Implementation of a particle filter on a GPU for nonlinear estimation in a manufacturing remelting process

Abstract

This paper discusses the use of modern methods for estimation in Vacuum Arc Remelting, a manufacturing process used in the production of specialty metals for aerospace applications. Accurate estimation in this process is challenging because the system is nonlinear and all available measurements are corrupted with noise. Particle filters are nonlinear estimators that sample a set of points, called particles, in the state space to construct discrete approximations of probability density functions. Real-time issues arise when using these methods in systems with low signal-to-noise ratios because of the large number of particles required to reach acceptable accuracy. In these cases, the throughput of the particle filter becomes critical, and parallelization becomes a necessity. This paper presents the implementation of a particle filter using a GPU with NVIDIA's CUDA technology, whose large number of processor cores allows massive parallelization.