基于时频分布特征稀疏优化建模的滑动扫描数据谐波噪声压制方法

Vibroseis techniques currently have been widely used in land seismic exploration because it is safe, environment-friendly, high-efficiency, and low-comsumption. However, the vibroseis acquired seismic data are unavoidably degraded by harmonic noise, which are becoming one of the major problems affecting the development of vibroseis techniques. This project focuses on the harmonic noise suppression for slip-sweep vibroseis acquired seismic data. By analyzing the time-frequency distribution characteristics of typical ground force signal, and the corresponding effective signals and harmonic noise in correlated seismic records, we propose to utilize their waveform divergency and construct sparse representation dictionaries for effective signals and harmonic noise, respectively. The two corresponding waveform dictionaries are then combined together as an overcomplete dictionary to perform the harmonic noise separation. The morphological component analysis (MCA) model, based on Lp norm (0 <p <1) sparsity constraint, will be used for the harmonic noise separation problem. And a fast iterative thresholding algorithm will be designed to solve this nonlinear optimization problem. Furthermore, the corresponding parallel implementation scheme will also be included for massive pre-stack seismic data processing. The study not only shows significant value in seismic application, but also promote the development of the theory and methods for sparse representation.

可控震源具有安全、环保、高效、低耗等优势, 已在陆上地震勘探中得到了广泛应用。但谐波噪声的存在严重降低了可控震源采集数据的品质,并成为制约可控震源勘探技术发展的主要因素之一。本项目聚焦于可控源滑动扫描接收地震数据中谐波噪声的压制。根据典型的地面力信号、相关后地震记录中有效信号及谐波噪声的时频分布特征的差异性，构造能分别稀疏表示有效信号及谐波噪声的波形字典，进而组成用于谐波噪声分离的超完备波形字典；建立基于Lp范数（0<p<1）稀疏性约束的形态分量分析（Morphological Component Analysis, MCA）模型，设计与之对应的快速阈值迭代算法；发展适合海量叠前地震谐波噪声压制的并行实现方法。该项研究不但具有重要的应用价值，而且能够推动稀疏表示理论与方法的发展。

本项目经过四年研究工作，完成了项目计划研究内容：提出了一种基于谐波噪声及有效信号波形特征的地震数据稀疏优化表示方法，该方法通过了典型模型及实际数据验证，保真性明显好于目前工业界常用方法；发展了常规基于L1范数约束的形态成分分析方法，提出了基于Lp范数约束形态成分分析方法，并在数据恢复及数据超分辨率两个方面测试了方法的正确性及有效性；为了适应地震信号海量特性，采用GPU进一步提高了本项目方法计算速度。在项目计划书之外，本项目还提出了适合于面波噪声、脚印噪声、风车噪声、DAS耦合噪声、多次反射折射干扰及探地雷达直达波干扰压制的超完备字典构造方法及基于形态成分分析的噪声压制技术，其中，采集脚印噪声及谐波噪声压制的超完备字典构造，采用了数据驱动策略的字典学习方法。另外，本项目及时跟进新的研究方向，在国际上率先探索了将深度学习用于地震资料噪声衰减、高精度曲率提取及提高分辨率等方面的研究。