论文详情
地震记录初至拾取质量往往受限于地震数据的复杂性,在陆地和浅海地震数据中尤为明显。为了更高效地拾取初至,提出了一种基于深度卷积生成对抗网络(DCGAN)的地震数据初至拾取方法,其关键在于构建一个适用于地震数据初至拾取的DCGAN,包含生成器与判别器两部分。生成器由一个全卷积神经网络(FCN)构成,用于学习地震炮集数据到初至波之间的特征映射;判别器由一个卷积神经网络(CNN)构成,用于辅助生成器训练。基于DCGAN的初至拾取方法的实现分为三步:数据预处理、网络训练和预测拾取。通过对不同卷积层数的网络结构的对比分析,确定了一个最优的DCGAN结构。一旦DCGAN的训练完成,利用其完成一炮地震数据的初至拾取仅需几秒的时间。将DCGAN方法应用于实际数据初至拾取并与现有初至拾取方法(如长短时窗比(STA/LTA)法和峰度赤池信息量准则(AIC)法)的拾取结果相比较,结果表明基于DCGAN的初至拾取方法的精度更高,能满足生产需要。
First-arrival pickup from seismic data is challenging because of data complexity,especially for land and shallow marine data.In order to effectively pick up the first-arrival time,a method based on a deep convolution generative adversarial network(DCGAN) was developed.The DCGAN is composed of a generator and a discriminator.The discriminator consists of a convolutional neural network(CNN),which is used to aid the training of the generator.The generator consists of a fully convolutional network(FCN),which is designed to learn a mapping to relate the seismic shot gathers to the first-arrivals.The DCGAN for first-arrival picking is realized in three steps:data preprocessing,training,and prediction.An optimized architecture of the DCGAN is determined by evaluating the convolution layers with different numbery.Once the training of the DCGAN is completed,the first-arrival picking using the DCGAN only requires a few seconds per shot gather.Application to real data showed that,compared with common methods such as the STA/LTA and AIC,the DCGAN provided better accuracy,thus it could fulfill the industrial requirements.
基金项目：中国石油化工股份有限公司科研项目(PE19007-6)资助。