GENERALIZED FEATURE-EXTRACTION FOR TIME-VARYING AUTOREGRESSIVE MODELS

In this paper, a feature extraction scheme for a general type of nonst ationary time series is described. A non-stationary time series is one in which the statistics of the process are a function of time: this t ime dependency makes it impossible to utilize standard globally derive d statistical attributes such as autocorrelations, partial correlation s, and higher order moments as features, In order to overcome this dif ficulty, the time series vectors are considered within a finite-time i nterval and ape modeled as time-varying autoregressive (AR) processes, The AR coefficients that characterize the process are functions of ti me that may be represented by a family of basis vectors, A novel Bayes ian formulation is developed that allows the model order of a time-var ying AR profess as well as the form of the family of basis vectors use d in the representation of each of tile AR coefficients to be determin ed. The corresponding basis coefficient are then invariant over the ti me window and, since they directly relate to the time-varying AR coeff icients, are suitable features for discrimination. Results illustrate the effectiveness of the method.