A Prognostic Modeling Methodology for Multistream Degradation-based Signals

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

High-valued engineering assets used in the manufacturing and service sectors are increasingly being instrumented with hundreds of sensors that are used for condition monitoring and predicting remaining lifetime, i.e., prognostics. The underlying idea of prognostics is that sensor data often contain unique fault-based features that can be utilized for prediction. Today, multiple sensors are increasingly being used to monitor different aspects of a degradation process in a single machine, thus, it is important to leverage the combined information embedded in these sensors. However, most of the existing prognostic models developed to date focus on single-sensor applications and do not account for any data quality issues. This award supports fundamental research to provide prognostic models for multistream sensor signals where data observations may be sparse and/or missing; an aspect that has consistently challenged the implementation of prognostics in real-world applications. This research will enable numerous industries in the manufacturing and service sectors to increase equipment availability, prevent catastrophic failures, and reduce maintenance costs. Research findings will also be used to advance engineering education by incorporating the findings of this work in graduate coursework. Almost all approaches available for degradation modeling are based on a key assumption that is rarely satisfied in reality; the quality of sensor data is high and states are observed on a continuous basis. This project will bridge the gap between theory and practice in the area of prognostics by relaxing the simplifying assumptions that have traditionally been the basis for developing prognostic models. This will be accomplished by incorporating parsimonious estimation methods and functional data analysis to model degradation signals with missing and sparse observations. Specifically, functional Principal Component Analysis will be used to model simultaneous variations of multivariate signal recorded by different sensors. Principal Components Analysis through Conditional Expectation will be used to address the challenges arising from the missing observations. Isotonic regression methods will be used to ensure monotonicity of the resulting models, which will be key in estimating remaining lifetime.