Extreme learning machine for time sequence classification

Abstract

In this paper, a new framework to effectively classify the time sequence is developed. The whole time sequence is divided into several smaller sub-sequence by means of the sliding time window technique. The sub-sequence is modeled as a linear dynamic model by appropriate dimension reduction and the whole time sequence is represented as a bag-of-systems model. Such a model is very flexible to describe time sequence originated from different sensor source. To construct the bag-of-systems model, we design the codebook by using the K-medoids clustering algorithm and Martin distance between linear dynamic systems. Such a technology avoids the problem that linear dynamic systems lie in non-Euclidean manifold. After obtaining the represented of time sequence, an extreme learning machine is utilized for classification. Finally, the proposed method is verified on some benchmark and shows that it obtains promising results.