Automated classification of signals with duration-dependent segments via class-specific features and Gibbs sampling

We study time series modelled as variable duration hidden Markov (VDHMM) segments. The observations within each segment can be statistically dependent and parameterized according to an unknown distribution. Gibbs sampling and class-specific (CLASP) feature methods, which are comfortable in high-dimensional and complex problems, are applied to the segmentation and, ultimately, the integrated classification of these segmented time series; that is, the segmentation is automatic, the segments are soft-classified according to the candidate VDHMM models, and a menu of VDHMM models themselves compete to claim the overall time series. Simulation results show their effectiveness, and, in particular, we highlight the benefits of incorporating the Markov structure of the segment types and the segment duration probability information. Techniques for estimation of the VDHMM models from training data are also given.