deep learning, phase detection, first-motion polarity, event classification, pre-trained model, Geophysics. Cosmic physics, QC801-809, Medicine (General), R5-920

The application of machine learning, particularly deep learning methods, is becoming increasingly widespread in seismology, achieving near-human accuracy in tasks such as phase detection and event classification. However, most neural network models in seismology currently focus on single tasks. Based on the CSNCD dataset released by the China Earthquake Networks Center, we have developed a bi-directional neural network pre-trained model for single-station data analysis. This model uses three-component seismic waveform data as input and employs convolutional neural networks and bi-directional Transformer models for feature extraction and processing. It not only performs routine tasks such as Pg, Sg, Pn and Sn phase detection, P-wave first-motion direction determination, and event type classification but can also be adapted to other seismic waveform data analysis tasks through transfer learning.