Simulation of platform-free inertial navigation system of unmanned aerial vehicles based on neural network algorithms

Abstract

The object of research is the process of controlling the trajectory of unmanned aerial vehicles (UAVs) in autonomous flight mode based on neural network algorithms. The study is based on the application of numerical-analytical approach to the selection of modern technical solutions for the construction of standard models of platformless inertial navigation systems (BINS) for micro and small UAVs, followed by support for assumptions. The results of simulation in the Matlab environment allowed to simulate the operation of the UAV control system based on MEMS technology (using microelectromechanical systems) and Arduino microcomputers. It was also possible to experimentally determine the nature of the influence of the structure of the selected neural network on the process of formation of navigation data during the disappearance of the GPS signal. Thus, to evaluate the effectiveness of the proposed solutions for the construction of BINS, a comparative analysis of the application of two algorithms ELM (Extreme Learning Machine)-Kalman and WANN (Wavelet Artificial Neural Network)-RNN (Recurrent Neural Network)-Madgwick in the form of two experiments. The purpose of the experiments was to determine: the study of the influence of the number of neurons of the latent level of the neural network on the accuracy of approximation of navigation data; determining the speed of the process of adaptive learning of neural network algorithms BINS UAV. The results of the experiments showed that the application of the algorithm based on ELM-Kalman provides better accuracy of learning the BINS neural network compared to the WANN-RNN-Madgwick algorithm. However, it should be noted that the accuracy of learning improved with the number of neurons in the structure of the latent level <500, which iincreases computational complexity and increases the learning process time. This can complicate the practical implementation using micro- and small UAV equipment. In addition, thanks to the simulation, the result of the study of the application of the proposed neural network algorithms to replace the input data instead of GPS signals to the input BINS, allowed to estimate the positioning error during the disappearance of GPS signals. Also, the application of the WANN-RNN-Madgwick algorithm allows to approximate and extrapolate the input signals of navigation parameters in a dynamic environment, while the process of adaptive learning in real time.