Application of Neural Networks to Seismic Signal Discrimination Research Findings
Abstract
Research focused on identification and collection of a suitable database, identification of parametric representation of the time series seismic waveforms, and the training and testing of neural networks for seismic event classification. It was necessary to utilize seismic events that had a high degree of reliability for accurate training of the neural networks. The seismic waveforms were obtained from the Center for Seismic Studies and were organized into smaller databases for training and classification purposes. Unprocessed seismograms were not well suited for presentation to a neural network because of the large number of data points required to represent a seismic event in the time domain. The parametric representation of the seismic events in some cases provided adequate information for accurate event classification, while significantly reducing the minimum size of the neural network. Various networks have achieved classification rates ranging from 88 percent classification of three class problem to 75 percent for the 5 class problem. The results vary dependent on the number of classes and the method of parametric transformations utilized. Multiple tests were performed in order to statistically average the training and classification rates. Test summaries presented and individual test results are given in the appendix.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 11, 1994
- Accession Number
- ADA285848
Entities
People
- Barbara Crist
- Don J. Smith
- G. A. Cipriani
- J. J. Fuller
- James A. Cercone
- John D Martin
- Larry Mccutchan
- Stephan Goodman
- V. S. Foster
- W. M. Clark