GNGTS 2014 - Atti del 33° Convegno Nazionale

44 GNGTS 2014 S essione 3.1 shallow manholes, easily influenced by human presence and activity on the square. We assigned 3 votes to all the other channels. During November 2013 - March 2014 we recorded more than 2000 events with different waveforms, duration and frequency content. Our first aim is therefore to analyze and classify these signals and to extract only the ones with a possible relation with fracture processes. Their localization and time evolution can give important information in respect to the most active zones in the rock mass and their progression to failure. First results of signal analysis and classification. The identification and classification of the recorded seismic signals is a major task for the monitoring of the stability at Madonna del Sasso. Each type of signal is related to different source processes, among which we would like to focus on those related to fracturing processes. Only after this classification the spatial localization and temporal distribution of the microseismic events could be used as an objective element to realize an early warning system. In order to achieve this goal, the classification of the recorded events should be done as close to real time as possible, with automatic procedures of analysis and classification of data. For these reasons, researches are involved in the development of automatic robust seismic event discrimination algorithms, enabling to reduce subjectivity and time of analysis and to concentrate only on a reduced number of signals. Many efforts have been made in past years for the automatic recognition of seismo-volcanic events: the recent trend 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. , 2009) scenarios. Several authors have successfully appliedHiddenMarkovModels (HMMs) to continuous volcano-seismic event recognition (Benítez et al. , 2007; Beyreuther et al. , 2008) rivalling in popularity with other techniques such as Artificial Neuronal Networks (ANNs) (Falsaperla et al. , 1996; Scarpetta et al. , 2005) and Support Vector Machines (SVMs) (Masotti et al. , 2006; Giacco et al. , 2009). We are currently analysing manually the recorded events and characterizing them in time and frequency domains (Fig. 2), in order to identify the key parameters on which make reliable distinctions among the nature of each signal and to devise an automatic classification procedure. The considered parameters are signal shape (in terms of amplitude, duration, kurtosis) and frequency content (range of maximum frequency content, frequency distribution in spectrograms). Particularly the kurtosis of the envelope was found to be a crucial parameter for the description of a signal shape (Hilbert et al. , 2014). Kurtosis parameter ( k ) is a quantitative measurement of the flatness or peakedness of a random-variable distribution compared to a normal distribution. For a random variable, it can be expressed as the ratio between the fourth central moment and the fourth power of the standard deviation of the expected distribution: E ( x – μ ) 4 k = –––––––– (1) σ 4 where μ is the mean of x , σ is the standard deviation of x and E(y) represents the expected value of the quantity y . In this way the kurtosis of a normal distribution is 3. Distributions that are flatter than normal have kurtosis values lower than three, sharper-distribution kurtosis is higher than 3. As a first result, we can clearly distinguish four main classes of recorded signals: microseismic events, regional earthquakes, electrical noises and calibration signals, still unclassified events (probably grouping rockfalls, quarry blasts, other anthropic and natural sources of seismic noise). Microseismic events (Fig. 2a) show impulsive and short duration signals which envelope has a triangle shape with a clear coda. The spectrogram has a specific aspect with a sharp energy increase followed by an exponential decay of the high-frequency content with time. These features are in good agreement with those reported by several authors (Burlini et al. , 2007; Helmstetter and Garambois, 2010; Levy et al. , 2011). As a result, they show very short duration (0.8-4 s) and very high kurtosis values (>10). Unlike the high frequency content expected for