Automatic classification and onset estimation of seismic P and S wave signals recorded at local seismic network using artificial neural networks

Automatic detection and onset estimation of seismic P and S wave signals are essential in automatic seismic event location systems. The problem of phase classification is studied by utilizing multiple parameters of different types each computed in several frequency bands. P and S wave signals have fundamentally different polarization properties. The input parameters depending on signal polarization in this study included rectilinearity, principal ellipticity, global polarization parameter, eigenresultants, quadratic resultant and predicted coherency. Several statistical parameters were also used. They included skewness, kurtosis and Jarque-Bera test. Variance in time of several of the parameters were added to the input database. Different amplitude ratios were used also. Statistical parameters were computed separately from vertical and horizontal channels. All parameters were computed at 6 different frequency range and time window combinations resulting 210 input parameters. The parameters were computed from 10634 seismic traces of local events creating 2.2M new time-series. Independent training, testing and validation datasets selected from these time-series consisted ~1.5M inputs each. Artificial Neural Networks (ANN) are a robust and efficient tool in classification using large amount of input parameters. A deep ANN with 4 hidden layers was used for classification of the P and S wave signals. Final validation showed that 98% of signals were classified correctly. Magnitudes of the events are usually in range 1.5>ML>0.5 and event to station distance less than 200 km. Similar method was applied to automatic onset estimation. The data set was recorded in local seismic network on Botnian Bay coast, Finland. The network is collecting information of microearthquakes in an area around a planned nuclear power plant.