Data-driven methods for prognostics and fault detection in industrial components

Recent years have seen an increasing focus by researchers on the issues explicated by the Industry 4.0 and Industry 5.0 paradigms, leading to define several maintenance strategies and methods to increase the diffusion of such paradigms. Nowadays, the key aspect for manufacturing industries is reduce the down-time of the production lines, optimising the machinery maintenance schedule in order to make maintenance actions faster and easier. In this way, industries see the maintenance costs reducing and the production volume increasing. To optimise the maintenance schedule of a machinery, it is usefull not only know if there are faults (fault detection) in a machinery component, but also having real-time information about its Health Condition (HC) and predictions about its time-to-failure (time-to-alarm) that is the time instant at which a machinery failure occurs (its critically damaged). This is the aim of fault prognostics. Formally, fault prognostics is the process of forecasting the time-to-failure or the time-to-alarm of industrial items using degradation models. From these predictions it is possible to derive the Remaining Useful Life (RUL) of components, that is the time before a failure occurs. Prognostics is a key point in the so called Predictive Maintenance (PM) strategy, that is one of the most widely adopted maintenance optimisation strategies. Due to the high complexity of modern industrial machinery, in most context it is necessary the development data-driven solutions to fault detection and fault prognostics. In fact, define mathematical models for such complex systems can be too difficult and leads to modeling errors that affects the degradation state predictions. As a result, statistical-based data-driven methods and Machine Learning (ML)-based data-driven methods have found fertile ground for their definition and development. Typically, the development of such methods require many run-to-failure acquisitions to understand the degradation of a machinery component behavior and create the related degradation model, accounting for possible sources of uncertainty, e.g., noise on data acquisition. However, such experiments are destructive and expensive for manufacters. Thus, decreasing the number of run-to-failure experiments is key in reducing Predictive Maintenance costs. This thesis presents a data-driven fault prognostics methods that requires only a single run-to-failure experiment, providing time-to-alarm predictions and so RUL ones, that are probabilistically guaranteed by the use of the scenario approach theory. In addition, it is proposed extensions of such method that allows to handle situations in which neither a single run-to-failure experiment is feasible, i.e., online prognostics situations, as well as a mathematical generalization of the method by exploiting the Wait-and-Judge (WJ) paradigm of scenario approach theory. Last but not least, real-world applications of the principles of data-driven fault detection in the context of leak detection and electric vehicles are presented.