【摘要】 破坏性地震对高速铁路安全会造成严重的影响。为使高速铁路地震预警系统快速获得可靠的Ⅰ级地震警报，基于日本K-NET台网记录的强震动数据，提出一种基于震级阈值(M=5.5)和峰值地震加速度阈值(PGA=40 cm/s~2)的深度学习高速铁路现地阈值地震预警方法。设计一种结合卷积神经网络、循环神经网络和注意力机制的混合神经网络架构，并在该网络架构的基础上，训练两个独立的模型，分别用于预测震级和峰值地震加速度是否超过阈值。同时，根据震级阈值和峰值地震加速度阈值设置4个警报预测级别。基于混合神经网络模型对震级和峰值地震加速度的预测，每个台站将发布相应的警报预测级别。警报预测级别4表示台站附近有高震级事件且存在潜在破坏。在P波到达后3 s,对于相同的测试数据集，和基线模型相比，混合神经网络模型对于震级预测和峰值地震加速度预测有更好的性能，且震级预测的准确率达到97.26%,PGA预测的准确率达到98.89%,警报预测级别的准确率到达96.31%。将该方法应用于2021年2月13日发生于日本福岛的7.3级地震，结果表明，在P波到达后10 s内，警报预测级别的准确率达到90%,平均预警时间超过19 s。更多还原

【Abstract】 Destructive earthquakes can have a serious impact on the safety of high-speed railways. To quickly and reliably obtain first-level earthquake alerts for the high-speed railway earthquake warning(EEW) system, based on the strong motion data recorded by the K-NET network in Japan, this paper proposed a deep learning on-site threshold EEW method for high-speed railway based on the magnitude threshold(M=5.5) and peak ground acceleration threshold(PGA=40 cm/s~2). A hybrid neural network architecture was designed combining convolutional neural network, recurrent neural network and attention mechanism. Based on the architecture, two separate models were trained to respectively predict the magnitude and PGA exceeding the thresholds. Meanwhile, four alert prediction levels were set based on the magnitude threshold and PGA threshold. Based on the predictions of magnitude and peak ground acceleration by the hybrid neural network models, each station will issue corresponding alert prediction levels. Alert prediction level 4 indicates the presence of a high magnitude event and potential damage near the station. At 3 s after the arrival of the P-wave, for the same test dataset, compared with the baseline models, the hybrid neural network models have better performance in magnitude prediction and peak ground acceleration prediction, with the accuracy of magnitude prediction of 97.26%, peak ground acceleration prediction prediction accuracy of 98.89%, and alert prediction level accuracy of 96.31%. To verify the reliability of the proposed method, the method was applied to the Off Fukushima earthquake(M=7.3) that occurred on February 13, 2021 in Japan. The results show that within 10 s after the arrival of the P-wave, the accuracy of the alert prediction level reaches 90%, with the average warning time exceeding 19 s.更多还原