Machine learning localization of ice fracture events from 3-axis geophone data from using a physics-based feature space

Three-axis geophones were deployed in an array for more than a month as a part of the 2021 Sea Ice Dynamics Experiment (SIDEx’21) with the objective of classifying and localizing flexural ice crack events. However, using conventional geophone localization methods requires detailed knowledge of frequency-dependent and time-dependent ice characteristics that were inadequately characterized by calibration data from the beginning of the deployment. As a result, we instead used a localization method with a physics-based feature space and machine learning. First, an observation feature vector is calculated from field or simulation data time series consisting of what we term delay angle (a time delay of arrival metric that abstracts out group speed), amplitude ratio (an amplitude metric that abstracts out attenuation and source level), and arrival angle (based on longitudinal/transverse wave relationships). To localize field data, OASES OASP was first used to generate training, test and evaluation time series datasets for a variety of environments and source locations; Monte Carlo was used to introduce noise into the traces, and feature vectors were extracted. The resulting SVM model is then used to estimate the location of each real-world event time series, using the calibration dataset for real-world performance evaluation. [Work funded by ONR.]