Lessons learned from a one-year-long deployment of a machine learning-based infrasound monitoring system

Since the deployment of the Hungarian infrasound array (PSZI) in May 2017, a large number of PMCC (Progressive Multichannel Correlation) detections have been collected and manually categorized using ground-truth information from independent sources. This dataset enabled the training and evaluation of machine learning (ML) models for infrasound signal classification.

The final ensemble model consists of a Random Forest model trained on PMCC-related features and a Convolutional Neural Network trained on spectrograms.  To automate infrasound signal processing, these were trained to distinguish detections originating from known sources from those of unknown origin.

 Based on the ensemble ML model, we designed a monitoring system to help with daily routine processing. We aimed to remove noise, such as detections associated with industrial activity from the daily list of detections and highlight those that are from signals of interest, for instance quarry blasts, thunderstorms and activity of the Etna. During the one-year-long test phase, the system achieved high accuracy in classifying quarry blast signals and successfully identified multiple eruptions of Mount Etna, highlighting its capability for automated infrasound signal source classification.