Contributions to prognostics and health-aware control of dynamic systems

Abstract

Maintenance is indispensable in every industry that uses machines and is inevitable regardless of the attention paid to it. The cumulative effect of its neglect or a non-optimized approach is a long or short-term shock on a company's growth. It, therefore, comes as no surprise, the overwhelming interest in the area of intelligent maintenance where comprehensive control units can detect, alert personnel, predict when a failure will occur, and manage faults to extend the life of components. Borne out of this is a new field of study, Prognostics and Health Management, which emerged only about two decades ago. This thesis contributes to this nascent field of study, and most importantly to a novel interest in the field of incorporating component prognostic information into control to manage the extent of influence degradation has on the efficient output of a plant. In sum, the thesis seeks to (1.) Contribute to component prognostics and how uncertainty can be efficiently handled and (2.) Promote the incorporation of prognostic information in a control scheme (i.e. Model Predictive Control). For prognostics, the thesis considers two critical components, a wind turbine blade composite material and an insulated gate bipolar transistor utilizing two different types of prognostic methods, the model and data-based methods respectively. Wind turbine blades are by far the most exposed component to damage predominately due to their level of mechanical activity in the turbine operation. Forces such as gyroscopic and gravity, debris in wind, and the effect of the stochastic nature of wind contribute to a gradual damaging effect culminating in a complete blade breakdown. Given that the blade material itself is innate, mathematical degradation equations dependent on material properties to predict the extent of the material damage in the absence of sensor information is used. Therefore, with a stiffness degradation algorithm aided by a zonotopic Kalman filter, the remaining useful life of a wind turbine's blade is predicted with a model-based prognostic technique. For the data-based technique on the insulated gate bipolar transistor, a run-to-failure data set from NASA Ames was used on a novel data-based Evolving Ellipsoidal Fuzzy Information Granule modelling (EFFIG) scheme to predict the remaining useful life of the component with appropriate metrics proving its effectiveness when compared to other methods and also offers an online recursive learning advantage and explainability characteristics, a sought-after quality in practice. The issue of uncertainty quantification has been one of the main conundrums in prognostics, thus in both cases, this thesis contributes to using set-based methodologies which are known to be easy to formulate and computational friendly to quantify and propagate. For health management by a suitable control scheme, two variants are proposed. First, a controller that takes a direct consideration of prognostics information via the stiffness degradation algorithm proposed as done in the model-based prognostics case and a health management scheme that considers the characteristic property of reliability of pumps in a network. Various contributions were made in this research undertaking. For, the reliability aware control, an MPC controller is designed that ensures reliability of the Drinking Water Networks using the Bayesian theorem, specifically for the Barcelona drinking water network. Even though some papers have indeed tackled the issue of network reliability in the literature, they fail to account for uncertainties that arise from consumer demand, which cannot be ignored. Also a more viable optimization problem against wind turbine's blade degradation is tackled with the health-aware control considering incorporated health information.