Engine noise separation through Gibbs sampling in a hierarchical Bayesian model

Abstract

Abstract An algorithm based on a hierarchical Bayesian model is introduced to separate sources highly overlapping in time and frequency and observed through correlated references. The method is applied to internal combustion (IC) engine signals with the aim of separating the contributions due to different physical origins. The results are compared to the ones provided by classical Wiener filter. The Bayesian context allows correlated references to be taken into account with no consequences on the identifiability of the sources, thanks to the possibility of providing some regularizing prior information in the form of Bayesian prior laws. Moreover, the credibility interval on the estimated sources derives directly from the adopted sampling strategy. Finally, it is shown in a simple case that the proposed algorithm can be rewritten as a weighted sum of the classical and cyclic Wiener filters proposed by Pruvost in 2009. As opposed to them, the present algorithm autonomously chooses one or the other depending on the characteristics of the analysed signals. Even if the development context is the separation of the sources in an IC engine, the presented method is general and can be applied to any source separation problem.