Abstract

In this paper, a novel algorithm is introduced for the analysis of long-memory patterns hidden in electromagnetic (EM) readings prior to earthquakes. The algorithm builds upon previous work on long-memory detection in EM measurements by fusing Support Vector Machine (SVM) classifiers with well-deployed power law fit tests and Rescaled-Range (R/S) time-series variability methods. To apply the algorithm, fractal power law in the wavelet domain is assessed so as to identify fractional Brownian motion (fBm) segments of continuously monitored pre-earthquake EM activity. The selected segments are then further processed through R/S Analysis in order to further refine the detection of prominent fBm behaviour. The combined output of the two methods is used to train a SVM classifier which is subsequently employed to verify similar fBm states in existing EM data and to allow for rapid fBm detection in large data sequences of unprocessed or newly incoming EM readings. The SVM classifier is added in a modular fashion, on top of pre-earthquake monitoring algorithms, and can be trained with a small fraction of a huge available dataset of EM readings. Three earthquake events in Greece, corresponding to different time occurrences and geographic locations, were investigated. For each of the three earthquakes, data collected by a nearby EM measurement station one month prior to the peak event were analysed by the proposed method. The results yielded an overall accuracy rate of at least 90% for the detection of specific, prominent fBm segments despite the fact that the fBm profile in the three investigated earthquake sequences was very different.

Keywords: Seismic effects; Spearman correlation coefficient; Support vector machine