Abstract

The problem of fault detection in nonlinear systems with application to trajectory tracking of nonholonomic wheeled mobile robots (WMRs) is addressed in this thesis. For the considered application, a nonholonomic wheeled mobile robot--having nonlinear kinematics--is required to follow a predefined smooth trajectory (in the absence of obstacles in the environment). This goal has to be accomplished despite the presence of faults that may occur in two of its major subsystems which are vital for navigation, namely the driving subsystem and the steering subsystem. These faults are assumed to be caused by actuator faults in either of these two subsystems. The problem addressed here is to detect the presence of faults and to determine the subsystem which has been affected by these faults. Toward this end, two different fault detection approaches are proposed and investigated. The first approach is based on system identification through Extended Kalman Filters (EKF) whereas the second one is based on system identification via artificial neural networks. In the former approach a novel method for residual generation is proposed while in the latter by utilizing the neural network's formal stability properties the desired performance can be guaranteed. Each of the proposed fault detection methods is studied subject to two different kinds of controllers (namely a dynamic linear controller and a dynamic feedback linearization based controller) and two different types of actuator faults (namely the Loss-of-Effectiveness fault and Locked-In-Place fault). In this way, the impact of the controller strategy on the fault detection approach is also investigated and evaluated.