Model-based Prediction of the Remaining Useful Life of the Machines

Accurate prediction of the remaining useful life (RUL) of machines is becoming mandatory in exploiting the asset in an efficient and secure way by avoiding the unplanned downtimes. In this paper we present an approach to the RUL prediction developed for a shot blasting machine by analyzing the recordings from inexpensive vibrational sensors. The key idea consists of (i) employing generalized Jensen-Renyi divergence (JRD) as a measure of change in the vibrational pattern (ii) exploiting the monotone relationship between JRD and the abrasive wear in rotor blades and (iii) using a Markov model to describe wear dynamics. The unknown parameters of the Markov model are obtained by expectation-maximization approach. Prediction of the remaining useful life is done by executing Monte Carlo simulations on the updated model and evaluation of the first passage time of the JRD index. The approach is validated experimentally by running the machine up to the failure, hence allowing for naturally evolving wear progression.