FierClass: A multi-signal, cepstrum-based, time series classifier

Abstract

The task of learning behaviors of dynamical systems heavily involves time series analysis. Most often, to set up a classification problem, the analysis in time is seen as the main and most natural option. In general, working in the time domain entails a manual, time-consuming phase dealing with signal processing, features engineering and selection processes. Extracted features may also lead to a final result that is heavily dependent of subjective choices, making it hard to state whether the current solution is optimal under any perspective. In this work, leveraging a recent proposal to use the cepstrum as a frequency-based learning framework for time series analysis, we show how such an approach can handle classification with multiple input signals, combining them to yield very accurate results. Notably, the approach makes the whole design flow automatic, freeing it from the cumbersome and subjective step of handcrafting and selecting the most effective features. The method is validated on experimental data addressing the automatic classification of whether a car driver is using the smartphone while driving.