Increasing Warning Times for the Onsite Earthquake Early Warning Algorithm
Wednesday, 17 December 2014
A significant source of alert latency in earthquake early warning (EEW) algorithms, which can possibly be reduced, is the duration of the time window used to estimate EEW parameters following a P wave detection. We investigate whether the detection time for the τc-Pd Onsite algorithm can be reduced by estimating EEW parameters using shorter waveform windows. The Onsite algorithm is currently one of three EEW algorithms implemented in ShakeAlert, the earthquake early warning system being developed for the west coast of the United States. Onsite uses the ground-motion period parameter (τc) and peak initial displacement parameter (Pd) to estimate the magnitude and expected ground shaking of an ongoing earthquake. The current implementation of Onsite requires a 3 second window of the P waveform before issuing magnitude estimates. We use a large suite of waveform records of local earthquakes (M>3) recorded in southern California to examine the performance for a range of window lengths between 1-5 s. We examine source-to-station distances up to 300 km and focus on the results for distances less than 100 km. We find that Pd shows a stronger correlation with magnitude than is observed for τc, for all tested window lengths. τc does not correlate with magnitude for earthquakes with M<4; however, we do not apply any quality metrics to remove bad or noisy records (e.g. Böse et al., 2012). The broadband waveforms give more reliable Pd and τc measurements than strong motion waveforms, at least for moderate sized earthquakes, likely because of higher signal to noise ratios. However, these results may be biased since the dataset does not include many waveforms for earthquakes with M>6. Overall, preliminary results suggest that it may be feasible to use shorter window lengths (1 sec) without significantly degrading the magnitude estimate in the Onsite algorithm. We will determine the optimal waveform window that provides both fast and reliable estimates of τc and Pd for use in the Onsite algorithm, and will work with spatially varying station density. Further, we will investigate if sliding window lengths can be used to improve the τc and Pd estimates continuously.