Sequential Monte Carlo: Enabling Real-time and High-fidelity Prognostics

Abstract

Uncertainty quantification and propagation form the foundation of a prognostics and health management (PHM) system. Particle filters have proven to be a valuable tool for this reason but are generally restricted to state-space damage models and lack a natural approach for quantifying model parameter uncertainty. Both of these issues tend to inhibit the real-world application of PHM. While Markov chain Monte Carlo (MCMC) sampling methods avoid some of these restrictions, they are also inherently serial, and, thus, MCMC can become intractable as model fidelity increases. Over the past two decades, sequential Monte Carlo (SMC) methods, of which the particle filter is a special case, have been adapted to sample from a single, static posterior distribution, eliminating the state-space requirement and providing an alternative to MCMC. Additionally, SMC samplers of this type can be run in parallel, resulting in drastic reductions in computation time. In this work, a potential path toward real-time, highfidelity prognostics using a combination of surrogate modeling and a parallel SMC sampler is explored. The use of SMC samplers to enable tractable parameter estimation for full-fidelity (i.e., non-surrogate-assisted) damage models is also discussed. Both of these topics are studied in the context of fatigue crack growth in a geometrically complex, metallic specimen subjected to variable amplitude loading.