A Hybrid State Particle Filter for Failure Prognosis in Deteriorating Systems

In this paper, we present a new approach for predictive analytics using time-series condition-monitoring data. A double hybrid state-space model was proposed, with a continuous latent state vector serving as a unit-less health indicator regarding the health status of a system and a discrete latent state vector denoting the operating condition of the system through time. We also considered two observation levels. The first level consisted of continuous noisy sensor observations. The survivability of the system was considered as a second level of observations. Instead of an explicit formula, a non-parametric method relating the continuous latent state to the sensor observations using a single-layer feed-forward neural network, called the Extreme Learning Machine (ELM), was considered and a new Bayesian iterative procedure using the Expectation-Maximization algorithm for ELM training and joint state-parameter estimation was presented. Finally, our method was tested on a number of simulated degradation data sets.