An improvement of SAX representation for time series by using wavelet packet decomposition and FastDTW
Yonghua Guo, Kai Zhu, Xu Wei Cheng · 2023
Aiming at the shortcomings of conventional SAX that using the mean value as the eigenvalue may lead to misclassification, this paper comprehensively considers the numerical distribution difference and morphological fluctuation characteristics of time series data. Firstly, wavelet packet decomposition (WPD) can decompose various types of non-smooth random signals into wavelet components at different scales. Firstly, the WPD is used to extract the features of the time series, and then the Fast Dynamic Time Warping (FastDTW) is used to take advantage of the metrics, and the FastDTW metrics are applied to the features to define the global trend indicators. Then, the SAX distance is combined with the trend distance, and an improved distance measure is proposed. Finally, the experimental results on different time series datasets show that the proposed method has the highest classification accuracy in 11 of the 20 datasets, which proves that the representation method proposed in the paper is significantly better than the original SAX representation and the improved SAX representation.