A new framework for pattern recognition of time-series data
Kewei Tang, Daewon An · 2004
Time-series or time-sequence data are a collection of the sequential measurements of some physical systems over a certain period of time, such as seismograph, speech signals, electroencephalograph, and so on. In this thesis, cluster analysis and classification of time-series or time-sequence data that are dynamic over the time are investigated. More specifically, a novel framework for identifying similarities and performing classifications of the data is proposed. The thesis proposes a novel ontological framework that can provide a theory of measuring similarity and dissimilarity among practical datasets. This new framework is then used to evaluate the effectiveness of a new computationally efficient method for clustering analysis and pattern classification. This new method transforms the original time series data into a multi-dimensional space from which prominent features of the data are extracted for cluster analysis and pattern classification. It is built upon the well known Principal Components Analysis (PCA) method and is therefore called Pseudo Principal Components Analysis (PPCA). When combined with the Continuous Wavelet Transform (CWT) method, we found that PPCA provides comparable or better results than other approaches at reduced computational requirements.