卷积神经网络的半监督层位追踪方法  

A semi-supervised horizon tracking method with convolutional neural network

作　　者：李沐阳 高建虎[1] 雍学善[1] 常德宽 LI Muyang;GAO Jianhu;YONG Xueshan;CHANG Dekuan(Research Institute of Petroleum Exploration&Development-Northwest,PetroChina,Lanzhou,Gansu 730020,China)

机构地区：[1]中国石油勘探开发研究院西北分院,甘肃兰州730020

出　　处：《石油地球物理勘探》2024年第5期938-947,共10页Oil Geophysical Prospecting

基　　金：中国石油集团科学研究与技术开发项目“人工智能地震处理解释新方法研究”(2021DJ3505)资助。

摘　　要：层位追踪是地震资料解释的关键步骤,通常由解释人员以人机交互方式进行,效率较低。卷积神经网络可以构建地震数据和训练标签的非线性映射关系从而完成层位追踪,由于人工解释结果获取困难,仅由少量标签训练的模型泛化能力较差。为此,提出一种基于卷积神经网络的半监督层位追踪方法,将层位追踪转化为层位断层间区域的图像分割。首先使用自编码器对无标签数据进行训练,之后将部分参数迁移至有监督学习网络后使用少量标签数据进行有监督学习,最后对整个工区的地震数据进行预测,提取分割结果边缘作为层位追踪结果。合成数据和实际数据的测试结果均表明,相较于有监督学习层位追踪方法,该方法具有较少的错误分割,由分割边界提取的层位与人工层位解释结果的误差较小,具有更好的泛化能力。Horizon tracking is a key step in seismic data interpretation.It is typically performed manually by interpreters in a human-computer interaction manner,which results in low efficiency.Convolutional neural network(CNN)can establish a nonlinear mapping relationship between seismic data and training labels to achieve horizon tracking.However,since it is difficult to obtain manually interpreted results,models trained merely with a few labels tend to have relatively poor generalization capability.Therefore,a semi-supervised horizon tracking method based on a convolutional neural network is proposed to transform horizon tracking into image segmentation between horizons and faults.First,the unlabeled data is trained by the autoencoder.Then a small amount of labeled data is used for supervised learning after part of the parameters are transferred to the supervised learning network.Finally,the seismic data of the whole working area is predicted,and the edge of the segmentation result is extracted as the horizon tracking result.The test results of both synthetic data and the real data show that compared with the supervised learning horizon tracking method,the proposed method presents less error segmentation and smaller errors between the horizon extracted from the segmentation edge and the artificial horizon interpretation results,and thus has better generalization capability.

关 键 词：层位追踪 地震资料解释 卷积神经网络 半监督学习 图像分割 

分 类 号：P631[天文地球—地质矿产勘探]