摘要
断层识别是地震数据解释的重要环节之一。深度学习技术的发展有效提高了断层自动识别的效率和准确性。然而,目前在断层的自动识别任务中,如何准确捕捉断层细微结构并有效抵抗噪声干扰仍然是一个具有挑战性的问题。为此,在HRNet网络的基础上,构建了一种基于解耦自注意力机制的高分辨率断层识别网络模型AHRFaultSegNet。对于自注意力机制解耦,结合空间注意力和通道注意力,代替HRNet中并行传播的卷积层,在减少传统自注意力机制计算量的同时,模型可以在全局范围内计算输入特征的相关性,更准确地建模非局部特征;对解耦自注意力使用残差连接来保留原始特征,在加速模型训练的同时,使模型能够更好地保持细节信息。实验结果表明,所提出的网络模型在Dice、Fmeasure、IoU、Precision、Recall等性能评价指标上均优于其他常见的断层自动识别网络模型。通过对合成地震数据与实际地震数据等进行测试,证明了该方法对断层细微结构具有良好的识别效果并且具有良好的抗噪能力。
Fault recognition is an essential step in seismic data interpretation.The development of deep learning has effectively improved the efficiency and accuracy of automatic fault recognition.However,in automatic fault recognition,it is still challenging to accurately capture subtle structures of faults and effectively resist noise interference.Thus,in this study,we propose a high-resolution fault recognition network model,AHRFaultSeg-Net,based on the HRNet network and decoupled self-attention mechanisms.The decoupling of self-attention mechanisms combines spatial attention and channel attention,replacing parallel convolution layers in HRNet.This reduces the computational amount of traditional self-attention mechanisms while allowing the model to calculate the relevance of input feature on a global scale,thus more accurately modeling non-local features.In decoupled self-attention,the residual connection is employed to preserve the original feature,speeding up model training and better maintaining detailed information.Experimental results demonstrate that the proposed network model outperforms other common automatic fault recognition network models in performance evaluation indexes such as Dice,Fmeasure,IoU,Precision,and Recall.Through fault recognition experiments on synthetic seismic data and actual seismic data,this method is proven to be effective in subtle fault structure identification and robust in noise resistance.
作者
李克文
李文韬
窦一民
朱信源
阳致煊
LI Kewen;LI Wentao;DOU Yimin;ZHU Xinyuan;YANG Zhixuan(Qingdao Institute of Software,College of Computer Science and Technology,China University of Petroleum(East China),Qingdao,Shandong 266580,China)
出处
《石油地球物理勘探》
EI
CSCD
北大核心
2024年第6期1225-1234,共10页
Oil Geophysical Prospecting
基金
国家自然科学基金项目“储层天然气水合物相变和渗流多场时空演化规律”(51991365)
山东省自然科学基金项目“基于多源数据融合的浊积岩有效储层预测方法”(ZR2021MF082)联合资助。
关键词
断层检测识别
深度学习
解耦自注意力机制
残差连接
fault detection and recognition
deep learning
decoupled self-attention mechanism
residual con-nection
