Noise-Tolerant Self-Embedded LSTM for Seismic Event Classification

João Paulo Canário, Ricardo Araújo Rios

Abstract: The modeling of temporal data has significant implications for a wide range of applications, such as the classification of volcano seismic events. Accurately identifying such events at the earliest stages allows for proactive measures to be taken to support communities located nearby volcanoes. However, the presence of noise in the original signals significantly diminishes the performance of event classifiers, requiring the implementation of preprocessing methods to remove noises before adjusting models, which is a time-consuming process that typically needs specialized knowledge and appropriate tools. In this work, we first developed a deep neural network capable of achieving superior results in predicting seismic events for the Llaima volcano in Chile. Secondly, we devised a new memory cell that is more robust to noise. Experiments indicate that our trainable memory cell can be applied to scenarios with different noise influences.