Process Mining Capabilities Extended to Time Series Analysis as Applied to Meteosat Water Vapor Images

Abstract

This paper presents two approaches for using Process Mining to the analysis of time series. They are applied to a time series of Meteosat full-disk water-vapor satellite images characterized by their pairwise similarity. Event logs constructed from this time series are processed with the Inductive Miner algorithm producing models in the form of Petri Nets and Directly-Follows Graphs (DFG). In the first approach, logs corresponding to five years of similarity indexes of daily images were grouped by astronomical seasons and Petri Nets models were discovered for each season. From them, trace fitness obtained with token-based replay conformance checking were used as features for characterizing the logs, and supervised machine learning algorithms were applied to obtain classification models. In the second approach, Gram matrices using different graph kernels were computed from the individual DFGs associated to each log, and used by support vector machines for constructing classification models. For all seasons, high-performance models were found, matching those obtained with state-of-the-art techniques, with the plus of providing a description of the time series data structure, demonstrating the potential of PM-based approaches. The analysis revealed the presence of some consistent miss-classification patterns that could be related to the ongoing process of climate change. The results obtained are preliminary and open interesting research directions.