Improved RBF Neural Network Control System Design for Helicopter of Large Envelope

Abstract

The controller design of helicopter was very complicated because of the strong coupling between channels and the complex nonlinear connection. To solve this problem, chose several state points to linearize the system on the condition of little perturbation. Based on H infinity mixed sensitivity theory, designed attitude angle control system, and utilized the error input and control output for sample collection, then built the RBF neural network which trained by the collected samples. Considering that different RBF neural networks need to select different learning rates, this will bring great inconvenience to the use of RBF. In order to solve such problem, a new type of dynamic optimal learning rate is derived, which will be optimized for each iteration. Tested the fully trained RBF neural network controller at non-design points. Simulation results show that the control system can track the attitude instruction excellently, the tracking speed is fast. And the neural network controller shows strong robustness and adaptivity in the whole flight envelope.