瑞雷波频散曲线的深度学习反演方法

瑞雷波频散曲线反演是获取地表横波波速的关键步骤, 现有线性反演方法的效果取决于初始模型的选择, 非线性反演也存在效率低、多解等问题.为了进一步提高瑞雷波频散曲线反演的速度与精度, 受深度学习卓越非线性映射能力启发, 本文提出了瑞雷波频散曲线的深度学习反演方法.文中首先基于近地表速度结构的遍历属性和演化特征的有序性, 提出了约束马尔科夫决策的样本数据构建方法; 然后设计了一种卷积神经网络衔接长短时记忆网络的混合网络结构(CNN-LSTM), 用于构建频散序列数据到速度结构的非线性映射关系, 该网络结构包含了3个局部特征学习模块和1个长短时记忆层; 再利用样本数据对混合网络进行训练; 最后进行反演预测.理论模型试验的频散曲线在无噪与含噪情况下, 拟合的平均相对误差分别不超过5.6%和8.9%, 表明本文所提方法具有较高的计算精度和良好的鲁棒性.最后, 将本文方法应用于2008年汶川M_W7.9地震白鹿镇同震地表破裂带的瑞雷波勘探中, 为其浅表同震变形的局部化效应提供了科学约束.

Deep learning inversion of Rayleigh dispersion curves

The inversion of the dispersion curve of Rayleigh wave plays a vital role in achieving the profile of the shear wave velocity. In terms of the current inversion methods, the linear inversion depends on the initial model; meanwhile, there are still some problems, such as low efficiency and diversity of solutions, existing in the nonlinear inversion. In order to further improve the speed and accuracy of the inversion of the dispersion curve, this study proposes an inversion method based on the deep learning, inspired by its excellent nonlinear mapping ability. Through the constrained Markov decision, we first established a method for building the sample data to guarantee the ergodicity and evolutive orderliness of the near-surface velocity profile. Then, we designed a hybrid network structure named CNN-LSTM, which joints the Convolutional Neural Network (CNN) and Long and Short-Term Memory network (LSTM). The network consists of three modules of local feature learning and one LSTM layer, so that it constructs a nonlinear mapping relationship from the dispersion sequence data to the velocity structure. Next, the hybrid network was trained by means of the massive sample data. Finally, inversion prediction was carried out with the trained network. As for the forward models, the average relative fitting errors are less than 5.6% (without noise) and 8.9% (with noise), respectively, indicating that this study's inversion method may provide arbitrary precision arithmetic and good robustness. In the end, we conducted Rayleigh wave survey to detect the structure of the co-seismic surface fracture zone in Bailu Town induced by the 2008 Wenchuan earthquake (MW7.9). The results are conducive to understand the velocity profile of the fracture zone and analyze the localization effect of the shallow co-seismic deformation.