Cluster Analysis for CTBT Seismic Event Monitoring
Mines at regional distances are expected to be continuing sources of small, ambiguous events which must be correctly identified as part of the Comprehensive Nuclear-Test-Ban Treaty (CTBT) monitoring process. Many of these events are small enough that they are only seen by one or two stations, so locating them by traditional methods maybe impossible or at best leads to poorly resolved parameters. To further complicate matters, these events have parametric characteristics (explosive sources, shallow depths) which make them difficult to identify as definite non-nuclear events using traditional discrimination methods. Fortunately, explosions from the same mines tend to have similar waveforms, making it possible to identify an unknown event by comparison with characteristic archived events that have been associated with specific mines. In this study we examine the use of hierarchical cluster methods to identify groups of similar events. These methods produce dendrograms, which are tree-like structures showing the relationships between entities. Hierarchical methods are well-suited to use for event clustering because they are well documented, easy to implement, computationally cheap enough to run multiple times for a given data set, and because these methods produce results which can be readily interpreted. To aid in determining the proper threshold value for defining event families for a given dendrogram, we use cophenetic correlation (which compares a model of the similarity behavior to actual behavior), variance, and a new metric developed for this study. Clustering methods are compared using archived regional and local distance mining blasts recorded at two sites in the western U.S. with different tectonic and instrumentation characteristics: the three-component broadband DSVS station in Pinedale, Wyoming and the short period New Mexico Tech (NMT) network in central New Mexico. Ground truth for the events comes from the mining industry and local network locations, respectively. The clustering techniques prove to be much more effective for the New Mexico data than the Wyoming data, apparently because the New Mexico mines are closer and consequently the signal to noise ratios (SNR's) for those events are higher. To verify this hypothesis we experiment with adding gaussian noise to the New Mexico data to simulate data from more distant sites. Our results suggest that clustering techniques can be very useful for identifying small anomalous events if at least one good recording is available, and that the only reliable way to improve clustering results is to process the waveforms to improve SNR. For events with good SNR that do have strong grouping, cluster analysis will reveal the inherent groupings regardless of the choice of clustering method.
- Research Organization:
- Sandia National Lab. (SNL-NM), Albuquerque, NM (United States); Sandia National Lab. (SNL-CA), Livermore, CA (United States)
- Sponsoring Organization:
- US Department of Energy (US)
- DOE Contract Number:
- AC04-94AL85000
- OSTI ID:
- 9576
- Report Number(s):
- SAND99-1406C; TRN: AH200124%%403
- Resource Relation:
- Conference: 21st Seismic Research Symposium: Technologies for Monitoring the CTBT, Las Vegas, NV (US), 09/21/1999--09/24/1999; Other Information: PBD: 3 Aug 1999
- Country of Publication:
- United States
- Language:
- English
Similar Records
Reducing Information Overload in Large Seismic Data Sets
Comparison of Single-Fired and Delay-Fired Explosions at Regional and Local Distances,