《Joint Segmentation and Classification of Time Series Using Class-Specific Features》.pdf

1056 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 34, NO. 2, APRIL 2004 Joint Segmentation and Classification of Time Series Using Class-Specific Features Zhen Jane Wang and Peter Willett, Fellow, IEEE Abstract—We present an approach for the joint segmentation example, an autoregressive (AR) model of low complexity if it and classification of a time series. The segmentation is on the basis is “ambient” and free of interesting artifacts; but it may con- of a menu of possible statistical models: each of these must be de- tain a short-duration signal (or even, perhaps, more than one) scribable in terms of a sufficient statistic, but there is no need for whose presence indicates a situation of interest. One can assume these sufficient statistics to be the same, and these can be as com- that there has been some effort to model each sort of such tran- plex (for example, cepstral features or autoregressive coefficients) as fits. All that is needed is the probability density function (PDF) sient signal, both in terms of selection of some sort of statistics of each sufficient statistic under its own assumed model—presum- that might be thought “sufficient” (mathematically, we shall as- ably this comes from training data, and it is particularly appealing sume that they are just so) and in terms of characterizing sta- that there is no need at all for ajoint statistical characterization of tistically the distribution of each such statistic based on some all the statistics. There is similarly no need for an a-priori specifica- training data. At a more specific level, perhaps one transient type tion of the number of sections,