CS229 - Project Final Report: Automatic earthquake detection from distributed acoustic sensing (DAS) array data

We attempt to automatically detect earthquake events in distributed acoustic sensing (DAS) data via a supervised learning approach. Detecting earthquakes with different magnitudes could potentially provide the ability of predicting major catastrophic events. Introduction Distributed acoustic sensing (DAS) is an emerging technology used to record seismic data that employs fiber optic cables as a probing system. By measuring the backscattered energy of a pulsing laser transmitted down a fiber optic cable, it is possible to measure the strain rate occurring within different sections of the cable [1]. DAS recording systems have been shown to measure data comparable with conventional geophones [2] and have been successfully used in exploration and earthquake seismology settings [3, 4]. Recently, a DAS array has been deployed beneath Stanford campus using existing telecommunication fiber optic cables. The data recorded from this array have the potential for near-surface imaging of the subsurface and early-warning earthquake monitoring. In order to accurately detect earthquakes for early warning we must extract earthquake signals from the surrounding urban and ambient noise that is constantly recorded by our DAS array. Therefore, in addition to classifying earthquake signals, we also attempted to classify urban and ambient noise present in the array dataset with the end goal of classifying each sample or windows of samples as one of these three types of signal. To perform this classification, we used a supervised learning approach, in which we train a classifier on labeled training data that consists of the processed amplitudes from our DAS array and their corresponding type of signal. Therefore, in our application we aimed to separate the signal into three classes, namely, ambient noise, urban noise, and earthquake signal. We first attempted this classification process using a single sample from our amplitude data (i.e., our feature is a scalar) with Gaussian Naive Bayes, softmax and kernelized support vector machine (SVM) supervised learning models. We found that using this feature space, our estimated models classified nearly all samples as ambient noise (the most dominant class) and not a single earthquake was detected. To improve the performance of our classifier, we then changed the feature space to overlapping windows of amplitude data. We found that using this feature space, we improved the precision and accuracy of the machine learning algorithms.