深度域自适应加权多模态多任务学习声阻抗反演

Depth domain adaptive weighted multimodal multitask learning acoustic impedance inversion

以往基于深度学习的叠后地震声阻抗反演通常仅限于利用单道地震数据,当地震数据中存在较强噪声或地下介质横向变化较大时反演结果中会面临横向连续性差和层位区分不清晰的问题.本文提出了一种深度域自适应加权多模态多任务学习声阻抗反演方法,采用GPU加速运算,在阻抗剖面的估计过程中加入井位置附近的空间背景信息,并在数据输入端加入初始阻抗模型.构建的网络包括阻抗反演和地震数据重建两个任务,网络训练过程中采用自适应权重调整策略能够同时优化各自输出的损失,实现数据增强,缓解网络的过拟合,提高网络的泛化能力;相比于传统最小二乘阻抗反演方法,该方法反演速度更快.针对井数据偏少的情况,本文引入其他可用训练数据并采用联合学习策略改善反演结果.模型和实际数据测试表明该方法能够在含有噪声的地震数据和测井数据有限的情况下估计深度域声阻抗,反演结果的横向连续性得到了明显的改善.

In the past,the post stack seismic acoustic impedance inversion based on depth learning was usually limited to using single trace seismic data.When there is strong noise in the seismic data,or there are few training well data,the inversion results will suffer from poor horizontal continuity and unclear horizon discrimination.In this paper,a depth domain adaptive weighted multimodal multitask learning acoustic impedance inversion method is proposed.GPU acceleration is employed for faster computation.In the estimation process of impedance profiles,spatial background information near well locations is incorporated,and an initial impedance model is added at the data input stage.The network includes impedance inversion task and seismic data reconstruction task.The adaptive weight adjustment strategy can optimize the loss of each output at the same time in the network training process.Data enhancement is realized,the overfitting of the network is alleviated,and the generalization ability of the network is improved.Compared with the traditional least squares impedance inversion method,the inversion speed of this method is faster.In view of the lack of well data,this paper introduces other available data and adopts joint learning strategy to improve the inversion results.The model and actual data tests show that this method can estimate the depth domain acoustic impedance with the noisy seismic data and limited logging data,and the horizontal continuity of the inversion results has been significantly improved.