Automatic 3D fault segmentation based on multi-scale feature fusion model with compound loss function

Faults serve as oil and gas storage space and transportation channels, so fault identification is significant to oil and gas exploration. Fault extraction methods based on manual identification or seismic body attributes are prone to recognition errors due to human factors or poor data quality. With the development of deep learning, researchers have proposed different network models to extract 3D faults. However, the traditional models still have room for improvement in fine-grained segmentation results and model robustness. Therefore, this study proposes a new multi-scale feature fusion network architecture named MAR-UNet. In order to solve the defect of insufficient fine granularity of traditional model segmentation results, this paper designs a local feature extraction module named Residual Sampling Convolution block (RSC block) and deploys it to MAR-UNet; at the same time, in order to improve the defect that the existing 3D model cannot effectively deal with complex spatial relationship features, this study designs a plug-and-play attention module named Mix Attention Mechanism (MAM) in the model. Finally, this paper proposes a compound loss function named Weight Focal-Dice loss for the model's weak robustness caused by sample imbalance. The results of ablation and cross-experiments show that the loss function proposed in this paper is suitable for accomplishing the fault segmentation task under the influence of sample imbalance, and the model proposed in this paper still shows good reliability and robustness when deploying the model to the actual workspace data.