Automatic seismic event recognition and later phase identification for broadband seismograms

Abstract

Knowledge of the patterns of frequently observed seismic phases associated with specific distances and depths have been well developed and applied by seismologists (see, e.g., Richter, 1958; Kulhánek, 1990). However, up till now, the expertise of recognizing seismic event patterns for teleseisms has not been translated into automatic processing procedure. A new approach is developed to automate this kind of heuristic human expertise in order to provide a means of improving preliminary event locations from a single site. An automatic interpretation system exploiting three-component broadband seismograms is used to recognize the pattern of seismic arrivals associated with the presence of a seismic event in real time accompanied by an identification of the individual phases. For a single station, such a real-time analysis can be used to provide a preliminary estimation of the location of the event. The inputs to the interpretation process are a set of features for detected phases produced by another real-time phase analyzer. The combinations of these features are investigated using a novel approach to the construction of an expert system. The automatic system exploits expert information to test likely assumptions about phase character and hence epicentral distance and depth. Some hypotheses about the nature of the event will be rejected as implausible, and for the remainder, an assessment is given of the likelihood of the interpretation based on the fit to the character of all available information. This event-recognition procedure provides an effective and feasible means of interprating events at all distances, and characterizing information between hundreds of different possible classes of patterns even when the observation is incomplete. The procedure is based on “assumption trees” and provides a useful tool for classification problems in which a number of factors have to be identified. The control set of expert knowledge used in testing hypotheses is maintained separately from the computational algorithm used in the assumption search; in consequence, the information base can be readily updated.