LSTM神经网络在地震资料初至拾取中的应用

A First Break Picking Method in Seismic Prospecting Using LSTM Recurrent Networks

本文引入RNN的升级算法LSTM神经网络技术,建立了一套海量数据、高精度的自动拾取地震资料初至流程。相比于其他神经网络方法,LSTM神经网络能够提取数据的时序特征,加强学习初至前噪音时序特征,从而提高初至拾取的精度,为地震资料的初至拾取提供一套新的思路。首先设计样本制作过程并建立、训练模型,通过模拟资料验证方法的有效性,应用于胜利油田浅海与西部山地地震勘探资料的初至拾取,取得理想效果,证明LSTM神经网络初至拾取具有较高的精度与适用性。

High precision,high efficiency and high automation are very important in first break picking of massive seismic data.In this paper,LSTM recurrent networks are introduced to establish a set of automatic first break picking process of massive seismic data with high precision.Compared with other neural networks methods,LSTM recurrent networks can extract the time series characteristics of data and enhance the learning and memory of noise characteristics before the first arrival,to improve the accuracy of first break picking.This provides a new idea for first break picking of seismic data.Firstly,the making process of samples is designed and the training models are established.The implementation process of the method is discussed through the simulation data.Then,it is applied to first break picking of seismic data from Shengli Oilfield in both the ocean and western mountainous areas.The results show that the method using LSTM recurrent networks has high accuracy and applicability.