A data-driven prognostic approach based on AR identification and hidden Markov models

In this work a data-driven prognostic approach based on AutoRegressive (AR) estimation and hidden Markov models (HMMs) is addressed. In particular, the approach is capable of achieving Prognostic and Health Management (PHM) tasks such as real time detection and Remaining Useful Life (RUL) estimation. The approach can be seen as composed of a training part (offline) and an exploitation part (online). The offline part relies upon the use of a scalar health indicator coming from the system identification field: the Itakura Saito (IS) spectral distance. In particular, raw acceleration data, gathered in an unsupervised framework from the machine, are modeled by AR processes and then transformed into IS. Then, HMMs are used to map such IS signals into a finite number of parameters. Moreover, in the training procedure of HMMs, a left-to-right clustering of unsupervised data, based on Mixture of Gaussians (MOG) distribution is proposed. During the online exploitation a simulation of a running signal is tested against trained ones in order to carry out PHM tasks in real time. Simulations have been performed using a public benchmark available in ”NASA prognostic data repository”. It contains run-to-failure tests on bearings, on which acceleration signals are gathered. In particular the gathering experiment simulates an industry application, under constant operating conditions. Results of simulations, performed on real time data, validate the proposed prognostic approach and make the combined use of IS an HMMs a reliable way in achieving PHM goals.