Abstracts

Appropriate feature-based representations aresignificant for time series analysis and subsequent machinelearning applications. A low-dimensional set of comprehensivefeatures is instrumental to improving the efficiency and accuracyof classification. The main contribution of this work is to developa new methodology to characterize seismic time series signals byextracting and selection statistical features from them. Thismethodology allows one to study earthquakes with muchlower-dimensional, yet informative, datasets. In this work, a largenumber of unbiased features were generated from raw time seriesusing the highly-comparative times series analysis (HCTSA)operation library. The similarity between each pair of features wasrepresented by the measure of maximal information coefficient(MIC). MATLAB functions were implemented to compute the similaritymatrix of the feature dataset generated by HCTSA. Affinitypropagation (AP) was used for clustering similar features andselecting exemplary features from different clusters. Theseindependent exemplary features were determined to characterize theoriginal data. The process was applied to data from realearthquakes and a representation of reduced features was generatedto characterize the original times series signals. Results showedthat MIC reflected a reliable measure of general associations(similarities) between features, and features which were associatedtended to be placed into the same cluster. The clustering resultsalso showed that the average distance within a cluster wasgenerally less than the distances between different clusters, whichdemonstrated that the selected exemplary features were relativelyindependent. This work opens a door to the use of feature-baseddatasets from seismic signals for model inversion and improvedsubsurface characterization.Advisors/Committee Members: Ruben Juanes (advisor).