基于卷积神经网络和叠加速度谱的地震层速度自动建模方法

CMP道集NMO叠加速度分析拾取的时间速度对不仅受到水平层状介质假设的限制,而且在复杂构造低信噪比数据的适用性方面受到限制。提出了基于卷积神经网络和叠加速度谱的地震层速度自动建模方法,不拾取时间速度对,而是将速度谱作为神经网络的输入数据,将时间域层速度作为标签数据,通过模拟大量随机速度模型和加入随机噪声建立强化测试集,基于L1正则化对卷积神经网络进行训练,得到可直接将速度谱映射为时间域层速度的神经网络模型。将时间域层速度作为标签数据可以增强速度谱和速度模型的空间匹配,使得速度谱与速度模型的空间映射更加紧密和有效。将速度谱作为神经网络模型的输入数据,代替了速度谱时间速度对的拾取,能够较好地克服复杂构造、噪声干扰对速度谱能量团聚焦性的影响。大量随机速度模型和随机噪声强化测试集,增强了深度学习速度建模网络的泛化能力和实用性。模型数据和实际资料测试结果表明,该卷积神经网络模型能够适应复杂低信噪比地震资料的自动速度建模,建模精度与人工拾取结果相当,建模效率提高100倍以上。

The pick-up of time-velocity pairs of common mid-point gathers by normal moveout stacking velocity analysis is limited by the assumption of horizontally-layered media.Its applicability to complex structures with a low signal-to-noise ratio (SNR) is therefore limited.To address this issue,a method for the automatic modeling of seismic interval velocity based on a convolutional neural network (CNN) and a stacking velocity spectrum was investigated.In this method,the velocity spectrum was an input to the CNN,and the interval velocity values in the time domain were used as label data.By testing several random velocity models and random noise levels,a CNN model that can directly map the velocity spectrum to the interval velocity in the time domain was established based on L1 regularization.The use of the interval velocities as label data could enhance the spatial matching of the velocity spectrum and velocity model,thus achieving improved spatial mapping.Rather than selecting time-velocity pairs in the velocity spectrum,inputting the latter into a CNN model could reduce disturbance due to structural complexity as well as the noise interference on the focalization of the energy group in the velocity spectrum.The numerous tests that were performed demonstrated the strong generalization ability and practical value of deep learning modeling.Both the numerical model and tests on actual data showed that automatic velocity modeling based on CNN is suitable for complex structures with low SNR.Furthermore,the modeling accuracy was equivalent to that achieved by manual picking and could be achieved in 1/100 of the time.

国家重点研发计划“华南火成岩地区深层地热资源探测技术”(2019YFC0604902)和国家自然科学基金企业创新发展联合基金(U19B6003)共同资助。