Knowledge-based interpretation of seismic data in the Intelligent Monitoring System

Abstract

Abstract The Intelligent Monitoring System (IMS) provides a new capability for automated and interactive analysis of data to detect and locate seismic events recorded by a network of seismic stations. IMS integrates emerging technologies in artificial intelligence, database management, computer graphics, and distributed processing into an operational system used for routine bulletin production and associated research tasks. The first version of IMS (Bache et al., 1990a,b; Bratt et al., 1990) was designed for detection and location of regional events recorded by the two high-frequency arrays in Norway (NORESS and ARCESS). This has been extensively revised and expanded to become IMS Version 2 which is designed to detect and locate all seismic events recorded by an arbitrary seismic network. Since March 1991 it has been operated continuously to process the data from four high-frequency arrays (adding FINESA in Finland and GERESS in Germany). For some periods data from as many as seven 3-component stations in Eurasia have also been included in the processing. The most important new element is ESAL (Expert System for Association and Location) which interprets signal detections to form and locate seismic events. It is programmed in the ART expert system shell which provides the knowledge representation framework and inference mechanisms for complex and knowledge-rich rule-based reasoning. The current version of ESAL represents knowledge through approximately 200 ART rules that are configured through about 300 user-specified parameters and tables. The IMS architecture and operational procedures are designed to facilitate acquisition of new knowledge for ESAL. Knowledge acquisition methods being used include: Bayesian analysis, training neural-nets, statistical analysis to estimate parameters configuring rules, computing fuzzy-logic membership functions, and formulating new rules. Only the Bayesian probabilities are discussed in detail here. They provide a compact representation of complex knowledge about station-specific differences in phase characteristics. As an example, we describe the rules used for automated identification of detected regional Sn, Lg, and Rg phases. Using a Bayesian analysis technique, we quantify the differences in S-phase characteristics. The data show that they fall into two classes with GERESS distinct from the three Fennoscandia arrays.