Time Series Classification Using Images

This work is a contribution to the field of time series classification. We propose a novel method that transforms time series into multi-channel images, which are then classified using Convolutional Neural Networks as an at-hand classifier. We present different variants of the proposed method. Time series with different characteristics are studied in this paper: univariate, multivariate, and varying lengths. Several selected methods of time-series-to-image transformation are considered, taking into account the original series values, value changes (first differentials), and changes in value changes (second differentials). In the paper, we present an empirical study demonstrating the quality of time series classification using the proposed approach.