CN-121302546-B - Unmanned aerial vehicle residual force prediction network training and control method based on deep learning

CN121302546BCN 121302546 BCN121302546 BCN 121302546BCN-121302546-B

Abstract

The invention provides an unmanned aerial vehicle residual force prediction network training and control method based on deep learning, and relates to the technical field of deep learning. The method comprises the steps of constructing a Phi network and a domain discriminator, inputting data in flight state characteristic data elements into the Phi network to obtain abstract characteristics, determining an optimal wind field exclusive linear coefficient by taking a minimized aerodynamic residual force prediction error model as a target, further obtaining predicted residual force, and inputting the abstract characteristics into the domain discriminator to obtain the wind condition type prediction probability. And training the Phi network by combining with a domain countermeasure learning mechanism, calculating the loss values of the Phi network and a domain discriminator, and alternately optimizing and adjusting the parameters of the Phi network and the parameters of the domain discriminator until the loss values are converged. The abstract features output by the Phi network are used for stripping wind condition interference, the special linear coefficients of wind fields are updated in real time, different dynamic wind fields are adapted, high-precision prediction of pneumatic residual force is realized, and the application capability of the unmanned aerial vehicle in a severe meteorological environment is ensured.

Inventors

He Yunhan
HU SHUNYU
JI ZIQIANG
LI WEI
XU YUN
QIU CHANG
SUI YIMING
DING RUIQING

Assignees

合肥工业大学
安徽智鸥驱动科技有限公司

Dates

Publication Date: 20260505
Application Date: 20251009

Claims (9)

1. The unmanned aerial vehicle residual force prediction network training method based on deep learning is characterized by comprising the following steps of: Acquiring flight data of an unmanned aerial vehicle, and preprocessing the flight data to obtain an original data set, wherein the original data set comprises a plurality of flight state characteristic data elements, and each flight state characteristic data element comprises a speed, a gesture quaternion, a motor PWM signal and an actual pneumatic residual force; inputting data in the flight state characteristic data element into a Phi network to obtain abstract characteristics, wherein the residual force prediction network comprises the Phi network and a domain discriminator; Determining an optimal wind field exclusive linear coefficient by taking a minimized aerodynamic residual force prediction error model as a target; Obtaining predicted residual force according to the abstract characteristics and the optimal wind field exclusive linear coefficient; inputting the abstract features into a domain discriminator to obtain the prediction probability of the wind condition category; determining a loss value of the Phi network according to the predicted residual force, the actual pneumatic residual force, the predicted probability of wind condition types and the double-objective loss function; determining a loss value of the domain discriminator according to the prediction probability of the wind condition category, the real wind condition label and the cross entropy loss function; Alternately optimizing and adjusting the Phi network parameters and the domain discriminator parameters according to the loss value of the Phi network and the loss value of the domain discriminator until the loss value of the Phi network and the loss value of the domain discriminator are converged; The pneumatic residual force prediction error model is as follows The optimal wind field exclusive linear coefficient is Wherein, the Representing the actual aerodynamic residual force at the kth wind condition, Representing the abstract feature corresponding to the ith flight status feature data element under the kth wind condition, A wind-farm-specific linear coefficient representing a kth wind condition, Represents the optimal wind-field-specific linear coefficient under the kth wind condition, Representing an ith flight status characteristic data element in a kth wind condition, Representing the total number of flight status characteristic data elements.
2. The deep learning-based unmanned aerial vehicle residual force prediction network training method of claim 1, wherein the acquiring unmanned aerial vehicle flight data and preprocessing the flight data to obtain an original data set, wherein the actual pneumatic residual force is Wherein m is the mass of the unmanned aerial vehicle, Acceleration, g is gravity acceleration, R is gesture rotation matrix, Is the total thrust vector under the machine body coordinate system, F is the total thrust, The motor thrust of the nth rotor wing of the unmanned aerial vehicle, The rated voltage of the nth rotor motor of the unmanned aerial vehicle is given, m is the total number of the rotor motors of the unmanned aerial vehicle, For the thrust generated by a motor with rated voltage of 14.8V, PWM is the PWM signal value of the motor, A first characteristic parameter of the motor at a rated voltage of 14.8V, A second characteristic parameter of the motor at a rated voltage of 14.8V, 、、 And Gesture quaternion of unmanned aerial vehicle respectively Corresponding to the numerical value of the corresponding code.
3. The deep learning-based unmanned aerial vehicle residual force prediction network training method of claim 1, wherein the dual objective loss function is Wherein, the Representing the actual aerodynamic residual force at the kth wind condition, Representing the abstract feature corresponding to the ith flight status feature data element under the kth wind condition, A wind-farm-specific linear coefficient representing a kth wind condition, Represents the optimal wind-field-specific linear coefficient under the kth wind condition, a represents the counterweights, Represents the ith flight status characteristic data element under the kth wind condition, K represents the total number of wind conditions, Representing the total number of flight status characteristic data elements, Representing the loss value of the domain discriminator.
4. The deep learning-based unmanned aerial vehicle residual force prediction network training method of claim 3, wherein the cross entropy loss function is Wherein, the K is a real wind condition label; representing a predictive probability of the discriminator; fully connected network of domain discriminator , To indicate a function, when the true wind condition label is k, =1, Otherwise =0; Is a standard basis vector; for the abstract features of the Phi network output, J is the total number of wind condition categories.
5. A method of unmanned aerial vehicle control, comprising: preprocessing the acquired flight data of the unmanned aerial vehicle, and inputting the flight data into a Phi network to obtain abstract features, wherein the Phi network is trained by the unmanned aerial vehicle residual force prediction network training method based on deep learning according to any one of claims 1-4; Determining an optimal wind field exclusive linear coefficient by taking a minimized aerodynamic residual force prediction error model as a target; Obtaining predicted residual force according to the abstract characteristics and the optimal wind field exclusive linear coefficient; and inputting the predicted residual force into a flight control PID controller to compensate the thrust of the motor.
6. The unmanned aerial vehicle control method of claim 5, wherein the optimal wind farm-specific linear coefficient is updated when a fluctuation in wind speed in two times of data acquisition is detected to be greater than a fluctuation threshold.
7. An unmanned aerial vehicle control system, comprising: the data acquisition module is used for preprocessing the acquired flight data of the unmanned aerial vehicle and inputting the flight data into the Phi network to obtain abstract features, wherein the Phi network is trained by the unmanned aerial vehicle residual force prediction network training method based on deep learning according to any one of claims 1-4; The coefficient confirmation updating module is used for determining an optimal wind field exclusive linear coefficient by taking the minimized aerodynamic residual force prediction error model as a target; The residual force prediction module is used for obtaining predicted residual force according to the abstract characteristics and the optimal wind field exclusive linear coefficient; And the thrust compensation module is used for inputting the predicted residual force into the flight control PID controller and compensating the thrust of the motor.
8. An electronic device comprising a memory and a processor; The memory is used for storing a computer program; The processor is configured to implement the deep learning-based unmanned aerial vehicle residual force prediction network training method according to any one of claims 1 to 4 or the unmanned aerial vehicle control method according to any one of claims 5 to 6 when executing the computer program.
9. A computer readable storage medium, characterized in that the storage medium has stored thereon a computer program which, when executed by a processor, implements the deep learning based unmanned aerial vehicle residual force prediction network training method according to any one of claims 1 to 4 or implements the unmanned aerial vehicle control method according to any one of claims 5 to 6.

Description

Unmanned aerial vehicle residual force prediction network training and control method based on deep learning Technical Field The invention relates to the technical field of deep learning, in particular to an unmanned aerial vehicle residual force prediction network training and control method based on deep learning. Background The unmanned aerial vehicle pneumatic residual force prediction technology is used as a key technology for improving the adaptability and the flight stability of the complex environment of the unmanned aerial vehicle, and the traditional prediction method combines physical modeling and data driving to try to realize accurate prediction by fitting the relation between the flight state and the pneumatic residual force. The traditional pneumatic residual force modeling mostly adopts a pure physical modeling or single working condition data driving mode, the data driving method depends on flight data under specific wind conditions, for example, a pure physical model is difficult to capture transient effects of complex airflow disturbance (such as gusts and turbulence) under the simplifying assumption, and when the wind conditions change, the single working condition data driving model is easy to cause sudden reduction of prediction precision due to deviation of data distribution, and has poor generalization capability on the wind conditions which are not seen, so that the real-time response requirement of a dynamic wind field can not be met. Disclosure of Invention The invention aims to solve the problems that the existing residual force prediction method adopts pure physical modeling or single working condition data driving, has poor generalization capability, and has suddenly reduced prediction precision under the influence of complex airflow disturbance. In order to solve the above problems, in a first aspect, the present invention provides an unmanned aerial vehicle residual force prediction network training method based on deep learning, including: Acquiring flight data of an unmanned aerial vehicle, and preprocessing the flight data to obtain an original data set, wherein the original data set comprises a plurality of flight state characteristic data elements, and each flight state characteristic data element comprises a speed, a gesture quaternion, a motor PWM signal and an actual pneumatic residual force; inputting data in the flight state characteristic data element into a Phi network to obtain abstract characteristics, wherein the residual force prediction network comprises the Phi network and a domain discriminator; Determining an optimal wind field exclusive linear coefficient by taking a minimized aerodynamic residual force prediction error model as a target; Obtaining predicted residual force according to the abstract characteristics and the optimal wind field exclusive linear coefficient; inputting the abstract features into a domain discriminator to obtain the prediction probability of the wind condition category; determining a loss value of the Phi network according to the predicted residual force, the actual pneumatic residual force, the predicted probability of wind condition types and the double-objective loss function; determining a loss value of the domain discriminator according to the prediction probability of the wind condition category, the real wind condition label and the cross entropy loss function; And alternately optimizing and adjusting the Phi network parameters and the domain discriminator parameters according to the loss value of the Phi network and the loss value of the domain discriminator until the loss value of the Phi network and the loss value of the domain discriminator are converged. Optionally, the acquiring flight data of the unmanned aerial vehicle and preprocessing the flight data to obtain an original data set, where the actual pneumatic residual force is Wherein m is the mass of the unmanned aerial vehicle,Acceleration, g is gravity acceleration, R is gesture rotation matrix,Is the total thrust vector under the machine body coordinate system, F is the total thrust,The motor thrust of the nth rotor wing of the unmanned aerial vehicle,The rated voltage of the nth rotor motor of the unmanned aerial vehicle is given, m is the total number of the rotor motors of the unmanned aerial vehicle,For the thrust generated by a motor with rated voltage of 14.8V, PWM is the PWM signal value of the motor,A first characteristic parameter of the motor at a rated voltage of 14.8V,A second characteristic parameter of the motor at a rated voltage of 14.8V,、、AndGesture quaternion of unmanned aerial vehicle respectivelyCorresponding to the numerical value of the corresponding code. Optionally, the pneumatic residual force prediction error model is The optimal wind field exclusive linear coefficient is Wherein, the Representing the actual aerodynamic residual force at the kth wind condition,Representing the abstract feature corresponding to the ith flight status f