Using neural networks to discriminate between genuine and spurious seismic events in mines

Careful observation has shown that mining-induced seismicity follows a multimodal distribution, which we assume to arise from many distinct physical processes. The two major modes however, arise from those seismic events that are associated in some way with geological features on the one hand, and those that are associated, among other things, with fracturing in the volume of extreme stress concentrations ahead of the stope faces, on the other. We call the former "genuine" events and the latter "spurious" events. Untangling these modes has been a major problem for those researchers wishing to work with unimodal seismic catalogs. Partial separation of the genuine events from a catalog can be obtained by a careful selection from a scatter diagram of log (radiated seismic energy) against log (scalar seismic moment) or equivalently by selecting a threshold value of magnitude say, from an inspection of the Gutenberg-Richter diagram. This threshold is usually considerably greater than the threshold of completeness that can be achieved by modern seismic networks on mines. The main objective of this paper will be the demonstration that a simple neural network can improve this separation. In this study, for example, simple elimination below the threshold of log (scalar seismic moment) = 9.5 resulted in 206 genuine events remaining in the catalog. After running the eliminated events through a trained neural network, an additional 72 genuine events were found, representing an increase of nearly 35%. This has important consequences for statistical hazard analysis and for the identification of active geological structures in mines.