Abstract
In this new era of troubled peace, fleets of military or civil contingency equipment are maintained in storage facilities to readily surge in response to military operations, humanitarian assistance, and disaster relief (HADR) or pandemic. However, the current maintenance approach is premised on traditional preventive or corrective maintenance. The regime is labour-intensive repetitive inspections. It is unproductive and not sustainable from both economic and technical workforce perspectives. With the advent of embedded instrumentation and data science, condition-based maintenance (CBM) has grown to be a popular strategy to minimize the phenomenon of “over-maintenance”. Condition-based maintenance (CBM) presents a more cost-effective approach to owning and maintaining these assets under long-term storage. However, research in prognostics and health management (PHM) for fleets in storage remains sporadic. The lack of large degradation data sets or established degradation model pose challenges to employing typical data-driven or model-based approaches to predict residual storage life (RSL) for storage prognosis. Since storage degradation is most aptly collected during the operations and maintenance (O&M) phase, the lack of technical expertise in storage facilities to identify and collect the appropriate degradation data to initiate the PHM study compounds the challenge. Even if the right expertise can be engaged, the slow rate of degradation implies only a few significant degradation data can be collected. With only sparse degradation data available, it limits our options to employ the existing common prognostic approaches. Furthermore, the diversity of initial storage states and storage conditions also exacerbated the degradation complexity in storage. There is no formal instruction to initiate and manage an equipment or component when placed under storage, vis-à-vis, the preventive and corrective maintenance instructions prescribed by the suppliers or manufacturers. In this work, the objective is to address the challenges of implementing condition-based maintenance for equipment fleets in storage through the development of a data-driven approach with the sparse condition and degradation data. My thesis explores multi-output Gaussian process regression (MOPGPR) to model non-monotonic RSL for energy storage systems (Li-ion batteries) and electronics (light-emitting diodes) under various storage conditions. We focused on battery and electronics due to their pivotal roles in the 4th Industrial Revolution (4IR) and the Green economy. The effects of training data selection and test data parameters, like detrending mean, on the prediction performance, vis-à-vis the computational resources, are examined. Due to the unique sensitivity of RSL prediction to the inherent component bias, a detrending bias suppression framework is developed and achieved mean absolute percentage error (MAPE) to be lower than 1%. Novel applications of MOGPR are also explored to perform degradation anomaly detection and fleet storage prognosis under new or untested conditions with limited or no know-how. A simple and scalable bootstrapping MOGPR PHM framework premised on sparse degradation data and limited computing resources, is designed to be executable by the non-sophisticated operators and maintainers, who manage a large fleet in storage. This research democratizes PHM and empowers a paradigm shift in the fleet storage industry, from being overly reliant on the prescribed maintenance instructions of preventive and corrective maintenance, to be able to implement an evidence-based design-it-yourself (DIY) regime of condition-based maintenance.