Joint detection of variance changes using hierarchical Bayesian analysis

This paper addresses the problem of detecting variance changes in time-series coming from two different sensors. The two sequences are modeled as zero-mean white Gaussian sequences with piecewise constant variances. Bayesian inference allows to define interesting priors which reflect the correlations between the two change-point sequences. Unfortunately, the Bayesian estimators for the change-point parameters cannot be expressed in closed-form. A Metropolis-within-Gibbs algorithm allows to generate samples distributed according to the posterior distributions of the unknown parameters. The hierarchical structure of the Bayesian model is also used to estimate the unknown hyperparameters.