Abstract:
An Artificial Neural Network (ANN) algorithm has been applied to the automatic phase picking of seismic signal. The aim of this work is to detect and accurately pick the onset time of seismic arrivals for a set of 3-component seismic data using ANN. A variety of features for signal detection and phase identification were analyzed in terms of sensitivity and efficiency. Comparing the performance of each feature in discriminating the P and S wave, eight features were selected as input attributes to ANN in which the first four attributes used for the ANN P-phase picker and next four attributes for ANN S-phase detector. They are as follows: (1) Ratio between short-term average and long-term average (STA\LTA). (2) Rectilinearity, (3) Ratio of maximum to minimum horizontal amplitude (hmxnm), (4) Long-axis incidence angle of polarization ellipsoid (Inc1), (5) Ratio between horizontal power to total power (Rh2t), (6) Planarity, (7) Ratio of horizontal-to-vertical power (Hvratp), (8) Short-axis incidence angle of polarization ellipsoid (Inc3).
These attributes were calculated in the frequency band of 1-10 Hz with a length of 0.5 sec moving window for seismic phase identification and another length of 4 sec moving window for P-wave picking. The detection and phase picking is achieved using Back propagation Neural Network (BPNN). The results of preliminary training and testing with a set of broadband seismic recordings shows that the ANN seismic phase picker can achieve a good performance in phase identification and onset-time estimation. In overall result, 72% correct rate of phase identification has been achieved by the both: the trained ANN P-phase detector and the trained ANN S-phase detector, and 53% of P-wave is precisely picked with onset time error less than 0.1 sec by trained ANN P-phase picker using STA \ LTA algorithm. The algorithms developed in the present study have been tested on the seismic data obtained from seismically active Garhwal Kumaon Himalayan region. The results provide accurate and robust automatic picks on a large experimental data.
