应用自注意力机制对抗网络进行海洋多次波压制方法研究

Marine multiple attenuation method based on SA⁃GAN

由于海面和海底两个强波阻界面的存在,海洋地震资料普遍发育强能量多次反射波,海洋多次波衰减贯穿着整个海洋地震资料处理的始终,是影响海洋地震资料成像品质最主要的因素之一。复杂海域情况下的多次波压制往往需要通过多方法多域分步组合的衰减策略,计算耗时而且多域多步骤会造成计算误差的累计,从而影响多次波的衰减效率和精度。为此,提出了一种基于自注意力机制对抗网络(SA-GAN)的海洋多次波压制方法。首先,针对特征数据利用多域分步组合法压制多次波获得标签数据集;其次,在U-Net生成器网络中引入自注意力机制(SA),构建基于SA-GAN网络的多次压制深度学习模型,并进行网络训练;最后,利用训练完备的SA-GAN网络对整体数据进行多次波压制处理。引入SA的U-Net生成器的GAN网络收敛速度快且计算稳定,在地震样本数据集上具有更好的数据泛化能力。与常规方法相比,本文提出的方法只需人工处理少量特征数据,网络训练后便可进行工区大量数据的多次波压制处理,避免了复杂多次波压制多方法串联组合的繁琐过程,为海洋实际地震数据的多次波压制提供了一种高效手段。模型和NH探区深水实际资料处理结果验证了本方法的有效性。

Due to the existence of two strong wave impedance interfaces at the sea surface and the sea bottom,strong multiples are commonly developed in marine seismic data.Marine multiple attenuation runs through the whole process of marine data processing,which is one of the most important factors affecting the imaging qua‐lity of marine seismic data.Multiple attenuation in complex marine conditions often requires many methods step by step in different domains.The calculation is time‐consuming and multi‐domain and multi‐step will cause the accumulation of calculation errors,which will affect the multiple attenuation efficiency and precision.In this pa‐per,a marine multiple attenuation method based on self‐attention generative adversarial networks(SA‐GAN)is proposed.Firstly,label datasets are obtained by suppressing multiples using the method step by step in diffe‐rent domains.Secondly,the self‐attention mechanism is introduced into the U‐Net generator network,and the multiple attenuation deep learning model based on SA‐GAN is constructed,with the network trained.Finally,the SA‐GAN with complete training is used to suppress the whole data.The GAN of the U‐Net generator with SA has fast convergence speed and stable computation,and it has better data generalization ability on seismic sample datasets.Compared with the conventional methods,the proposed method only needs to manually pro‐cess a small amount of feature data,and the network can be trained to perform multiple attenuation of a large number of data in the working area,which avoids the tedious process of multi‐method series combination for com‐plex multiple attenuation and provides an efficient means for multiple attenuation of actual marine seismic data.The model and the actual data processing of the NH deep water exploration area verify the effectiveness of this method.