UAV flight are used to generate flight data that meet the data standards. The data collection platform is RflySim, and the programming language is Python.

[0047] As a toolchain, the RflySim platform contains many software applications for the development of unmanned systems, such as: Python 3.8 environment, MATLAB / PSP toolbox, FilghtGear, QGroundControl, and self-developed RflySim3D, CopterSim, etc. for unmanned systems. The process of using this platform for fixed-wing UAV simulation and obtaining flight data is as follows: ① Run SITLRun.bat, set the number of drones, use Pixhawk as the control module, UE4 as the drone simulation environment, and QGroundControl as the drone ground station to configure the drones. In CopterSim, set the drone model to be simulated, as well as the starting position and attitude. Initialize a fixed-wing drone, correctly connect the flight controller, and start controlling autonomous flight.

[0048] ② Configure the corresponding Python 3.8 environment and write appropriate control algorithms for comprehensive data collection.

[0049] ③ The data sampling frequency is 60 Hz. Obtain the control state and motion state data during flight. Each piece of data includes the throttle, flap, aileron, elevator, and rudder command data obtained by each motor, the linear velocity, angular velocity, position, and attitude angle of the fixed-wing drone. A total of 100,000 flight data of the fixed-wing drone are collected.

[0050] The pseudo-code for controlling the flight of a fixed-wing drone using the Rflysim interface is as follows:

[0051] Based on the comprehensive data generation, unbiased data screening is then performed on the data.

[0052] This application proposes an unbiased data screening method to produce a uniformly distributed dataset. The continuous space is equally divided into n intervals. Based on the interval with the least amount of data, the number of data in this interval is randomly selected as the final data for other intervals to ensure that the amount of data in each interval is consistent. The uniform distribution of data 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 dataset, the model may assign higher or lower weights to these situations, resulting in biases in the prediction results for other situations.

[0053] Its algorithm pseudo-code is as follows:

[0054] Based on the above scheme, a set of methods for comprehensive data generation and unbiased data screening of fixed-wing drones is proposed, obtaining a large amount of high-quality drone flight data, comprehensively covering the maneuvering behaviors of drones in various flight states, which is beneficial to constructing a drone maneuvering simulation model and improving the authenticity and credibility of the simulation.

[0055] 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.

[0056] In one embodiment, before cascade training the four neural network sub-models respectively using the training sets corresponding to the four neural network sub-models, this application proposes to optimize the hyperparameters of the fixed-wing UAV simulation model based on the hybrid enhanced particle swarm optimization algorithm; the hybrid enhanced particle swarm optimization algorithm includes: Regarding the hyperparameter set as a particle; the hyperparameter set included in each particle is divided into two parts: a fixed dimension set and a variable dimension set. Among them, the elements in the fixed dimension set include the initial learning rate and the number of network hidden layers; the variable dimension set is the number of nodes in each hidden layer; Initialize the position and velocity of each particle, and calculate the fitness of each particle; Update the position and velocity of each particle according to the fitness, and update the individual optimal solution and the global optimal solution; Loop and iterate the above steps until the preset stop condition is reached; After the iteration is completed, obtain the optimization result of the hyperparameters of the fixed-wing UAV simulation model.

[0057] Cascade training the four sub-models respectively using the training sets corresponding to the four sub-models, including: Perform cascade training on the four sub-models respectively based on the optimized hyperparameters, test the trained model based on the test set, and if the test passes, verify the model that passes the test based on the validation set. If the test fails, continue the training. Since the training samples of each sub-model include features and labels, supervised learning can be used for training.

[0058] The specific algorithm process of the particle swarm algorithm can refer to the related technology, that is, the calculation of fitness, the update of the optimal solution, and the entire iteration process in the algorithm can all refer to the related technology of the particle swarm algorithm, which will not be elaborated here. The focus of this proposal is to propose the variability of dimensions, that is, to propose two types of sets: a fixed dimension set and a variable dimension set, that is, to use the particle swarm algorithm for calculation based on the two sets. The hybrid enhanced particle swarm optimization algorithm proposed in this application will be introduced in detail below.

[0059] In order to overcome the problem that the solution dimension of the standard particle swarm algorithm is immutable, a new method called the hybrid enhanced particle swarm algorithm is proposed. This method realizes the variability of the solution dimension by using the relationship between the number of hidden layers and the solution dimension. Divide the particle solution into two sets, one set is the fixed dimension set , and the other set is the variable dimension set , which can be expressed as: (5) The fixed dimension set The elements in it are the initial learning rate and the number of hidden layers of the network 。Fixed dimension set can be expressed as: (6) Variable dimension set The elements in are the number of nodes in each hidden layer 。Variable dimension set can be expressed as: (7) where is the dimension of the variable set The number of hidden layers in the fixed dimension set determines the dimension of the variable dimension set That is: (8) The pseudo-code of the hybrid enhanced particle swarm optimization algorithm is as follows:

[0060] The calculation process of optimizing the hyperparameters of the neural network by the particle swarm optimization algorithm in this application is as Figure 4 shown. Since there is an inconsistency between the UAV state and the control quantity in the collected flight data, that is, the current state and the control quantity of a collected flight data may not be at the same moment, a long short-term (LSTM) neural network is added to eliminate this error, that is, PSO-LSTM. The algorithm calculation steps are as follows: The first step: Data processing. After normalizing the original data set, it is divided into a training set, a test set and a validation set. The second step: Optimize the hyperparameters by the particle swarm optimization algorithm. The third step: Model evaluation. Evaluate the optimal neural network model through the validation set, and use MSE as the evaluation index. Although the training time of a single neural network is short, when optimizing the neural network by the particle swarm optimization algorithm, multiple neural networks need to be trained according to the number of particles for each iteration, and in the framework proposed in this paper, a total of 4 neural network models are required, so it takes a lot of time. This paper uses the characteristic that the particle swarm optimization algorithm is suitable for parallel distributed computing, so the computer cluster can be used for distributed computing of the network.

[0061] The pseudo-code of optimizing the hyperparameters of the neural network by the parallel particle swarm optimization algorithm is as follows:

[0062] A hybrid enhanced particle swarm optimization algorithm proposed based on the present invention. Using this optimization algorithm to find the optimal solution of the neural network by simulating the search and movement process of particles in the solution space, and determining the optimal hyperparameters of the deep neural network. It has better performance in the generation of the maneuvering simulation model and improves the performance of the neural network model.

[0063] In a specific implementation manner, the flight data used in the present invention is generated by a fixed-wing unmanned aerial vehicle in the Rflysim simulation platform released by the Reliable Flight Control Group of Beihang University. The Pixhawk is used as the control module, QGroundControl is used to configure the unmanned aerial vehicle ground station, and Unreal Engine 4 is used as the unmanned aerial vehicle simulation environment. In this experiment, various flight data of the unmanned aerial vehicle is obtained from the data storage interface of Rflysim, and the data sampling frequency is 60 Hz. Each piece of data collected includes the throttle, flap, aileron, elevator, and rudder command data obtained by each motor, the linear velocity, angular velocity, position, and attitude angle of the fixed-wing unmanned aerial vehicle. 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: 1. Divide the fixed-wing unmanned aerial vehicle maneuver model into 4 deep neural network sub-models. The key technologies for the research on the cascaded fixed-wing unmanned aerial vehicle maneuver deep neural network model are shown in Figure 5 ; 2. According to the connection method shown in the Figure 3 network, combine each sub-model to determine the input and output of different maneuver models; 3. Data processing: After normalizing the original data set used to train the simulation model, divide it into a training set, a test set, and a validation set; 4. Use the improved particle swarm optimization algorithm to optimize the hyperparameters; 5. Adopt a cascaded training method. The predicted output of the previously trained network is used as part of the training sample features of the next network. Select appropriate evaluation metrics to evaluate the optimal neural network model. The experimental results are shown in Figures 6 - 17 , Figures 6 - 17 which shows the comparison results between the true value and the prediction ( Figures 6 - 8 shows the comparison results of the speed; Figures 9 - 11 shows the comparison results of the angular velocity; Figures 12 - 14 shows the comparison results of the position; Figures 15 - 17 shows the comparison results of the attitude angle).

[0064] As shown in Figure 18 , corresponding to the above-mentioned method for training a fixed-wing unmanned aerial vehicle simulation model based on a cascaded neural network, the present application also proposes a device for training a fixed-wing unmanned aerial vehicle simulation model based on a cascaded neural network. The fixed-wing unmanned aerial vehicle simulation model includes four cascaded neural network sub-models; the device includes: A data acquisition module 710, configured to acquire a fixed-wing unmanned aerial vehicle simulation data set. Each piece of data in the fixed-wing unmanned aerial vehicle simulation data set includes various control parameters and various operating states of the fixed-wing unmanned aerial vehicle; wherein, the various motion states include various intermediate motion states and various final motion states; The training set splitting module 720 is used to split the fixed-wing UAV simulation data set to obtain the training sets corresponding to four neural network sub-models, and the four neural network sub-models include two dynamics sub-models and two kinematics sub-models; wherein, each training sample in the training set corresponding to each dynamics sub-model uses at least two control parameters and at least one motion state at the first moment as eigenvalue, and at least one intermediate operation state at the second moment as label value; each training sample in the training set corresponding to each kinematics sub-model uses at least one final motion state at the first moment and at least one intermediate motion state at the second moment output by at least one dynamics sub-model as eigenvalue, and at least one final motion state at the second moment as label value; the second moment is the next moment of the first moment. The training module 730 is used to perform cascade training on the four neural network sub-models respectively by using the training sets corresponding to the four neural network sub-models to obtain the trained fixed-wing UAV simulation model.

[0065] In one implementation, the data acquisition module 710 is specifically configured to, in the UAV simulation platform, control the fixed-wing UAV to perform simulation flight based on all types of control parameter combinations, acquire the control parameters and operation states during the simulation flight of the fixed-wing UAV, and obtain the first fixed-wing UAV simulation data set. The continuous space of the first fixed-wing UAV simulation data set is equally divided into n intervals, and the interval with the least number of data is used as the reference interval. For other intervals, randomly select the data with the same number of data as the reference interval corresponding to these intervals as the final data of other intervals to ensure that the data volume in each interval is the same, so as to obtain the fixed-wing UAV simulation data set.

[0066] In one implementation, the control parameters include: throttle input command, flap input command, aileron input command, elevator input command, rudder input command; the motion states include: linear velocity, position, angular velocity, attitude angle; wherein, the intermediate motion states include linear velocity and angular velocity, and the final motion states include position and attitude angle.

[0067] In one implementation, the two dynamics sub-models include a first dynamics sub-model and a second dynamics sub-model; the two kinematics sub-models include a first kinematics sub-model and a second kinematics sub-model. The input of the first dynamics sub-model is the throttle input command, flap input command at the second moment and the linear velocity at the first moment, and the output is the linear velocity at the second moment. The input of the second dynamics sub-model is the aileron input command, elevator input command, rudder input command at the second moment and the angular velocity at the first moment, and the output is the angular velocity at the second moment. The input of the first kinematic sub-model is the linear velocity at the second moment output by the first dynamic sub-model, the angular velocity at the second moment output by the second dynamic sub-model, and the position at the first moment, and the output is the position at the second moment; The input of the second kinematic sub-model is the angular velocity at the second moment output by the second dynamic sub-model and the attitude angle at the first moment, and the output is the attitude angle at the second moment.

[0068] In one implementation, the training module 730 is further configured to optimize the hyperparameters of the fixed-wing UAV simulation model based on the hybrid enhanced particle swarm optimization algorithm before cascaded training of the four neural network sub-models using the corresponding training sets of the four neural network sub-models respectively; the hybrid enhanced particle swarm optimization algorithm includes: Regarding the hyperparameter set as a particle; the hyperparameter set included in each particle is divided into a fixed dimension set and a variable dimension set. Among them, the elements in the fixed dimension set include the initial learning rate and the number of network hidden layers; the variable dimension set is the number of nodes in each hidden layer; Initialize the position and velocity of each particle, and calculate the fitness of each particle; Update the position and velocity of each particle according to the fitness, and update the individual optimal solution and the global optimal solution; Loop and iterate the above steps until the preset stop condition is reached; After the iteration is completed, obtain the optimization result of the hyperparameters of the fixed-wing UAV simulation model; Perform cascaded training on the four sub-models respectively based on the optimized hyperparameters.

[0069] In one embodiment, a computer device is provided. The computer device may be a server. The computer device includes a processor, a memory, a network interface, and a database connected through 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 includes 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 an external terminal through a network connection. When the computer program is executed by the processor, it implements a method for training a fixed-wing UAV simulation model based on a cascaded neural network.

[0070] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, it implements the steps of the method described in any of the above embodiments.

[0071] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in any of the above embodiments are implemented.

[0072] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and 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 DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0073] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0074] The above-described embodiments merely represent several implementation manners of the present application. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for training a fixed-wing UAV simulation model based on a cascaded neural network, characterized in that the fixed-wing UAV simulation model includes four cascaded neural network sub-models; the method includes: Obtain a fixed-wing UAV simulation data set, each piece of data in the fixed-wing UAV simulation data set includes a variety of control parameters and a variety of operating states of the fixed-wing UAV; among them, the variety of motion states includes a variety of intermediate motion states and a variety of final motion states; Split the fixed-wing UAV simulation data set to obtain training sets corresponding to the four neural network sub-models. The four neural network sub-models include two dynamics sub-models and two kinematics sub-models; among them, each training sample in the training set corresponding to each dynamics sub-model uses at least two control parameters and at least one intermediate motion state at the first moment as eigenvalue, and at least one intermediate operation state at the second moment as the label value; each training sample in the training set corresponding to each kinematics sub-model uses at least one final motion state at the first moment and at least one intermediate motion state at the second moment output by at least one dynamics sub-model as eigenvalue, and at least one final motion state at the second moment as the label value; the second moment is the next moment of the first moment; Use the training sets corresponding to the four neural network sub-models to perform cascaded training on the four neural network sub-models respectively to obtain the trained fixed-wing UAV simulation model.

2. The method according to claim 1, wherein The obtaining of the fixed-wing UAV simulation data set includes: In the UAV simulation platform, control the fixed-wing UAV to perform simulation flight based on all types of control parameter combinations, and obtain the control parameters and operation states during the simulation flight of the fixed-wing UAV to obtain the first fixed-wing UAV simulation data set; Equally divide the continuous space of the first fixed-wing UAV simulation data set into n intervals, use the interval with the least number of data as the reference interval, and randomly select the data with the same number of data as the reference interval for other intervals as the final data for other intervals to obtain the fixed-wing UAV simulation data set.

3. The method according to claim 1 or 2, characterized in that the control parameters include: throttle input command, flap input command, aileron input command, elevator input command, rudder input command; the motion states include: linear velocity, position, angular velocity, attitude angle; among them, the intermediate motion states include linear velocity and angular velocity, and the final motion states include position and attitude angle.

4. The method according to claim 1, characterized in that, The two dynamics sub-models include a first dynamics sub-model and a second dynamics sub-model; the two kinematics sub-models include a first kinematics sub-model and a second kinematics sub-model; The input of the first dynamics sub-model is the throttle input command, flap input command at the second moment and the linear velocity at the first moment, and the output is the linear velocity at the second moment; The input of the second dynamics sub-model is the aileron input command, elevator input command, rudder input command at the second moment and the angular velocity at the first moment, and the output is the angular velocity at the second moment; The input of the first kinematic sub-model is the linear velocity at the second moment output by the first dynamic sub-model, the angular velocity at the second moment output by the second dynamic sub-model, and the position at the first moment, and the output is the position at the second moment; The input of the second kinematic sub-model is the angular velocity at the second moment output by the second dynamic sub-model and the attitude angle at the first moment, and the output is the attitude angle at the second moment.

5. The method according to claim 1, characterized in that, The method further includes: Before cascaded training of the four neural network sub-models using the corresponding training sets of the four neural network sub-models respectively, optimizing the hyperparameters of the fixed-wing UAV simulation model based on the hybrid enhanced particle swarm optimization algorithm; the hybrid enhanced particle swarm optimization algorithm includes: Regarding the hyperparameter set as a particle; the hyperparameter set included in each particle is divided into a fixed dimension set and a variable dimension set. Among them, the elements in the fixed dimension set include the initial learning rate and the number of network hidden layers; the variable dimension set is the number of nodes in each hidden layer; Initializing the position and velocity of each particle, and calculating the fitness of each particle; Updating the position and velocity of each particle according to the fitness, and updating the individual optimal solution and the global optimal solution; Iteratively repeating the above steps until a preset stop condition is reached; After the iteration is completed, obtaining the optimization result of the hyperparameters of the fixed-wing UAV simulation model; The cascaded training of the four sub-models using the corresponding training sets of the four sub-models respectively includes: Performing cascaded training on the four sub-models respectively 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 includes four cascaded neural network sub-models; the device includes: A data acquisition module, configured to acquire a fixed-wing UAV simulation data set, and each piece of data in the fixed-wing UAV simulation data set includes various control parameters and various operating states of the fixed-wing UAV; among them, the various motion states include various intermediate motion states and various final motion states; A training set splitting module, configured to split the fixed-wing UAV simulation data set to obtain the corresponding training sets of the four neural network sub-models. The four neural network sub-models include two dynamic sub-models and two kinematic sub-models; among them, each piece of training sample in the training set corresponding to each dynamic sub-model uses at least two control parameters and at least one motion state at the first moment as eigenvalue, and at least one intermediate operation state at the second moment as label value; each piece of training sample in the training set corresponding to each kinematic sub-model uses at least one final motion state at the first moment, at least one intermediate motion state at the second moment output by at least one dynamic sub-model as eigenvalue, and at least one final motion state at the second moment as label value; the second moment is the next moment of the first moment; A training module, configured to perform cascaded training on the four neural network sub-models using the corresponding training sets of the four neural network sub-models respectively to obtain the trained fixed-wing UAV simulation model.

7. The device according to claim 6, characterized in that The data acquisition module is specifically configured to, in a UAV simulation platform, control a fixed-wing UAV to perform simulation flight based on all types of control parameter combinations, acquire the control parameters and operation status of the fixed-wing UAV during the simulation flight, and obtain a first fixed-wing UAV simulation data set; Divide the continuous space of the first fixed-wing UAV simulation data set into n intervals equally. Taking the interval with the least number of data as the reference interval, randomly select the data with the same number of data as that of the reference interval for other intervals as the final data of other intervals to ensure that the amount of data in each interval is consistent, thereby obtaining the fixed-wing UAV simulation data set.

8. The device according to claim 6 or 7, wherein The control parameters include: throttle input command, flap input command, aileron input command, elevator input command, rudder input command; the motion states include: linear velocity, position, angular velocity, attitude angle; among them, the intermediate motion states include linear velocity and angular velocity, and the final motion states include position and attitude angle.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

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

Images