Active Fault Detection using Model Predictive Control

Actuator faults are critical to detect as they reduce the ability of the controller to influence the system, in addition to causing unwanted system behaviour. Incipient actuator faults are therefore important to detect at an early stage in order to rectify the fault before losing the ability to do so as the fault increases in severity. Detection algorithms using parameter estimation are well suited for detecting incipient faults, as they are able to detect small deviations from the normal dynamics. However, all estimation algorithms needs sufficiently descriptive data in order to correctly estimate the system parameters. This thesis proposes an active fault detection algorithm using parameter estimation, which aims at increasing the detectability of incipient actuator faults. The estimate used in the fault detection algorithm is improved upon, by ensuring the input sufficiently excites the system, and this is achieved by constructing a persistently exciting controller. The proposed controller uses the framework provided by model predictive control, and includes the previously applied input in the constraints used within the optimization problem in the controller. Numerical simulations were done where the proposed persistently exciting controller is compared to using a nominal controller with Gaussian white noise added to the input as an auxiliary excitation signal. The persistently exciting controller reduces the amount of false alarms when compared to using white noise for excitation, but is not able to detect the fault at an earlier stage