基于Swin Transformer的地震相识别模型  

Seismic facies identification model based on Swin Transformer

作　　者：硕良勋 李志轩 柴变芳 王天意 郑晓东 SHUO Liangxun;LI Zhixuan;CHAI Bianfang;WANG Tianyi;ZHENG Xiaodong(School of Information Engineering,Hebei GEO University,Shijiazhuang,Hebei 050031,China;School of Earth Science,Hebei GEO University,Shijiazhuang,Hebei 050031,China;Intelligent Sensor Network Engineering Research Center of Hebei Province,Sijiazhuang,Hebei 050031 China)

机构地区：[1]河北地质大学信息工程学院,河北省石家庄市050031 [2]河北地质大学地球科学学院 [3]河北省智能传感物联网技术工程研究中心

出　　处：《天然气工业》2024年第12期63-72,共10页Natural Gas Industry

基　　金：河北省重点研发计划项目“基于时空大数据及深度学习的地质灾害风险识别关键技术研究(编号:22375415D)”;河北省石家庄市科技局项目“融合频谱分析和深度学习的地震反射模式提取及地震解释研究”(编号:241790607A)。

摘　　要：地震相识别是油气勘探开发过程中的一项重要技术,但该技术长期存在方法模型训练速度较慢、预测耗时、解释结果人为主观性较强,以及各层特征提取忽略多尺度特征等问题。为此,针对目前地震相识别精度不够且计算成本高的问题,构建了一个基于Swin Transformer的地震相识别模型(Seismic Facies Identification based on Swin Transformer,SFI-ST),首先联合卷积神经网络,利用编码器和解码器不断捕捉地震相细节特征,然后采用两种不同的数据集测试并评估模型的有效性,同时考虑到数据集划分对模型的影响,针对不同划分比例进行性能分析对比,最后对模型进行了消融实验以及抗噪性分析。研究结果表明:①编码器使用的Swin Transformer模块具有较好的特征提取能力,基于较小移动窗口进行特征提取的策略保证模型更快地学习高分辨率地震剖面特征,在各移动窗口使用自注意力机制计算特征的方法保证模型在较大视野下更准确地提取局部特征;②Swin Transformer使用逐层特征融合的方式,在提升特征提取速度的同时保证模型获取更多尺度的特征;③融合Swin Transformer和卷积神经网络模块实现各层特征提取,增强了模型对轮廓、边缘的提取能力。结论认为:①SFI-ST模型应用于两工区数据上的平均交并比分别为73.2%和77.6%,相较于其他主流深度学习算法至少分别提升了10.7%和6.0%,SFI-ST模型运行时间分别为0.62 h和2.88 h,相较于其他主流深度学习算法至少减少了15.1%和24.2%;②SFI-ST模型一定程度上解决了现有地震相智能识别方法识别速度慢、精度低等问题,为地震相识别提供了新方法,在技术上助力了油气勘探开发进程。Seismic facies identification is an important technology in the process of oil and gas exploration and development,but it has been facing several problems for a long time.For example,its method model consumes more time in training and prediction and provides interpretation results of strong subjectivity,and the multi-scale features are neglected in the feature extraction of each layer.To deal with low seismic facies identification accuracy and high calculation cost,this paper establishes a Seismic Facies Identification model based on Swin Transformer(SFI-ST).Firstly,the SFI-ST model,combined with the convolutional neural network,captures the detail features of seismic facies continuously using encoder and decoder.Then,the effectiveness of the model is tested and evaluated by using two kinds of data sets.Additionally,considering the influence of data set partition on the model,the performances at different partition proportions are comparatively analyzed.Finally,an ablation experiment and anti-noise analysis are conducted on the model.The following results are obtained.First,the Swin Transformer module used in the encoder has a better capability of feature extraction.The feature extraction strategy based on a small moving window ensures the model to learn the features of high resolution seismic profiles faster,and the application of the self-attention mechanism to compute the features in each moving window ensures the model to extract the local features more accurately under a larger view.Second,in the mode of layer-by-layer feature fusion,the Swin Transformer improves the feature extraction speed while ensuring the model to extract multi-scale features.Third,the fusion of Swin Transformer and convolutional neural network module realizes the feature extraction at each layer,and strengthens the model's capability of extracting outlines and edges.In conclusion,the mean intersection over union of SFI-ST model for the data of two working areas is 73.2%and 77.6%,which is at least 10.7%and 6.0%higher than that of other

关 键 词：地震相识别 语义分割模型 Swin Transformer 多尺度特征 油气藏预测 

分 类 号：TE19[石油与天然气工程—油气勘探]