Autopilot design using neural network for tilt-rotor unmanned aerial vehicle with nacelle-mounted wing extension

Abstract

The mathematical dynamics model of the tilt-rotor UAV that has nacelle mounted wing extensions (WE) is presented based on CFD analysis using FLUENT and DATCOM. The advantage of the aerodynamic performance of the WE is compared to the performance of the original tilt-rotor UAV in a trim analysis as well as simulation. Two different types of neural network controllers are designed for the inner loop and outer loop of the tilt-rotor and its WE variant. One is the linear parameterized sigma-pi neural network (SPNN) and the other is the nonlinear parameterized single hidden layer perceptron neural networks (SHL-NN). In order to improve the control performance of outer loop, Pseudo-Control Hedging (PCH) is applied to the outer loop as well as the inner loop of neural networks controls. The dynamic inversion on a linear model of the original tilt-rotor at hover conditions is used as a baseline. Both of SPNN and SHL-NN adaptive controllers minimize the error of the inversion model. This error typically occurs due to the use of an approximate tilt-rotor model for helicopter mode instead of the WE model throughout the flight envelope from helicopter to airplane mode. The waypoint navigation and the automatic hover guidance are applied to the most outer loop of the neural network controller for the autonomous flight, which consists of nacelle conversion and reconversion as well as automatic take-off and landing. The fast dynamic reference commands generated by the autonomous waypoint guidance are inputted to the outer loop control in order to make the PCH of the outer loop effective. Lastly, the nonlinear simulation results are compared under turbulent wind conditions, in which the WE is more negatively affected than the original tilt-rotor model.