Health index construction with feature fusion optimization for predictive maintenance of physical systems

In any prognostics and health management framework, the key to accurate remaining useful life prediction is finding a signal that directly quantifies the health status of a physical asset. When a direct measurement of an asset’s health is not available, researchers and practitioners develop virtual health indices, fused from many different signals, to quantify a unit’s health status. However, current metrics and methods used to engineer virtual health indices struggle to deal with large signal noises and do not consider health index performance on a population of units wholistically. In response to these challenges, we propose a new trend metric, which more accurately quantifies the monotonicity of a signal in the presence of noise, as well as a new probabilistic method, called meta-probability, for comparing health indices across a population of units to better understand the unit-to-unit variation in health index performance. To demonstrate the utility of our proposed trend metric and probabilistic comparison tool, we formulate a multi-objective optimization problem with the goal of creating the best virtual health index across a population of units by fusing many sensor signals and derived features. The proposed metrics and optimization scheme are evaluated in two case studies considering rolling element bearing run-to-failure data. The proposed optimization method, which considers the newly proposed monotonicity metric as one of the objectives, is found to create a health index that is more optimal for a larger percentage of the units in the population than five existing health index construction methodologies reported in the literature.