Paper Publications
Current position: Home > Scientific Research > Paper Publications
BuriedHillNet-A three-dimensional deep learning model of identifying buried hill faults based on expert interpretation samples and forward modeling samples
- Release time:2024-06-21
- Hits:
Abstract:
Identifying three-dimensional faults within the buried-hill fracture reservoir is a challenge due to the deep buried nature of the fracture and the relative low signal-to-noise ratio and resolution of seismic data in the study area. Without expert interpretation samples of research area, identifying three-dimensional faults is often not ideal relying solely on theoretical forward modeling labels for training. For obtaining expert interpretation samples, a three-dimensional annotation tool is created with manual expert annotation. An optimized BuriedHillNet network model with strong generalization ability was proposed to improve performance on three aspects (frequency domain,multiscale, and region of interest). Firstly, a three-dimensional cyclic residual convolution and a fast Fourier residual convolution are used to extract the time domain, frequency domain, and spatial domain features of the data. Secondly, dense void convolution is applied to reduce the loss of detail features for enhancing multiscale features by expanding the receptive field. Thirdly, three-dimensional coordinate attention is introduced to focus on the region of interest (ROI) from the beginning. Furthermore, dual-feature residual attention is used to fuse deep and shallow semantic features for solving the problem of the loss of semantic features with upsampling. After that, the trained model is applied to identify the internal fault system of the buried hill in JiNing Sag with the low signal-to-noise ratio data of the isolated island. The ideal identification results verify the effectiveness and applicability of this method. This model offers an efficient and reliable technical approach for identifying internal fault systems within the buried hill.Translation or Not:
no