METHODOLOGY FOR FLEET UNCERTAINTY REDUCTION WITH UNSUPERVISED LEARNING

Abstract

Operational and environmental variance can skew reliability metrics and increase uncertainty around lifetime estimates. For this reason, fleet-wide analysis is often too general for accurate predictions on heterogeneous populations. Also, modern sensor based reliability and maintainability field and test data provide a higher level of specialization and disaggregation to relevant integrity metrics (e.g., amount of damage, remaining useful life). Modern advances, like Dynamic Bayesian Networks, reduce uncertainty on a unit-by-unit basis to apply condition-based maintenance. This thesis presents a methodology for leveraging covariate information to identify sub- populations. This population segmentation based methodology reduces fleet uncertainty for more practical resource allocation and scheduled maintenance. First, the author proposes, validates, and demonstrates a clustering based methodology. Afterwards, the author proposes the application of the Student-T Mixture Model (SMM) within the methodology as a versatile tool for modeling fleets with unclear sub-population boundaries. SMM’s fully Bayesian formulation, which is approximated with Variational Bayes (VB), is motivated and discussed. The scope of this research includes a new modeling approach, a proposed algorithm, and example applications.