An STFT-LSTM System for P-Wave Identification

Abstract

The identification of P-waves in data with substantial noises is essential in seismic phase recognition, which is the foundation of seismic localization and an essential part of automatic seismic monitoring and processing. We treat the identification of P-waves as a sequence classification problem and build a long short-term memory (LSTM) network to model the difference between P-waves and noises. LSTM is designed to process time series data and has the ability to model the connections among short and long time steps; such temporal changes are characteristic of P-waves. In addition to temporal changes, spectral characteristics also distinguish between P-waves and noises. To leverage this difference, we introduce the short-time Fourier transform (STFT) at the front-end of the network to retain the implicit information, as well as to reduce the burden of the network. The results show that a deep neural network with STFT fails to function when the P-waves are not strong, but in the proposed STFT–LSTM system, the average recall for the P-wave detection is 90%, and the precision is approximately 75%. These results suggest that the proposed STFT–LSTM system is capable of distinguishing P-waves from background noises.