False Event Screening Using Data Mining in Historical Archives

Abstract

Analysts working at the International Data Centre (IDC) in support of treaty monitoring through the Comprehensive Nuclear-Test-Ban Treaty Organization (CTBTO) spend a significant amount of time reviewing hypothesized seismic events produced by an automatic processing system to ensure a high-quality event bulletin, which is then made available to the member states of the CTBT. Such a system is characterized as forming signal detections from the waveforms recorded at the International Monitoring System (IMS) stations, performing association or grouping of the detections, and finally converging on a set of hypothesized events that can account for the groups of detections. For each detection, a suite of features is automatically measured ( e.g. , arrival time, azimuth, signal to noise ratio [SNR], etc.) and recorded in a relational database. These features are then used by the association algorithm to form event hypotheses, where an acceptable hypothesis is one that explains a suite of features (within acceptable tolerances) across the network. Thus, the quality of the automatically built events is directly dependent on the comprehensiveness and quality of these features. A principal characteristic of current monitoring systems is that all events produced by automatic processing must undergo a varying degree of analyst review. Due to the desire to detect any potential nuclear event, such systems are by design very sensitive, and their tuning is heavily biased toward a low probability of missing a legitimate event. This bias comes with the tradeoff of a high rate of hypothesizing an event that later is not included in the final event bulletin ( i.e. , a “false” event). This approach reflects the inherently greater cost of missing a real event versus removing false events and results in analysts screening out approximately half of the events hypothesized during automatic association. Data mining methods, which rely in some form on archives …