Unsupervised System for Discovering Patterns in Time-Series

Within this paper we present a framework for discovering patterns in time-series by unsupervised feature selection and unsupervised, self-organised clustering. The proposed unsupervised feature selection algorithm is determining the feature relevance for a variety of transformations to select a set of features to build the feature space. We propose to take the phase space as basis and extend it by the selected features. The core of the presented approach is the self-organised clustering in feature space with Multi-SOMs and MultiNeural-Gas. For lack of prior knowledge about the real number of clusters, K is estimated by using a combination of common clustering analysis coefficients that have been adapted for M-SOMs and M-NGas. The presented approach of unsupervised feature selection, feature space construction and subsequent unsupervised clustering has been evaluated with time series from a robotic application with promising results.