基于深度嵌入网络的地震相聚类技术

Seismic facies clustering technology based on deep embedding network

早期基于机器学习的地震相聚类分析依赖地震属性种类的选择与组合,结果具有很强的主观性,而数据驱动下的深度学习可以规避该缺陷。因此,利用深度学习技术,采用自编码网络架构,通过嵌入编码(Embedding Code)对地震数据进行抽象表示;引入聚类损失函数与重建损失函数,建立联合损失函数并优化,使学习到的地震特征既能重建地震数据,又具有较好的聚类能力。鄂尔多斯盆地A致密气探区实际应用结果表明:经过500次迭代后,嵌入编码已具有明显的聚类特征,同时能很好地恢复原始地震信号,相对误差小于5%;与均方根振幅属性相比,基于深度嵌入网络的地震相聚类技术计算的地震相图刻画河道更准确、细节更丰富;比K-Means聚类算法预测结果的井震符合率更高,可达89.3%。

The early seismic facies clustering based on machine learning relies on the selection and combination of seismic attributes,leading to strong subjectivity of results.Nevertheless,this defect can be overcome by data-driven deep learning.Therefore,with deep learning technology,the autoencoder network is adopted to generate embedding code that can be used for abstract representation of seismic data.The clustering loss function and the reconstruction loss function are introduced to build a combined loss function,which is then optimized so that the seismic features learned can not only be used to reconstruct seismic data but also have favorable clustering ability.The proposed method is applied to a tight gas exploration area A in Ordos Basin.The following observations are drawn from the results:After 500iterations,the embedding code has noticeable clustering features,and the original seismic signals can be well reconstructed with a relative error of less than 5%;compared with that in the case of the root-mean-square(RMS)amplitude attribute,the seismic facies map calculated by the seismic facies clustering technology based on deep embedding network delineates channels more accurately with richer detail;compared with the K-means clustering algorithm,the proposed technology delivers a prediction result that has a higher coincidence rate between seismic data and well logging data,which can reach 89.3%.