Parallel System Architecture (PSA): An efficient approach for automatic recognition of volcano-seismic events

a r t i c l e i n f o Automatic recognition of volcano-seismic events is becoming one of the most demanded features in the early warning area at continuous monitoring facilities. While human-driven cataloguing is time-consuming and often an unreliable task, an appropriate machine framework allows expert technicians to focus only on result analysis and decision-making. This work presents an alternative to serial architectures used in classic recognition systems introducing a parallel implementation of the whole process: configuration, feature extraction, feature selection and classification stages are independently carried out for each type of events in order to exploit the intrinsic properties of each signal class. The system uses Gaussian Mixture Models (GMMs) to classify the database recorded at Deception Volcano Island (Antarctica) obtaining a baseline recognition rate of 84% with a cepstral-based waveform parameterization in the serial architecture. The parallel approach increases the results to close to 92% using mixture-based parameterization vectors or up to 91% when the vector size is reduced by 19% via the Discriminative Feature Selection (DFS) algorithm. Besides the result improvement, the parallel architecture represents a major step in terms of flexibility and reliability thanks to the class-focused analysis, providing an efficient tool for monitoring observatories which require real-time solutions. The analysis of volcanic seismicity is the most widely used method for monitoring volcano activity. Precursory seismicity allows the state of a volcano in an eruptive episode to be uncovered (Chouet, 1996). The main tasks carried out in a volcanic observatory such as early warning alerts, development of risk plans, source location, earthquake cataloguing or any seismic study, require the prior detection and classification of volcanic events. Traditionally, signal cataloguing has been carried out by expert technicians who manually classify events based on their knowledge and experience. These catalogues contain a record or label for each event saving information such its arrival time, duration and type. However, this manual classification has to deal with multiple factors that can negatively affect the accurate labelling of the data. The huge amount of records that modern seismometers can gather, low signal to noise ratio (SNR) events, the increase of the activity prior to an eruption requiring fast decisions, etc. (Ibáñez et al., 2009). Therefore, the recent trend in modern observatories is to complement the human work with automatic recognition systems providing support in early warning (Aspinall et al., 2006) or continuous volcano monitoring (Cortés et al., 2009b) …