Data Mining on Extremely Long Time-Series
Scott Simmons, Louis Jarvis, David E. Dempsey, Andreas W. Kempa-Liehr · 2021 International Conference on Data Mining Workshops (ICDMW) · 2021
The crux of machine learning for time-series revolves around the engineering of relevant features as inputs into a predictive model. Such features are designed to capture the informative characteristics of an input signal. For many real world applications it is required that these features are not only informative but also interpretable for domain experts. Recent work has led to the automation of time-series feature engineering. In this paper, we are demonstrating that for very large and complex time-series, the dynamics of time-series features itself are sources of discriminative and interpretable time-series features. This concept is evaluated for a predictive maintenance use case (time-series classification) and the estimation of a vibrational sensor signal from a co-located sensor (time-series regression). Both use cases show promising preliminary results for the efficacy of extracting time-series features from the dynamics of time-series features.