Fault recognition using 3D convolutional neural network with global information extraction
Received date: 2023-08-20
Online published: 2024-09-29
Copyright
Accurate fault identification is crucial to oil and gas exploration and development. Traditional fault identification technology based on coherence volume attribute has poor effects in complex structural zones. Conventional convolutional neural network based on image segmentation is also difficult to make up for the feature information lost in down sampling. Therefore, building a global information extraction attention mechanism can not only introduce information extraction in the concatenation part of the U-Net full convolutional network structure, compensates for the lack of information in the downsampling process and enhances the network's learning ability. It can also enhance the bottom level feature information and improve interpretation accuracy by using information scaling at the bottom level of the network. Moreover, this attention block does not add additional parameter information and has a low memory requirement. The experimental results show that the test accuracy of the neural network model with attention mechanism reaches 96%, and the loss function converges to 7%. The description of the main fault of the actual seismic data is better than the conventional U-Net. The attention mechanism of global information extraction provides a new idea for 3D fault intelligent recognition based on convolutional neural network.
LongLong QIAN , BinPeng YAN . Fault recognition using 3D convolutional neural network with global information extraction[J]. Progress in Geophysics, 2024 , 39(4) : 1532 -1543 . DOI: 10.6038/pg2024HH0319
图11 (a) 标签;(b)U-Net识别结果;(c)GIE-Net识别结果Fig 11 (a)Label; (b)U-Net recognition result; (c)GIE-Net recognition result |
表1 不同模块对U-Net模型的影响Table 1 The impact of different modules on U-Net model |
| 模块 | 评价指标 | ||||||||
| IS | IE | IOU | Accuracy | Precision | Recall | Dice | Time | ||
| U-Net | 0.7252 | 0.9785 | 0.8285 | 0.8534 | 0.8407 | 70 | |||
| √ | 0.7339 | 0.9792 | 0.8288 | 0.8651 | 0.8466 | 93 | |||
| √ | 0.6459 | 0.9718 | 0.7942 | 0.7758 | 0.7849 | 98 | |||
| √ | √ | 0.6888 | 0.9752 | 0.8036 | 0.8283 | 0.8158 | 126 | ||
表2 与其他网络模型的对比Table 2 Comparison with other network models |
| 模型 | IOU | Accuracy | Precision | Recall | Dice | Time |
| GIE | 0.6888 | 0.9752 | 0.8036 | 0.8283 | 0.8158 | 126 |
| CAM | 0.7164 | 0.9776 | 0.8174 | 0.8529 | 0.8348 | 105 |
| CBAM | 0.6636 | 0.9726 | 0.7829 | 0.8132 | 0.7978 | 252 |
图13 实际地震数据断层识别(a)标签;(b)CAM识别结果;(c)CBAM识别结果;(d)GIE-Net识别结果. Fig 13 Fault identification of actual seismic data (a) Label; (b) CAM identification results; (c) CBAM identification results; (d) GIE-Net recognition results. |
感谢伍新明教授在开源软件上做的贡献,感谢审稿专家提出的修改意见和编辑部的大力支持.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
/
| 〈 |
|
〉 |