Dynamic Time Warping for Classifying Long-term Trends in Time Series

This paper explores the potential of dynamic time warping (DTW) for recognizing different segments in time series data characterized by their long-term trends and curvature. To perform classification, a set of reference data for each class is required, where each time series in the reference set represents a typical shape of that class. The classification process involves computing the DTW distance between a given time series and each reference time series, then assigning the time series to the class with the minimum distance. Experiments on both simulated and real-world time series data from two different use cases demonstrate that DTW can correctly classify the different segments. Additionally, the paper investigates whether incorrectly classified phases could indicate data security issues. Additional experiments are performed to assess the number of data points required to reliably classify a segment correctly. These experiments highlight the limitations and emphasize the importance of selecting good reference data.