Residual Denoising Diffusion Model for Seismic Fault Recognition Under Data Imbalance

Summary Accurate seismic fault identification is crucial for interpreting complex geological structures and efficiently exploring hydrocarbon resources. However, severe data imbalance in seismic datasets, especially in weak-signal regions and complex geological environments, significantly limits traditional fault recognition methods. To address this issue, this study introduces an innovative Residual Denoising Diffusion UNet (RDD-UNet), which integrates a lightweight 3D residual UNet architecture with a diffusion probabilistic model, reformulating fault detection as a conditional generative task to enhance accuracy and generalization under imbalanced data conditions. The advantage of RDD-UNet lies in its residual denoising diffusion mechanism, accurately capturing the sparse spatial distribution of faults while precisely preserving critical boundary details. The model incorporates depthwise separable convolutions and residual connections, significantly improving computational efficiency for processing large-scale 3D seismic datasets. Experiments involved initial training on synthetic seismic datasets, followed by rigorous evaluations on real-world F3-3D and Kerry-3D seismic surveys. Comparative results demonstrate that RDD-UNet surpasses traditional UNet, ResUNet, and basic diffusion models in terms of accuracy, noise resilience, and generalization, particularly excelling in detecting small-scale faults within complex geological settings. This research pioneers the application of residual denoising diffusion models in seismic fault recognition, providing an efficient and robust solution that significantly advances automated seismic interpretation technologies.