RECURSIVE BAYESIAN AUTOREGRESSIVE CHANGEPOINT DETECTOR FOR SEQUENTIAL SIGNAL SEGMENTATION

摘要

The contribution addresses a sliding window modification of the Bayesian autoregressive change-point detector (BACD) enabling the sequential localization of signal changes (change-point detection). The modification consists in using the simplified data-dependent Bayesian evidence normalizing the classical BACD formula and in the recursive evaluation of these two functions. The suggested approach seems to be computationally effective and numerical stable as shown by experiments. Apart from the evaluation of the algorithm accuracy two illustrative examples with modelled signals are given. One application to the violin signal segmentation demonstrates the algorithm performance - even relatively weak and gradual signal changes can be detected.