Statistical Analysis of Sensor Network Time Series at Multiple Time Scales

Modern sensor networks often collect data at multiple time scales in order to observe physical phenomena that occur at different scales. Whether collected by heterogeneous or homogenous sensor networks, measurements at different time scales are usually subject to different dynamics, noise characteristics, and error sources. We explore the impact of these effects on the results of statistical time series analysis methods applied to multi-scale time series data. As a case study, we analyze results from GPS time series position data collected in Japan and the Western United States, which produce raw observations at 1Hz and orbit corrected observations at time resolutions of 5 minutes, 30 minutes, and 24 hours. We utilize the GPS analysis package (GAP) software to perform three types of statistical analysis on these observations: hidden Markov modeling, probabilistic principle components analysis, and covariance distance analysis. We compare the results of these methods at the different time scales and discuss the impact on science understanding of earthquake fault systems generally and recent large seismic events specifically, including the Tohoku-Oki earthquake in Japan and El Mayor-Cucupah earthquake in Mexico.