Real-time degradation modeling and residual life prediction for component maintenance and replacement
Abstract
Condition based maintenance (CBM) policies and degradation modeling approaches play a proactive role in maintenance and rely on the current state of the device using sensory information from condition monitoring of components. Condition monitoring (CM), an inherent component of CBM, is the process of collecting real-time sensory information from a functioning device in order to reason about the health of the device and predict its future performance. To effectively use CM information, it is useful to characterize a component degradation signal, which is a characteristic pattern in the sensor information that evolves with respect to component's operating time and condition. Our approach considers monitoring a single component, a thrust ball bearing, which is an inherent component in every type of rotational mechanical assembly. The goal of this research is to develop decision support models that use real-time condition monitoring information and degradation signals to improve maintenance and replacement decisions. We develop Neural Network models that utilize vibration-based sensory information to estimate failure times of partially degraded thrust bearings at various points in their service life. Empirical Bayes and Maximum Likelihood Estimation approaches are then used to compute bearing residual life distributions, which are continuously updated with every data acquisition. We also develop a random coefficients model that combines degradation characteristics of the bearing's population along with real-time vibration information from a specific bearing with the objective of computing remaining life distributions. Bayesian updating is then used to revise residual life distributions with each new sensory acquisition from a partially degraded bearing. The remaining life distributions computed using the neural network models and the Bayesian degradation models are used in conjunction with a simple age replacement model developed in Renewal Theory to compute optimal replacement times that correspond to the minimum long-run average replacement cost per unit time. Results from the replacement model change according to the updated remaining life distributions with each data acquisition process. We believe the major contribution of this research will be in decision making under large scale condition monitoring to support maintenance, replacement, and inventory policies when hundreds of components are being monitored simultaneously.
Degree
Ph.D.
Advisors
Lawley, Purdue University.
Subject Area
Industrial engineering|Mechanical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.