Advanced fusion diagnosis with evolving data streams for health management of complex engineered systems

Abstract

Effective health diagnostics for engineered systems provide tremendous benefits such as improved safety, reliability, and reduced costs of operation and maintenance. Extensive literature review on health diagnostics and prognostics of complex engineered systems have identified following three important challenge problems. - Because of the implicit relationships between different health states of a system and sensory signals in complex engineered systems, it is difficult to develop a generally applicable yet robust failure diagnosis algorithm. - A challenge for health diagnostics of dynamic environments is that the relationships between sensory signals and their health states may change or examples which belong to a novel health state appears over time. These changes may degrade the performance of the trained health diagnosis algorithm. -Although querying procedure can provide valuable information for dynamic health diagnostics and prognostics, sometimes the high costs of labeling make it difficult to acquire the true health states of new incoming examples immediately. To address these important challenges, three research solutions are proposed: a classification fusion with concurrent subset algorithm selection for health diagnostics, an adaptive fusion learning approach to build a robust classification model, an active learning with holding method for health diagnostics of data streams. In this dissertation, in order to demonstrate the effectiveness of the proposed research solutions numerical examples and different real-world applications are employed as case studies.