基于MultiRes-Unet神经网络的三维断层识别研究OA北大核心CSTPCD

Faults,which extend for dozens of meters to dozens of kilometers with fault throw changing from a few meters to tens of meters,may exhibit quite different seismic responses,e.g.discontinuous reflections,suddenly increasing or decreasing events,and blank or distorted reflections.Fault responses merely account for a tiny percentage of total seismic responses;this means that fault predictions may be quite snatchy and somewhat inaccurate.A 3D neural network,MultiRes-Unet3D,is a plausible solution to multi-resolution fault characterization.In view of the small proportion of fault responses,a weighted cross-entropy loss function is used in the learning process to balance among different terms and improve the credibility of fault detection.3D synthetic seismic data sets and fault labels are generated through forward modeling.The MultiRes-Unet3D is built,trained,and validated based on Tensor-flow,and then the network model trained is applied to 3D seismic data for fault identification.The results show good spatial conti-nuity of fault identification and credible fault boundary detection.The MultiRes-Unet3D has good generalization performance and could be applied to seismic data with different fault features.This technique can save the cost in time and labor of fault interpreta-tion and yield objective results.