基于分块字典学习理论的地震数据去噪

随着油气勘探观测环境愈发复杂,采集的地震数据常常掺杂各种噪声信号,导致勘探目标引起的有效微弱信号被覆盖,严重影响高精度的地震勘探数据解译,因而有效的压制地震勘探数据噪声显得越发重要。本文采用字典学习策略,将复杂地震数据进行分块,通过分块数据的字典学习获取字典原子,构建高精度的字典学习地震数据稀疏表示,通过两次迭代更新字典原子,进行数据去噪。将本文的字典学习算法应用于含随机噪声的模拟数据和实测地震勘探数据处理,验证该算法的可行性及有效性。结果表明,本文算法有效去除了随机噪声,保留了有效信号同相轴,提高了信噪比,可为复杂含噪地震数据的去噪处理提供新的技术手段。     

Simultaneous Denoising and Interpolation of Seismic Data via the Deep Learning Method

Utilizing data from controlled seismic sources to image the subsurface structures and invert the physical properties of the subsurface media is a major effort in exploration geophysics. Dense seismic records with high signal-to-noise ratio (SNR) and high fidelity helps in producing high quality imaging results. Therefore, seismic data denoising and missing traces reconstruction are significant for seismic data processing. Traditional denoising and interpolation methods rarely occasioned rely on noise level estimations, thus requiring heavy manual work to deal with records and the selection of optimal parameters. We propose a simultaneous denoising and interpolation method based on deep learning. For noisy records with missing traces, we adopt an iterative alternating optimization strategy and separate the objective function of the data restoring problem into two sub-problems. The seismic records can be reconstructed by solving a least-square problem and applying a set of pre-trained denoising models alternatively and iteratively.We demonstrate this method with synthetic and field data.     

模型约束下的在线字典学习地震弱信号去噪方法

本文针对噪声成分和噪声结构的复杂性及弱信号的特征,发展了最新的在线字典学习去噪方法.在线字典学习去噪方法是以数据驱动的方式,反复进行学习构建字典方式,求得信号的稀疏性解以实现对信号的去噪,在此基础上,提出了数据驱动与模型驱动联合的模型约束下的在线字典学习去噪方法,先通过模型驱动方式获得一个较优质的学习样本以构建字典再进行去噪处理.通过和传统小波变换进行理论地震合成记录的效果对比,在高噪声比例的弱信号情况下远远优于传统的时频域去噪方法.实际数据去噪处理表明,模型约束下的在线字典学习去噪方法是一种有效的去噪方法,这种联合去噪方式能在高噪声背景下有效地提取出弱信号,具有广阔的推广应用前景.     

双稀疏字典和FISTA的地震数据去噪

地震数据的随机噪声去除是地震数据处理中的一项重要步骤,双稀疏字典提供了两层稀疏模型,比单层稀疏模型可以更好地去除噪声.该方法首先利用contourlet变换对地震数据进行稀疏表示,然后在contourlet域中使用快速迭代收缩阈值算法（fast iterative shrinkage-thresholding algorithm,FISTA）对初始字典系数进行更新,接着采用数据驱动紧标架（data-driven tight frame,DDTF）在contourlet域中得到DDTF字典并通过FISTA得到更新后的字典系数,最后通过DDTF字典和更新后的字典系数获得新的contourlet系数,并对新的contourlet系数进行硬阈值和contourlet反变换得到去噪后的数据.通过模拟数据和实际数据的实验证明：与固定基变换去噪方法相比,该方法可以自适应地对地震数据进行稀疏表示,在地震数据较为复杂时得到更高的信噪比;与字典学习去噪方法相比,该方法不仅拥有较快的去噪速度,而且克服了字典学习因为缺少先验约束造成瑕疵的缺点.     

Random noise suppression for seismic data using a non-local Bayes algorithm

For random noise suppression of seismic data, we present a non-local Bayes (NLBayes) filtering algorithm. The NL-Bayes algorithm uses the Gaussian model instead of the weighted average of all similar patches in the NL-means algorithm to reduce the fuzzy of structural details, thereby improving the denoising performance. In the denoising process of seismic data, the size and the number of patches in the Gaussian model are adaptively calculated according to the standard deviation of noise. The NL-Bayes algorithm requires two iterations to complete seismic data denoising, but the second iteration makes use of denoised seismic data from the first iteration to calculate the better mean and covariance of the patch Gaussian model for improving the similarity of patches and achieving the purpose of denoising. Tests with synthetic and real data sets demonstrate that the NL-Bayes algorithm can effectively improve the SNR and preserve the fidelity of seismic data.     

基于学习型超完备字典的地震数据去噪(英文)

基于变换基函数的方法,是地震去噪处理中最常用的技术之一,它利用地震数据在某种基函数变换域内的稀疏性和可分离性来达到剔除噪声的目的。但传统的做法是事先选定一组固定的变换基并在对应域内进行处理,其效果往往并不十分令人满意。为了探索新的改进方法,我们引入学习型超完备冗余字典,即根据地震模型数据进行学习和训练,以寻求最优的稀疏表示字典,而不是只选用固定的变换基。本文在字典学习中融入全变差最小化策略以压制伪吉布斯现象。我们选用离散傅里叶变换作为初始变换,并以随机噪声为例,对单一的全局变换、未经学习的超完备冗余字典和学习型超完备冗余字典做了比较。结果表明,利用经过训练的超完备冗余字典,在对地震数据进行稀疏表示的同时,也达到了有效去除噪声的目的,可视性和信噪比都得到了明显提高。我们也比较了均匀和不均匀字典子块的效果,结果表明,不均匀的字典子块更利于地震数据去噪。     

基于数据增广和CNN的地震随机噪声压制

卷积神经网络(Convolutional Neural Network,CNN)是一种基于数据驱动的学习算法,简化了传统从特征提取到分类的两阶段式处理任务,被广泛应用于计算机科学的各个领域.在标注数据不足的地震数据去噪领域,CNN的推广应用受到限制.针对这一问题,本文提出了一种基于数据生成和增广的地震数据CNN去噪框架.对于合成数据,本文对无噪地震数据添加不同方差的高斯噪声,增广后构成训练集,实现基于小样本的CNN训练.对于实际地震数据,由于无法获得真实的干净数据和噪声来生成训练样本集,本文提出一种直接从无标签实际有噪数据生成标签数据集的方法.在所提出的方法中,我们利用目前已有的去噪方法从实际地震数据中分别获得估计干净数据和估计噪声,前者与未知的干净数据具有相似纹理,后者与实际噪声具有相似的概率分布.人工合成数据和实际数据实验结果表明,相较于F-X反褶积,BM3D和自适应频域滤波算法,本文方法能更好地压制随机噪声和保护有效信号.最后,本文采用神经网络可视化方法对去噪CNN的机理进行了探索,一定程度上解释了网络每一层的学习内容.     

基于深度卷积神经网络的地震数据随机噪声压制

本文以提高地震数据的成像质量为目标,提出一种智能的卷积神经网络降噪框架,从带有噪声的地震数据中自适应地学习地震信号。为了加速网络训练和避免训练时出现梯度消失现象,我们在网络中加入残差学习和批标准化的方法,并采用了ReLU激活函数和Adam优化算法优化网络。此外,Marmousi和F3数据集被用来对网络进行训练和测试,经过充分训练的网络不仅能在学习中保留地震数据特征,而且能去除随机噪声。首先充分地训练网络,从中提取出随机噪声,并保留学习到的地震数据特征,之后通过重建地震数据估算测试集中的波形特征。合成记录和实际数据的处理结果显示了深度卷积神经网络在随机噪声压制任务中的潜力,并通过实验验证表明了深度卷积神经网络框架有很好的去噪效果。      

Inversion-based data-driven time-space domain random noise attenuation method

Conventional time-space domain and frequency-space domain prediction filtering methods assume that seismic data consists of two parts, signal and random noise. That is, the so-called additive noise model. However, when estimating random noise, it is assumed that random noise can be predicted from the seismic data by convolving with a prediction error filter. That is, the source-noise model. Model inconsistencies, before and after denoising, compromise the noise attenuation and signal-preservation performances of prediction filtering methods. Therefore, this study presents an inversion-based time-space domain random noise attenuation method to overcome the model inconsistencies. In this method, a prediction error filter (PEF), is first estimated from seismic data; the filter characterizes the predictability of the seismic data and adaptively describes the seismic data’s space structure. After calculating PEF, it can be applied as a regularized constraint in the inversion process for seismic signal from noisy data. Unlike conventional random noise attenuation methods, the proposed method solves a seismic data inversion problem using regularization constraint; this overcomes the model inconsistency of the prediction filtering method. The proposed method was tested on both synthetic and real seismic data, and results from the prediction filtering method and the proposed method are compared. The testing demonstrated that the proposed method suppresses noise effectively and provides better signal-preservation performance.     

Application of alternative digital rock physics methods in a real case study: a challenge between clean and cemented samples

High‐resolution three‐dimensional images are used in digital rock physics to numerically compute rock physical properties such as permeability and elastic moduli. These images are not widely available, and their preparation is both expensive and time consuming. All of these issues highlight the importance of alternative digital rock physics methods that are based on two‐dimensional images and use different approaches to compute effective properties of three‐dimensional samples. In addition, the scale of study in both standard and alternative digital rock physics is very small, which applications of its results are questionable at wells or reservoir scale. The aim of this study is to use two‐dimensional images and alternative digital rock physics techniques for computing seismic wave velocity and permeability, which are compared with well and laboratory data. For this purpose, data from one well in a reservoir located in the southwestern part of Iran are used. First, two clean (carbonate) and two cemented (limy sandstone) samples were collected from well cores at different depths. Then, two‐dimensional images by scanning electron microscope and conventional microscope were captured. In the next step, two alternative digital rock physics methods, namely, empirical relations and conditional reconstruction, have been employed to compute P‐wave velocity and permeability of a three‐dimensional medium. Results showed that, in clean (mono‐mineral) samples, velocity values were reasonably close to well data. However, permeability values are underestimated compared with laboratory data because laboratory data were obtained at ambient pressure, whereas alternative digital rock physics results are more representative of reservoir pressure conditions. Nevertheless, permeability–porosity trends are valid for both samples. In the case of cemented samples, a two‐scale procedure, along with a method for two‐scale computation and grain‐cement segmentation, is presented and developed. Results showed that P‐wave velocity is overestimated probably due to random sampling in this method. However, velocity–porosity trends are in agreement with well data. Moreover, permeability results obtained for cemented samples were also similar to those obtained for the clean samples.     

Application of randomized sampling schemes to curvelet‐based sparsity‐promoting seismic data recovery

Reconstruction of seismic data is routinely used to improve the quality and resolution of seismic data from incomplete acquired seismic recordings. Curvelet‐based Recovery by Sparsity‐promoting Inversion, adapted from the recently‐developed theory of compressive sensing, is one such kind of reconstruction, especially good for recovery of undersampled seismic data. Like traditional Fourier‐based methods, it performs best when used in conjunction with randomized subsampling, which converts aliases from the usual regular periodic subsampling into easy‐to‐eliminate noise. By virtue of its ability to control gap size, along with the random and irregular nature of its sampling pattern, jittered (sub)sampling is one proven method that has been used successfully for the determination of geophone positions along a seismic line. In this paper, we extend jittered sampling to two‐dimensional acquisition design, a more difficult problem, with both underlying Cartesian and hexagonal grids. We also study what we term separable and non‐separable two‐dimensional jittered samplings. We find hexagonal jittered sampling performs better than Cartesian jittered sampling, while fully non‐separable jittered sampling performs better than separable jittered sampling. Two other 2D randomized sampling methods, Poisson Disk sampling and Farthest Point sampling, both known to possess blue‐noise spectra, are also shown to perform well.     

Low-rank seismic denoising with optimal rank selection for hankel matrices

Based on the fact that the Hankel matrix representing clean seismic data is low rank, low-rank approximation methods have been widely utilized for removing noise from seismic data. A common strategy for real seismic data is to perform the low-rank approximations for small local windows where the events can be approximately viewed as linear. This raises a fundamental question of selecting an optimal rank that best captures the number of events for each local window. Gavish and Donoho proposed a method to select the rank when the noise is independent and identically distributed. Gaussian matrix by analysing the statistical performance of the singular values of the Gaussian matrices. However, such statistical performance is not available for noisy Hankel matrices. In this paper, we adopt the same strategy and propose a rule that computes the number of singular values exceed the median singular value by a multiplicative factor. We suggest a multiplicative factor of 3 based on simulations which mimic the theories underlying Gavish and Donoho in the independent and identically distributed Gaussian setting. The proposed optimal rank selection rule can be incorporated into the classical low-rank approximation method and many other recently developed methods such as those by shrinking the singular values. The low-rank approximation methods with optimally selected rank rule can automatically suppress most of the noise while preserving the main features of the seismic data in each window. Experiments on both synthetic and field seismic data demonstrate the superior performance of the proposed rank selection rule for seismic data denoising.     

低信噪比地震资料FDOC-seislet变换阈值消噪方法研究

地震数据本质上是时变的,不仅有效同相轴表现出确定性信号的时变特征,而且复杂地表和构造条件以及深部探测环境总是引入时变的非平稳随机噪声.标准的频率-空间域预测滤波只适合压制平面波信号假设下的平稳随机噪声,而处理非平稳地震随机噪声时,需要将数据体分割为小窗口进行分析,但效果不够理想,而传统非预测类随机噪声压制方法往往适应性不高,因此开发能够保护地震信号时变特征的随机噪声压制方法具有重要的工业价值.压缩感知是近年出现的一个新的采样理论,通过开发信号的稀疏特性,已经在地震数据处理中的数据插值以及噪声压制中得到了应用.本文系统地分析了压缩感知理论框架下的地震随机噪声压制问题,建立了阈值消噪的数学反演目标函数;针对时变有效信息具有的可压缩性,利用有限差分算法求解炮检距连续方程,构建有限差分炮检距连续预测算子(FDOC),在seislet变换框架下,提出一种新的快速稀疏变换域———FDOC-seislet变换,实现地震数据的高度稀疏表征;结合非平稳随机噪声不可压缩的特征,提出了一种整形迭代消噪方法,该方法是一种广义的迭代收缩阈值(IST)算法,在无法计算稀疏变换伴随算子的条件下,仍然能够对强噪声环境中的时变有效信息进行有效恢复.通过对模型数据和实际数据的处理,验证了FDOC-seislet稀疏变换域随机噪声迭代压制方法能够在保护复杂构造地震波信息的前提下,有效地衰减原始数据中的强振幅随机噪声干扰.     

主动源与被动源地震数据插值及联合数据成像

本文提出了两种情况下主动源数据和被动源数据的插值方法,并研究了两种数据在偏移成像中的互补效果.基于互相关法被动源数据重构原理,本文提出了结合多域迭代去噪技术的重构方法.提出了两种时间域主动源和被动源数据的插值方法,分别是共炮点域能量匹配插值和共检波点域最小平方匹配插值.然后对获得的主动源和被动源联合地震数据进行叠前深度偏移成像.在被动源活跃度不是很高的地区进行被动源地震勘探时,少量的主动源地震数据可以有效控制和补充被动源数据的成像效果.在稀疏炮点的主动源勘探中,有效利用被动源的信息能够在成像中增加更多的细节信息,提高成像质量.     

A multi-data training method for a deep neural network to improve the separation effect of simultaneous-source data

Within the field of seismic data acquisition with active sources, the technique of acquiring simultaneous data, also known as blended data, offers operational advantages. The preferred processing of blended data starts with a step of deblending, that is separation of the data acquired by the different sources, to produce data that mimic data from a conventional seismic acquisition and can be effectively processed by standard methods. Recently, deep learning methods based on the deep neural network have been applied to the deblending task with promising results, in particular using an iterative approach. We propose an enhancement to deblending with an iterative deep neural network, whereby we modify the training stage of the deep neural network in order to achieve better performance through the iterations. We refer to the method that only uses the blended data as the input data as the general training method. Our new multi-data training method allows the deep neural network to be trained by the data set with the input patches composed of blended data, noisy data with low amplitude crosstalk noise, and unblended data, which can improve the ability of the deep neural network to remove crosstalk noise and protect weak signal. Based on such an extended training data set, the multi-data training method embedded in the iterative separation framework can result in different outputs at different iterations and converge to the best result in a shorter iteration number. Transfer learning can further improve the generalization and separation efficacy of our proposed method to deblend the simultaneous-source data. Our proposed method is tested on two synthetic data and two field data to prove the effectiveness and superiority in the deblending of the simultaneous-source data compared with the general training method, generic noise attenuation network and low-rank matrix factorization methods.     

二进小波变换方法的地震信号分时分频去噪处理

应用二进小波变换理论，对地震记录剖面进行小波分解，在小波变换域，对各分量剖面进行分时分频相关分析，并依此进行加权处理，最后通过小波反交换算法恢复出地震记录剖面．经实验该方法对衰减地震信号的噪声具有比较好的效果．     

3D simultaneous seismic data reconstruction and noise suppression based on the curvelet transform

Seismic data contain random noise interference and are affected by irregular subsampling. Presently, most of the data reconstruction methods are carried out separately from noise suppression. Moreover, most data reconstruction methods are not ideal for noisy data. In this paper, we choose the multiscale and multidirectional 2D curvelet transform to perform simultaneous data reconstruction and noise suppression of 3D seismic data. We introduce the POCS algorithm, the exponentially decreasing square root threshold, and soft threshold operator to interpolate the data at each time slice. A weighing strategy was introduced to reduce the reconstructed data noise. A 3D simultaneous data reconstruction and noise suppression method based on the curvelet transform was proposed. When compared with data reconstruction followed by denoizing and the Fourier transform, the proposed method is more robust and effective. The proposed method has important implications for data acquisition in complex areas and reconstructing missing traces.     

基于Seislet-TV双正则化约束的地震随机噪声压制方法

人工地震数据总是受到随机噪声的干扰,地震数据时-空变的特性使得常规去噪方法处理效果并不理想,容易导致有效信号的损失.目前广泛应用的预测滤波类方法存在处理时变数据能力不足的问题.随着压缩感知理论的不断完善,稀疏变换阈值算法能够解决时变地震数据噪声压制问题,但是常规的稀疏变换方法,如傅里叶变换,小波变换等,并不是特殊针对地震数据设计的,很难提供地震数据最佳的压缩特征,同时,常规阈值算法容易导致去噪结果过于平滑.因此开发更加有效的时-空变地震数据信噪分离方法具有重要的工业价值.本文将地震数据信噪分离问题归纳为数学基追踪问题,在压缩感知理论框架下,利用特殊针对地震数据设计的VD-seislet稀疏变换方法,结合全变差(TV)算法,构建seislet-TV双正则化条件,并利用分裂Bregman迭代算法求解约束最优化问题,实现地震数据的有效信噪分离.通过理论模型和实际数据测试本文方法,并且与工业标准FXdecon方法进行比较,结果表明基于seislet-TV双正则化约束条件的迭代方法能够更加有效地保护时-空变地震信号,压制地震数据中的强随机噪声.     

基于SW统计量的自适应时频峰值滤波压制地震勘探随机噪声研究

由于金属矿区地震记录中随机噪声性质复杂且信噪比低,常规降噪方法难以达到预期的滤波效果.时频峰值滤波(TFPF)方法是实现低信噪比地震勘探记录中随机噪声压制的有效方法,但其在复杂地震勘探随机噪声下时窗参数优化问题仍难以解决.本文充分利用地震勘探噪声的统计特性,结合Shapiro-Wilk(SW)统计量辨识地震勘探记录中的微弱有效信号,提出基于SW统计量的自适应时频峰值滤波降噪方法(S-TFPF).在S-TFPF方案中,对于有效信号集中区,S-TFPF方法根据信号频率特征,选择有利于信号保持的较短时窗长度;对于噪声集中区,按噪声方差自适应增加时窗长度,增强随机噪声压制能力.S-TFPF应用于合成记录和共炮点记录的滤波结果表明,与传统时频峰值滤波方法相比,S-TFPF方法可以有效抑制低信噪比地震勘探记录中的随机噪声,更好地恢复出同相轴.