Discrimination of the Near-Fault Pulse-Like Seismic Motions Based on Convolutional Neural Network

Identification of near-fault pulse-like seismic motions are an important part of engineering seismic design. In recent years, the rapidly developing deep learning algorithm, with its powerful data feature mining and image recognition ability, can quickly and accurately constrain the attributes of earthquake events. By inputting seismic wave data and seismic time-frequency data, a convolutional neural network (CNN) is used to obtain the waveform and spectral features of the seismic events, thereby constructing an automatic classification network model based on the CNN algorithm for near-fault pulse-like earthquakes and non-near-fault pulse-like earthquakes. Based on the observed data of near-fault pulse earthquakes and non-near-fault pulse-like earthquakes recorded in the NGA-West2 strong-motion database, a deep learning training model is constructed and the model is trained to determine known near-fault pulse earthquakes. The results show that the CNN based on time-frequency data has a high recognition accuracy, with a recognition rate of up to 91.59%.