Transformer模型和迁移学习在地震P波和噪声判别中的应用研究

Research on the application of transformer model and transfer learning in earthquake P-wave and noise discrimination

准确可靠地区分地震和噪声信号对于地震危险性分析和地震预警至关重要.然而, 无处不在且复杂的噪声信号使这项任务充满挑战.针对中国和日本数据的差异, 本研究在深度学习模型训练过程中采取了不同的策略来区分地震和噪声信号.首先, 鉴于日本数据丰富, 直接训练一个Transformer模型, 该模型在日本的判别准确率为99.82%.其次, 为缓解数据不平衡, 对中国地震数据采用了随机滑动波形窗进行增强.还使用中国数据对预先训练的日本模型进行了微调, 以更好地适应中国数据集.经过微调后, 模型在中国的判别准确率为99.47%.结果表明, 使用原始波形训练的深度学习模型进行地震事件判别时能够取得很高的准确率.此外, 迁移学习模型在门源6.9级地震和漾濞序列震中得到了良好的验证, 表明迁移学习在台网稀疏地区的应用是有效的, 这为地震学和地震预警提供了一种潜在的方法.

It is crucial for seismic hazard analysis and earthquake early warning (EEW) to discriminate earthquake and noise accurately and reliably. However, ubiquitous and complex noise signals make this a challenging task. Aiming at the differences between Chinese and Japanese datasets, this study adopts different strategies to discriminate earthquake and noise signals during the deep learning model training process. Firstly, because of the abundant data in Japan, we directly trained a Transformer model and the model has a discrimination accuracy of 99.82% in Japan. Secondly, to alleviate the data imbalance, the Chinese seismic data were augmented by a random sliding window of the waveform. We also fine-tuned the pre-trained Japanese model using Chinese data to better fit the Chinese dataset. After fine-tuning, the model has a discrimination accuracy of 99.47% in China. The results demonstrate that the deep learning model trained by the raw waveform can achieve high accuracy in seismic event discrimination. In addition, the transfer learning model has been well tested in the Menyuan M6.9 earthquake and Yangbi sequence earthquakes, which indicates that transfer learning is effective when applied to areas with sparse seismic networks. This provides a potential method for seismology and EEW.