Abstract
Robust controllers are typically designed using the upper bounds of the magnitudes of model uncertainties and exogenous disturbances to ensure the stable operation of a system across a broad range of conditions. This may produce a conservative controller with unnecessarily reduced performance. Artificial neural network techniques can provide online estimates of uncertainties and disturbances, which can be counteracted by subtracting the estimates on the control input channel. This relaxes conservativeness and increases system performance. We propose a novel Multiple-Input Multiple-Output (MIMO) Higher Order Sliding Mode Control (HOSMC) trajectory tracking controller for a general class of highly-coupled systems. Uncertainties and exogenous disturbances are estimated using two different neural networks: a Multilayer Perceptron Neural Network (MLPNN) and a Radial Basis Function Neural Network (RBFNN). Real-time learning rules for the neural network weights are derived within the adaptive MIMO super-twisting SMC structure through a stability analysis. The controllers are implemented in simulations of a 72 kg heavy-lift Coaxial Octorotor Uncrewed Aerial Vehicle performing an infinity maneuver while ascending. The turbulent fluctuations within an atmospheric boundary layer model are used to generate wind disturbances. Uncertainties include sudden changes in mass and propeller faults. The RBFNN-based controller slightly outperforms the MLPNN-based controller. When compared with a conventional MIMO HOSMC controller, the Integral Squared Error of both controllers is about 82% lower, and the Integral Absolute Error and Integral Time-weighted Absolute Error of both is about 64% lower. A first step towards practical implementation is demonstrated via Hardware-in-the-Loop simulations on a commercial flight controller.