Detecting teleseismic events using artificial neural networks

Abstract

Artificial neural networks (ANNs) have been trained to recognize seismic signal onsets from vertical channel data. ANNs were trained using previously analysed events and noise samples recorded at 3 short period stations in central Finland. Separate nets were trained for each station. Comparisons were made between different net configurations and between different types of neural nets. The input to the nets consisted of four different STA/LTA values computed in seven frequency bands. The training data base was obtained from P-wave signals of 193 teleseismic events. The ANNs were trained to give high output values at onset and low output for noise and coda of events. After training the ANNs could produce a time series in which the signal onsets were shown as sharp peaks. When compared with the Murdock–Hutt detector the ANN detector could find 25% more events of the Reviewed Event Bulletins (REB) of the International Data Center. The detectors were tuned to produce the same total number of detections. When tuned to detect the same number of REB events as the Murdock–Hutt detector, the ANN detector produced over 50% less detections indicating a smaller false alarm rate. The type of ANN used was a multi-layer-perceptron (MLP) with one hidden layer. MLPs with two hidden layers and with a linear output layer were also tested but they clearly gave weaker results. Also, partially recurrent Elman and Jordan networks were tried but they showed a weaker detection capability than MLP. This type of detector could be used as a post-detector processor using outputs of other detectors of different types as its input combining the best features of each of the detectors.