基于对偶贝叶斯U-Net的波阻抗不确定性反演方法研究

Research on acoustic impedance uncertainty inversion method based on dual Bayesian U-Net

1.中国石油大学(北京)油气资源与探测国家重点实验室,北京102249;
2.中国石油大学(北京)海洋石油勘探国家工程实验室,北京102249;
3.长江科学院水利部岩土力学与工程重点实验室,湖北武汉430010;
4.中海石油(中国)有限公司海南分公司,海南海口570100

1. State Key Laboratory of Petroleum Resources and Prospecting,China University of Petroleum (Beijing),Beijing 102249,China;
2. National Engineering Laboratory of Offshore Oil Exploration,China University of Petroleum (Beijing),Beijing 102249,China;
3. Key Laboratory of Geotechnical Mechanics and Engineering of Ministry of Water Resources,Changjiang River Scientific Research Institute,Wuhan 430010,China;
4. Hainan Branch Company,China National Offshore Oil Corporation,Haikou 570100,China.

传统的深度神经网络通常只能实现确定性的预测,无法对反演结果进行不确定性分析,即无法对反演结果的可靠性进行评价。为实现标签数据不足条件下精确的波阻抗反演和对反演结果的不确定性分析,提出了一种基于对偶贝叶斯U-Net的波阻抗不确定性反演方法。首先,开展基于对偶贝叶斯U-Net、前沿深度学习反演方法和传统不确定性反演方法的模拟数据实验,对比分析3种方法的反演精度。然后,将对偶贝叶斯U-Net和传统不确定性反演方法的反演可靠性和对于含噪数据反演的鲁棒性进行对比分析。最后,将对偶贝叶斯U-Net应用于实际地震资料波阻抗反演中。模拟数据实验结果表明,该方法的对偶贝叶斯U-Net在少量标签数据条件下具有较高反演精度并对含噪数据反演有较强鲁棒性。此外,不确定性分析表明,该方法的反演结果可靠性强。实际数据测试结果表明,对偶贝叶斯U-Net能在实际工区数据反演中获得合理并可靠的反演结果。

A traditional deep neural network can achieve a high-precision deterministic inversion of acoustic impedance,without uncertainty analysis of the results.This means that the reliability of the inversion cannot be evaluated.To solve this problem,a novel approach called dual Bayesian U-Net is proposed for acoustic impedance uncertainty inversion.This method performs well in the context of a small number of labels.Moreover,it realizes uncertainty analysis on inversion results.The experiments on synthetic data demonstrate that dual Bayesian U-Net could rival additional deep neural networks in the accuracy of inversion and is superior to traditional uncertainty inversion methods in accuracy,robustness,and reliability,particularly for noisy data.Field data tests show that the proposed method can obtain reasonable and reliable inversion results.

国家重点研发计划(2019YFC0312003)、中国石油天然气集团有限公司“物探应用基础实验和前沿理论方法研究”(2022DQ0604-04)和中国石油天然气集团有限公司中国石油大学(北京)战略合作科技专项(ZLZX2020-03)共同资助。

10.12431/issn.1000-1441.2025.64.01.011