Item

Abstract

The huge volume of infrasound detections asks for machine learning techniques for the automatic classification of signals. Our objective was to assess machine learning algorithms in identifying signals. Since 2017, the Hungarian infrasound array has collected approximately one million detections, processed with the Progressive Multi Channel Correlation method. Of these, we categorised some 14,000 detections from quarry blasts, storms and a power plant. These detections constitute the dataset for machine learning training, validation and testing. After pre-processing, features were extracted from the waveforms both in the time and frequency domain, to characterize the physical properties of the signals. We also defined PMCC related features to measure the similarity between the detections. For training, two classifiers were selected, Random Forests and Support Vector Machines. Hyperparameter tuning was performed with three-fold cross-validation using grid search. As a metric, f1 score was selected, and the confusion matrices were analysed. The goal was to separate the detections labelled as quarry blasts from the storm and power plant classes. The results reach 0.88 f1 scores, and high true positive rate for the quarry blasts, which show promising step in the direction of infrasound signal classification via machine learning.