A buried hill fault detection method based on 3D U-SegNet and transfer learning

Abstract

Faults control the formation and development of fault-block buried hill oil and gas reservoirs. Therefore, if the fault can be accurately detected, the well location can be accurately deployed, which is beneficial to improving recovery efficiency and production. Conventional convolutional neural networks (CNNs) have difficulty obtaining a large number of real fault samples. These approaches usually train on artificially synthesized samples, bringing about challenge for accurate prediction of different faults, especially for buried hill faults. To address it, we propose a detection method for buried hill faults based on a 3D hybrid network U-SegNet combined with transfer learning. First, we construct two traditional semantic segmentation networks, i.e., U-Net and SegNet, to form a new 3D U-SegNet network. Next, we train this network using plenty of 3D synthetic seismic data. Finally, we implement transfer learning by model fine-tuning using a few labeled real data. The method is applied to the buried hill field data, and the result proves that the U-SegNet network, after transfer learning, can reduce the cases of fault miss and false detection and can effectively improve the accuracy, which verifies the value of this research for complicated buried hill fault detection.