Methods of time series data mining: diseño e implementación de un detector de actividad volcánica del sangay.

Abstract

The Sangay is one of the most active volcanoes in Ecuador. Due to sporadic ash fall and climatic conditions, it is difficult to maintain continuous seismo-acoustic monitoring using using the closest seismic station to the volcano. In this study, we assess if a regional network can help fill in the observational gaps when nearest station measurements are not available. To this end, we design and implement a binary detector using signals from seismometers. A previously-generated dataset of times corresponding to labeled events and non-events from Sangay were taken between 2017 and 2020. The detector was implemented independently for seven stations located more than 50 km from the Sangay's summit, taking the two horizontal components (North and East) of the seismometer. To reduce the number of data points, two autoencoders were trained at each station with the Sangay events: one with the time series and the other with the spectral power density. We find that 10 dimensions are sufficient to adequately reconstruct the input data. A support vector machine (SVM) classifier was trained for each station for activity detection using the data from the autoencoders and other statistical indicators, obtaining a detection accuracy greater than 98 % at all stations. Finally, we asigned the probability of wheter or not the measurement in a station correspond to an event from the Sangay using 5-fold cross-validation, with which the minimum probability of detection of the majority of events and non-classified events was estimated.