Recognizing explosion sites with a self-organizing network for unsupervised learning

Abstract A self-organizing neural network model has been developed for identifying mining explosion locations in different environments in Finland and adjacent areas. The main advantage of the method is its ability to automatically find a suitable network structure and naturally correctly identify explosions as such. The explosion site recognition was done using extracted waveform attributes of various kind event records from the small-aperture array FINESS in Finland. The recognition was done by using P–S phase arrival differences and rough azimuth estimates to provide a first robust epicentre location. This, in turn, leads to correct mining district identification where more detailed tuning was performed using different phase amplitude and signal-to-noise attributes. The explosions studied here originated in mines and quarries located in Finland, coast of Estonia and in the St. Petersburg area, Russia. Although the Helsinki bulletins in 1995 and 1996 listed 1649 events in these areas, analysis was restricted to the 380 (ML≥2) events which, besides, were found in the reviewed event bulletins (REB) of the CTBTO/UN prototype international data centre (pIDC) in Arlington, VA, USA. These 380 events with different attributes were selected for the learning stage. Because no `ground-truth' information was available the corresponding mining, `code' coordinates used earlier to compile Helsinki bulletins were utilized instead. The novel self-organizing method was tested on 18 new event recordings in the mentioned area in January–February 1997, out of which 15 were connected to correct mines. The misconnected three events were those which did not have all matching attributes in the self-organizing maps (SOMs) network.