UAV Maneuverability Model Based On RNN-LSTM

The application scenarios are continuing expandding from computer vision,medical diagnostics to automated driving with the continuous development of deep learning.In the field of UAV,machine learning has been gradually used to achieve precise control of UAV.Neural network was used to adjust the internal PID to achieve more stable flight control for UAV at first.However,due to the limitation of the overall computing power of the UAV flight control board,it is impossible to run more complicated control algorithms,which limits the application range of the neural network.Through investigation and research,it was found that the relationship between UAV body information and control maneuverability in the current environment could be constructed by means of neural network,so that the control amount of the UAV in the specified UAV body could be predicted in the real flight environment,and achieved more precise control of the UAV.A new RNN-LSTM model for UAV flight control is applied for the above analysis of UAV control.For this purpose,the construction of the UAV control,data acquisition and analysis ground station platform,the acquisition of data for neural network model training,the training of RNN-LSTM models and the verification experiments applied to UAV control are carried out.The specific content of this article includes the following aspects:1)A C#-based UAV control,data acquisition and analysis ground station was developed for the control of UAV flight and the collection and analysis of UAV data.The ground station controls the UAV to carry out various repetitive flight experiments in multiple directions through the UAV control and data acquisition platform with real-time and high-precision positioning.Then acceleration and power of the aircraft and the corresponding UAV control amount at each moment were collected.After data pre-processing and standardization,a flight data set with 12800 sets of variable relationships was finally constructed.2)A new RNN-LSTM model applicable to UAV flight control was applied,and model training and verification were compared.The model considered the complex relationship between the acceleration,speed and power of the UAV and the control information of the UAV's motion flight,and useed multiple layers of hidden layers to achieve more accurate analysis.At the same time,the Adam algorithm was used to realize the automatic adjustment of the learning rate,and the actual training speed of the network was improved without reducing the accuracy.After the model training,compared with the BP neural network training results,the model had higher precision in predicting the control amount under the specified body information of the drone in the test set.3)Based on the optimized RNN-LSTM model,the flight verification of the specified path was controlled by the UAV in the real scenario.The RNN-LSTM model was loaded on the UAV control and data acquisition and analysis ground station platform,and corresponding control amount was calculated according to the straight line and quadrilateral flight planning path in the platform.The RNN-LSTM model predicted the control amount of the UAV based on the speed,power and set acceleration of the UAV at the corresponding time,and useed the control amount to control the UAV to obtain the actual flight trajectory data.By comparing the actual flight trajectory data(coordinates,acceleration and time)with theoretical calculations(flight distance,acceleration and time,etc.),the predicted acceleration was basically in accordance with the theoretical calculation value,and the flight distance had an error of 7%.After the construction,training and real scene verification of the RNN-LSTM model,the UAV body information obtained from the model's predicted control information was in good agreement with the experimental theoretical values,it could be used to specify UAV control under the body information.