A gaussian process-enabled MCMC approach for contaminant source characterization in a sensor-rich multi-story building

Abstract

This paper presents contaminant source localization and characterization in a sensor-rich multi-story building with a large-scale domain. Bayesian framework infers the posterior distribution of source location and characteristics from the sensor network with the model uncertainty and inaccurate prior knowledge. A Markov Chain Monte Carlo method with a Metropolis-Hastings algorithm provides samples extracted from the posterior distribution. A computationally efficient Gaussian process emulator allows Markove Chain Monte Carlo sampling to use a physics-based model with tractable computational cost and time. The posterior distribution obtained by the proposed method through hypothetical contaminant release in a four-story building with total 156 subzones and sensors approaches true values of parameters of interest closely and shows the efficacy for parameter inference in a large-scale domain.