Prognosis of ball bearings with non-characteristic degradation using recurrent neural network

Abstract

Breakdowns in rotary machines are often related to bearing failure. In recentyears many prognostic and diagnostic models of bearings have been developedto prevent unexpected shutdowns. Estimating the remaining useful lifetime(RUL) of bearings have in many cases demonstrated to be problematic dueto complex degradation mechanisms.This thesis explores alternative pre-processing and neural network basedmodels for RUL prediction of bearings. Predictive models are generatedfrom six training bearings, where vibration data is sampled from start toend of life (EOL). The models are further tested on 11 test bearings forRUL predictions. Both training and test bearings are degraded in a uniquemanner as they are subjected to loads above the bearings’ specification. Thecombination of limited training data and abnormal wear patterns suggestthat accurate predictions of RULs are quite difficult to obtain.