基于神经网络图像特征识别的中国强震台站谱比曲线场地分类

摘要:

随着各国地震观测台网的布设与地震记录的不断积累，采用强震动记录的单台水平/竖向谱比曲线进行场地放大特征分类已在多国强震台网应用. 但是由于场地地震反应的不确定性，谱比曲线分类大多依靠主观判断与经验阈值，存在分类结果不稳定、可重复性差、难以适应大规模数据处理等问题. 为了解决该问题，本研究采用机器学习神经网络图像识别技术，通过提取强震记录谱比曲线形状特征进行中国数字强震台网的台站场地分类. 首先以钻孔资料较为完备和强震台站较为密集的日本KiK-net与K-NET台站作为训练数据集，依据中国抗震规范和30 m土层等效剪切波速 (VS30)进行场地分类并进行标签标记，训练得到了基于广义回归神经网络 (GRNN) 以及卷积神经网络 (CNN)的分类模型. 基于经过质量核查筛选的中国数字强震台网 2008 至 2020 年收录的11078 组三分量强震记录，对满足计算条件的323个强震台站进行谱比曲线图像识别分类，并结合部分台站的钻孔与现场调查资料进行了分类效果验证. 结果表明，基于GRNN的中国抗震规范场地分类结果的总样本准确率，以及基于CNN的VS30场地分类总样本准确率约为60%左右，证明基于日本数据训练的分类模型可在中国应用，具有较好的泛化性. 作为我国台站场地类别主要组成的中国抗震规范 II 类场地达到了65.7%的召回率，以及高达77.4%的精确率. 同样，占台站大部分比例的以VS30为依据的美国抗震规范C、D类场地的预测结果召回率分别达到 85.1% 和61.8%. 场地软硬和放大特征差异较大的III类和I类场地 (D类场地和A + B类场地)的混淆概率较低，证明神经网络图像识别技术较好把握了不同类别场地谱比曲线的形状特征，具有较强的鲁棒性. 本研究的相关结果已应用于开源的中国强震记录数据库项目 (China-Flatfile).

=

Abstract:

With the deployment of seismic observation station networks and the continuous accumulation of seismic recordings in various countries, classification of site amplification features using the Horizontal and Vertical Spectral Ratio (HVSR) curves from strong motion station has been successfully applied in many countries. However, due to the uncertainty of the seismic site response, the classification of spectral ratio curves mainly relies on subjective judgment and empirical thresholds, which leads to unstable classification results, poor reproducibility and difficulty in handling large-scale datasets. In this study, the machine learning and neural-network image recognition technique is used to extract the shape features of spectral ratio curves of strong motion recordings and applied in site classification of National Strong Motion Observation Network System in China. The KiK-net and K-NET stations in Japan, which have complete borehole data and densely distributed stations, are used as the training datasets. The site classification and corresponding labeling are carried out according to China seismic code and 30-meter soil averaged shear wave velocity (VS30). The classification models based on Generalized Regression Neural Network (GRNN) and Convolutional Neural Network (CNN) are obtained after training these data. 11078 sets of three-component strong earthquake recordings ranging from 2008 to 2020 are screened in National Strong Motion Observation Network System. After quality check and screening of the ground motion dataset, 323 stations satisfied the computational requirements are classified based on the spectral ratio method and image recognition technology. The classification performance is verified by borehole drilling data or field survey regarding some strong motion stations. The results show that the overall accuracy of the China seismic code site classification results based on GRNN, as well as the overall accuracy of the VS30 site classification based on CNN is about 60%. It proves that the classification model trained based on the Japanese data can be applied in China and having good generalization performance. The identification of Class II sites in China seismic code, the main component of station sites in China, achieves a recall rate of 65.7%, as well as a precision rate of up to 77.4%. Similarly, the prediction results of C and D sites in the United States seismic code based on VS30, which account for the majority of station sites, reach a recall rate of 85.1% and 61.8%, respectively. Due to the significant differences in site stiffness and amplification characteristics, the confusion probability of Class III and Class I sites (D and A + B sites) is low. It proves that the neural network image recognition technique grasps the shape features of the spectral ratio curves of the different categories of sites, and the classification results is robustness. The relevant results of this study are applied in the open-source China-Flatfile project.