Supervised frequency change detection using MCMC methods

Classical abrupt change detection needs to fix a threshold often difficult to be tuned. We propose to learn this value on a training set of signals, considering then the detection as a supervised classification problem. We describe here a bayesian approach. As a closed form of this bayesian learning is untractable, a stochastic simulation method (MCMC) is proposed leading to a good approximation of the posterior probability of change. The method is applied first to an academic problem for which an analytical solution exists and allows comparisons. Then a generalization of the composite hypothesis, corresponding to the most of real cases, is proposed and applied to abrupt frequency change detection in noisy multicomponent signals. Presented results show the influence of training set size on the performances of detection.