基于先验约束的深度学习地震波阻抗反演方法

Deep learning seismic impedance inversion based on prior constraints

不同于传统的深度学习反演方法,文中提出一种基于先验约束的深度学习地震波阻抗反演方法:参照地震相类型分割待反演区域,且将区域分割结果作为一种明确的空间约束条件监控网络模型的反演过程;将蕴含丰富低频信息的初始模型作为一种标签以丰富反演结果的低频信息;并使用一种强抗噪性激活函数提高网络模型对噪声数据的适应能力。为降低标签数据的获取难度并保证网络的反演精度,还采取半监督学习方式对网络模型进行训练。将所提方法应用于Marmousi 2模型,测试结果表明反演效果良好且具有较强抗噪性能;随后将该方法成功地应用于M油田实际勘探数据。

We propose a deep learning seismic impedance inversion method based on constraints of prior information.Different from traditional deep learning inversion methods,the inversion area is segmented based on the category of seismic face and segmentation regions are applied as an explicit spatial constraint to constrain the inversion process of the network model.Then the initial model is set as a label to enrich the low-frequency information of the inversion result.Finally,a strong anti-noise activation function is used to improve the adaptability of the network model to noisy data.To reduce the difficulty of acquiring label data and ensure the inversion accuracy of the network,semi-supervised learning is adopted to train the network model.The proposed method is tested on the Marmousi2 model,and the test results indicate that it has a good inversion effect and anti-noise performance.Subsequently,it is successfully applied to the real exploration data of an oilfield.