Seismic signal pattern recognition

In this study several artificial neural network pattern recognition techniques have been developed and implemented to discriminate between natural earthquakes and underground explosions using entire seismic signatures recorded at regional distances. Seismic events collected from the regional array (NORESS) in Norway and the single station (WMQ) in China are used for the purpose of testing various seismic signal discrimination methods. The starting point of this study is the transformation of observed time-domain seismic signals into spectrum-normalized and noise-corrected frequency-velocity or frequency-slowness spectral images. Artificial neural network (ANN) pattern recognition models are then designed and applied to these noise-corrected composite seismic images.;Recent developments indicate that artificial neural networks are appropriate for solving difficult problems in signal discrimination and classification. To test the ANN models, the entire composite seismic images or feature images are used as input to the ANN models. Thus, the geophysical information content of the entire seismic image can be fully utilized. Several design strategies using neural network pattern recognition methods for seismic event identification are investigated and applied to actual data from several geographical regions. These include multilayer perceptron, image compression neural network and reference image identification. The ANN techniques for seismic event identification are the principal focus in this study.;For data sets of 11 natural earthquakes and 11 mining (chemical) explosions recorded by the NORESS array in Norway and 15 natural earthquakes and 15 nuclear explosions recorded by the single station WMQ in China, each event group (earthquake or explosion) is found to be separable by the ANNs in the feature space chosen. The recognition results of three different neural network methods to seismic event identification show that these neural network approaches are all very effective and suitable for near-real-time, automatic event recognition. They have the advantages that entire seismic signatures rather than small subsets of observations are used in the recognition and once trained, the ANNs can be applied automatically, eliminating or minimizing the need for human intervention in the identification process.