Artificial Neural Networks for Seismic Data Interpretation

Abstract : This is the second Semiannual Technical Summary of the MIT Lincoln Laboratory Artificial Neural Networks for Seismic Data Interpretation project. The effort during this period has concentrated upon phase labeling and event recognition networks for use in the DARPA/NMRO Intelligent Monitoring System (IMS) for seismic surveillance. Perception networks were developed with standard IMS parameters as inputs, improving upon the phase labeling performance of the expert IMS. These networks were developed and tested using data sets containing 5,000 to 10,000 arrivals. An initial version of the expert system achieved a 79% success rate compared with 86% for the neural network. A more recent and improved version of the expert system achieved rates of 87% compared with 90% for an updated neural network. Phase labeling experiments with sonograms and three-component autoregressive modeling for signal representation did not lead to further improvement. Two event labeling experiments were carried out using three-component autoregressive signal models with Radial Basis Function classification networks and involved on the order of 200 events. Success rates were 96.6% for an earthquake/nonearthquake classification experiment and 91% for a Kola Peninsula event recognition experiment. These results, using only autoregressive waveform representations, are encouraging but very preliminary.