Pipeline Processing with an Iterative, Context-Based Detection Model
Technical Report,19 Mar 2013,22 Dec 2015
NORSAR Kjeller Norway
Pagination or Media Count:
Under existing detection pipelines, seismic event hypotheses are formed from a parametric description of the waveform data obtained from a single pass over the incoming data stream. The full potential of signal processing algorithms is not being exploited due to simplistic assumptions made about the background against which signals are being detected. A vast improvement in the available computational resources allows the possibility of more sensitive and more robust context-based detection pipelines which glean progressively more information from multiple passes over the data. Under this contract we have evaluated iterative components at different levels of the pipeline hierarchy. At the level of the raw waveforms, we have evaluated schemes for the detection and cancellation of noise transients which can both reduce the detection capability of a station and supply the parametric datastreams with phase detections which may result in spurious event hypotheses. At the single array level, the sensitivity to sites of monitoring interest can be diminished by energy arriving from other directions, for example from ongoing aftershock sequences. We demonstrate how energy from the direction of interest can be enhanced at the expense of the nuisance energy using adaptive beamforming and empirical matched field processing. The aftershock scenario can have significant consequences for the generation of near real-time bulletins both due to the increased number of events and the deterioration of fully automatic event bulletins due to spurious phase association. We provide proof-of-concept of a system for spawning a targeted process for accurate aftershock characterization in a given source region such that all associated phases are removed from the parametric datastreams. New iterations of Global Association, or equivalent algorithms, would then read in pre-screened detection lists and have a lower likelihood of generating false events.