The development of remaining oil plays an important role in increasing the late production, and low-grade faults seriously impacts residual oil exploration and development. Low-grade faults have a small fault displacement, brief extension, and strong concealment, which hinders their prediction using the typical semantic segmentation network. To intelligently identify low-grade faults, we designed a codec target edge detection technique. For the network to fully learn the low-grade faults information, we constructed the encoder using dilated convolution. Next, we introduced an attention mechanism to the decoder to improve capture of location information from a shallow network and semantic information from a deep network. Finally, the multi-scale fusion decoder outputs fault information of different scales, which further improves the identification accuracy of low-grade faults. The training model is applied to simulated data and actual seismic data through ablation experiments. The results show that this method can effectively identify low-grade faults and overcomes problems such as blurred fault cross-location, thicker edge contour lines, lower detection accuracy, and less training data. Compared with conventional a Holistically-Nested Edge Detection (HED) and a semantic segmentation (UNet), fault misidentification is reduced, fault continuity is increased, and fault accuracy is improved. providing technical support for the exploration and development of remaining oil and increasing the recovery rate of old oil fields.

Low-grade faults play an important role in controlling oil and gas accumulations, but their fault throw is small and difficult to identify. Traditional low-grade fault recognition methods are time-consuming and inaccurate. Therefore, this study proposes a combination of a simulated low-grade fault sample set and a self-constructed convolutional neural network to recognize low-grade faults. We used Wu’s method to generate 500 pairs of low-grade fault samples to provide the data for deep learning. By combining the attention mechanism with UNet, an SE-UNet with efficient allocation of limited attention resources was constructed, which can select the features that are more critical to the current task objective from ample feature information, thus improving the expression ability of the network. The network model is applied to real data, and the results show that the SE-UNet model has better generalization ability and can better recognize low-grade and more continuous faults. Compared with the original UNet model, the SE-UNet model is more accurate and has more advantages in recognizing low-grade faults.

Deep learning shows great promise for seismic parameter inversion, but its success hinges on large, high-quality datasets, which are often scarce in this field. To address this, we propose a novel method for constructing a subsurface parameter model, which can be used to build a training dataset. Our approach leverages kernel density estimation to capture the distribution characteristics and inherent relationships between P-, S-wave velocity, and density. We then use the Metropolis–Hastings sampling algorithm to generate reference points and apply transformations to create realistic parameter models. By constructing parameter models based on actual data distributions, this method inherently avoids the inclusion of abnormal samples, making it particularly beneficial for training inversion neural networks. Unlike conventional approaches that focus on single-velocity influences, our method considers the interplay between multiple parameters, resulting in a more realistic representation. This approach effectively addresses data scarcity, mitigates data leakage, and provides a robust foundation for deep-learning-based inversion.