Pseudo principal components analysis for feature extraction and pattern recognition of time-series data

We proposed a novel method to extract a feature from time-series data by principal components analysis (PCA) with time-delay embedding, and showed its usefulness in pattern recognition. We first resampled from the original time series data and constructed a new data with time-delay embedding. Then we applied PCA to the new data to get a pseudo principal component (PPC), which now represents the newly constructed data and hence the original time series data as well. The PPC was used as a feature vector for the original data, and the pattern classification of was performed upon PPC. In order to improve the performance of the classification, we incorporated with the continuous wavelet transform (CWT) to the newly constructed data before we take the PPCs. The results showed that the new method is useful for classification tasks of time series data, and that the performance is improved when well combined with the CWT technique.