Measuring the accuracy of time series reduction methods based on modified dynamic time warping distance calculations

Abstract

Representation of sensor data in the form of time series is a crucial aspect of numerous related tasks such as comparison, reduction, clustering, and classification. Time series representation methods included in most programming languages/integrated development environments support dimensionality reduction, data preprocessing, and feature extraction for time series data, as do several normalization techniques. This research study focused on 14 different methods of dimensionality reduction from the TSepr (R Studio) package on eight different time series, which are collections of sensor data of varying lengths. The similarity of the reduced time series and the original time series is compared using a modified version of dynamic time warping with time alignment measurement. These methods are further combined with the Gaussian kernel function to normalize the distance between variously aligned series. The results showed that perceptually important points (PIP) and piecewise linear approximation (PLA) were found as the best methods for TS reduction with a minimum deviation (error term) as low as 5 – 12%. The results also indicate that PIP performs significantly differently compared to seasonal decomposition, while there are no significant differences between PIP and the other methods (PLA, FEACLIPTREND, and FEACLIP). In addition, this research study demonstrated the development of an interactive web-based application in which time series are stored in csv files, and the distance between them is calculated through the chosen reduction method.