Development and tuning of automatic control methods for nonlinear systems using computational intelligence techniques with emphasis on the control of unmanned aerial vehicles

Abstract

The objective of this dissertation is to develop and fine-tune automatic control techniques for nonlinear systems, with a focus on unmanned aerial vehicle control, through the application of computational intelligence methods. Specifically, this research focuses on utilizing radial basis function networks (RBFNs), known for their simplicity and fast training, and cooperative particle swarm optimization (CPSO), recognized for its improved optimization results through collaboration among multiple swarms. An important issue faced in this dissertation is the problem of controlling nonlinear systems by utilizing nonlinear control methodologies, primarily backstepping control and model predictive control. Backstepping control offers robustness, and stability for non-strict feedback systems, whereas the model predictive control (MPC) method involves formulating and solving an optimization problem at discrete time steps, enabling accurate prediction of future system behavior and control in complex dynamic systems with constraints and disturbances. The main nonlinear systems that are investigated in this dissertation are unmanned aerial vehicles (UAVs), with a specific focus on quadrotor vehicles. Controlling the quadrotor, especially concerning trajectory tracking, presents a significant challenge due to its inherently nonlinear and underactuated nature, characterized by intercoupled terms. In this thesis the trajectory tracking problem was addressed by developing a new nonlinear backstepping controller which integrates RBF neural networks. Backstepping controllers are based on first-principles equations to face the significant challenge of effectively handling inherent nonlinearities, but are vulnerable to unmodeled dynamics and uncertainties in practical applications. To tackle this challenge, the thesis proposes a novel solution which integrates a backstepping controller with RBF networks for handling uncertainties during quadrotor trajectory tracking, thus offering a data-driven approximation for handling unmodeled uncertainties. In addition to developing an effective tracking control strategy for a quadcopter, it is equally important to properly tune its control parameters, especially when more than one controller is used for regulating the system. To this end, in this thesis, a novel CPSO optimization framework is designed for optimizing the tuning parameters of a quadrotor trajectory tracking control scheme. The control framework included two subsystems: an MPC controller for position tracking and a PID scheme for attitude stabilization. This approach involves collaborative optimization of the numerous controllers tuning parameters, resulting in improved tracking performance, enhanced robustness and efficient optimization within reasonable timeframes. In tandem with the development of an algorithm for the optimal tuning of a quadcopter's control parameters, two additional cooperative particle swarm algorithms were also devised to address and resolve high-dimensional non-convex problems. To this end, two novel CPSO frameworks were formulated to address the problems related to optimal reactive power flow (ORPF) management in smart distribution grids and critical parameter identification in WWTPs. To be more specific a CPSO optimization and control framework was designed in order to tackle the reactive power flow (RPF) problem of photovoltaic-heavy distribution networks. Furthermore, in response to the estimation of critical parameters challenges faced in wastewater treatment processes (WWTPs), a new CPSO-identification framework was proposed that can be used for solving a nonlinear optimization problem. This thesis also addresses another crucial issue concerning the modeling and control of nonlinear time-varying systems. In this context, the challenge lies not only in choosing between linear and nonlinear models but, more importantly, in ensuring that the model employed can adapt its parameters so as to effectively track changes in the system's dynamics. In this thesis, a new nonlinear control framework is presented in which adaptive neural network models are incorporated. A comprehensive framework for nonlinear adaptive control is developed, ensuring satisfactory control performance across various operation regions. The control law of the closed-loop system is proven to be asymptotically stable using Lyapunov stability theory. Two case studies are conducted within this framework, involving a nonlinear autoregressive exogenous (NARX) system and a time-varying continuous stirred tank reactor (CSTR). The strategies presented in this dissertation are evaluated across a range of case studies, including simulated scenarios. The effectiveness of these proposed schemes is showcased through comparisons with other approaches documented in the bibliography.