GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 1.1 - POSTER GNGTS 2023 Fig. 2 – Architecture of the CFM: the deep Convolutional neural network for First Motion polarity classification. We decided to feed the network with 1.6-s-long time windows. We demeaned each window subtracting the mean of the 4s preceding the P-wave arrival. We emphasized the portion of the initial wave by setting a cutoff threshold in order not to neglect any of the smaller oscillations arising from the signal. The threshold is different for each seismogram. To decide it, the amplitude of the largest peak among those that preceded the arrival time by at least 0.05 seconds was considered in each trace. The threshold coincides with 20 times the value of this amplitude. Each seismogram was normalized to its respective threshold value. The portion of the signal exceeding this threshold was cut off. Layer Stage Channels Kernel size 0 Input - - 1 Conv1D 32 5 2 Conv1D + Pooling 64 4 3 Conv1D + Pooling 128 3 4 Conv1D 256 5 5 Conv1D + Pooling 128 3 6 Fully connected - - 7 Output - - Tab. 1 – Summary of the CNN used Results We trained the network with approximately 130 ˙ 000 waveforms, the portion of Dataset_1 outside the orange and magenta boxes in Fig. 1a. The portions in the boxes have been used as validation