【作者】 刘彦萍；

【导师】 李月；

【作者基本信息】 吉林大学 ， 通信与信息系统， 2013， 博士

【摘要】 地震勘探是油气、矿藏资源开发的一种重要的物探方法。它利用人工方法（用炸药或非炸药震源）激发地震波，根据岩石的弹性来研究地震波在地层中的传播规律以查明地下地质构造。在地震勘探时，由于人为、环境、仪器等各种因素的影响，会使采集到的地震资料中包含有大量的噪声。这些噪声与有关地下构造和岩性的信息之间互相交织在一起，对有用信息造成了不同程度上的干扰。因此，这些资料不宜被直接用来做地质解释，需要对其进行数字处理，从中提取出相关有用信息然后再进行一系列的后续处理。地震勘探中的“三高”（即高信噪比（high signal-to-noiseratio，SNR）、高分辨率（high resolution）和高保真度（high fidelity））及“一准”（准确成像）资料处理手段对于有用信息的获取具有重要作用。地震信号降噪是地震资料数字处理的关键步骤之一。有效地压制地震勘探资料中的噪声，提高地震资料的信噪比、分辨率和保真度对于地质构造的解释及油气、矿藏资源的开发具有重要的意义。本文主要研究地震信号降噪中的随机噪声压制，目前已有很多行之有效的方法用来消减地震勘探随机噪声，而时频域方法是近些年来被验证和广泛应用的一类方法。本文主要介绍了三种时频域滤波方法，它们是线调频小波（Chirplet）、经验模态分解（Empirical Mode Decomposition, EMD）和时频峰值滤波（Time-Frequency PeakFiltering, TFPF）方法。在众多的时频域方法中，TFPF方法表现得尤为突出。TFPF是一种非常有效的随机噪声消减方法，起初被应用于生物医学中新生儿脑电信号的增强中。近年来，该方法已经被吉林大学现代信号处理实验室深入研究并应用于地震勘探随机噪声的消减方面。本文针对传统的一维TFPF方法在滤波方面存在的不足进行改进，提出了三种改进方案。首先针对TFPF方法本身存在的一对矛盾：长窗长能更有效地压制随机噪声，但是对有效信号的幅值损失很大；反之，短窗长对有效信号的幅值保护很好，但在随机噪声压制方面表现得有些能力不足。也就是说窗长的选择对TFPF方法的限制较大，如果能在选择窗长方面具有灵活性，那么将会取得更好的滤波效果，从而我们需要在压制随机噪声和保护有效信号这两方面达到一个很好的权衡。本文采用EMD方法来辅助TFPF以取得在噪声压制和幅值保持方面的一个很好的权衡，具体方案为利用EMD方法的分解特性，即它能够将原始含噪信号按照频率从高到低分解成一系列本征模态函数（Intrinsic Mode Functions, IMFs），这些模态都是组成原始信号的各个分量，然后通过计算各个模态间的互相关程度大小来判断需要进行滤波的模态，进而选择不同窗长的TFPF进行处理，最后将滤波后的模态和剩余模态相加得到最终的滤波信号。此方法为TFPF的窗长选择提供了很大的灵活性，能够对原信号中不同的频率分量有针对性地选择窗长，即对于噪声主导的模态分量选择长窗长进行滤波，反之，对于信号主导的模态分量选择短窗长进行滤波，对于纯信号模态不进行滤波，这样不但能够有效地消减随机噪声而且能够很好地保护有效信号的幅值。通过对模拟地震信号和实际地震数据的处理实验，我们取得了较为理想的效果，从而验证了该方法的优势。接着本文将传统的一维TFPF方法发展为时空二维TFPF方法，这样对于更有效地压制地震勘探随机噪声，恢复有效同相轴以及减小滤波偏差方面具有很大的改善。我们所研究和处理的地震资料表现为时间和空间上的二维特性，那么选择符合其时空二维特性的滤波方法一定会取得较为理想的效果。本文所研究的时空二维TFPF方法是利用地震同相轴的横向相关性，借助Radon变换来实现时空域滤波。Radon变换是一种沿着预先定义好的路径，如直线、抛物线或者双曲线等路径将原始数据进行叠加的方法。鉴于Radon变换能将原始地震记录中的同相轴识别出来，表现为在Radon域中同相轴被聚焦为不同位置的能量子波，这样其实是对各同相轴的走向起到一个表征作用，也就是为TFPF处理提供方向性，即达到我们想要沿着有效同相轴方向进行TFPF处理的目的。这种做法打破了传统TFPF方法沿着地震道滤波的局限，为地震勘探随机噪声的压制提供了新的途径。我们通过对模拟地震信号和实际地震资料进行实验，验证了该方法的实用性和有效性并将其适用性进行了推广，提出了局部时空二维TFPF方法。本文最后介绍了局部时空二维TFPF方法的原理和应用。局部方法比全局方法适用性更广，对于处理同相轴情况较为复杂的地震资料更加有效。全局Radon变换对于具有规则几何形状同相轴的地震资料是非常适用的，但是对于含有不规则几何形状的同相轴或者是相交在一起的、其倾斜或弯曲程度差异很大的同相轴的地震资料，全局Radon变换已显得能力不足了，需要采用局部方法来跟踪同相轴的特征变化趋势。局部Radon变换其实是全局Radon变换加时空窗的形式。加窗后的Radon变换使得局部区域内的同相轴情况简单化，便于采用简单的叠加路径进行计算。我们先采用局部Radon变换对原始地震记录中的同相轴进行跟踪识别，然后在局部Radon域内进行TFPF处理。这种方法是对基于Radon变换的时空二维TFPF方法的进一步推广，增加了其在地震勘探随机噪声压制方面的普适性和灵活性，为处理较复杂的地震资料提供了新的途径。更多还原

【Abstract】 Seismic prospecting is an important kind of geophysical method for oil, gas and mineralresources. It utilizes manual methods (using explosive or non-explosive hypocenters) tostimulate seismic waves and study the propagation law of seismic waves in strata toascertain the underground geological structure according to rock elasticity. In seismicprospecting, there are lots of noises in collected seismic data because of the influencescaused by the behaviors of people, environments, and instruments and so on. These noisesare interlaced with the related information of underground structure and lithology, andcause interference for the valid information to varying degrees. So this kind of seismicdata could not be used to do geological explanation directly. It must be conducted digitalprocessing and extracted related useful information, and then taken a series of subsequentprocessings. The data processing means of “three highs”(high signal-to-noise ratio (SNR),high resolution and high fidelity) and “one accuracy”(accurate imaging) are important forthe acquisition of valid information. The seismic noise reduction is one of the keyprocedures of seismic data processing. To suppress the seismic noise effectively andimprove the SNR, resolution and fidelity of seismic data are of great importance for theinterpretation of geological structure and exploitation of oil, gas and mineral resources.This article mainly studies the seismic random noise reduction. At present, there are somany effective methods to suppress the seismic random noise, and time-frequency domainalgorithms belong to the kind of method which has been verified and widely used. In thisarticle, we mainly introduce three algorithms of time-frequency domain, Chirplet,empirical mode decomposition (EMD) and time-frequency peak filtering (TFPF). Amongnumerous time-frequency domain methods, TFPF is especially prominent. TFPF is ahighly effective method in random noise reduction, and the conventional1-D TFPF wasapplied to the enhancement of newborn electroencephalogram signals originally and hasbeen applied in the seismic random noise reduction by the modern signal processinglaboratory of Jilin University.This article aims at the shortage in filtering aspect of the conventional1-D TFPF andtakes improvement measures by proposing three improved approaches. First, it aims at apair of contradiction in TFPF itself that is long window length can suppress random noise effectively but produces large attenuation for the amplitude of valid signal, on the contrary,short window length can preserve the amplitude of valid signal well but presents lack ofcapability in random noise reduction. That is to say, there is a large restriction to TFPF bythe selection of window length. If it can be flexible in the window length selection, it willget better filtering results and we need to obtain a good trade-off between the randomnoise reduction and the valid signal preservation. This article adopts EMD method toassist TFPF to obtain the good trade-off between noise suppression and amplitudepreservation. The concrete scheme is that it utilizes the decomposition characteristic ofEMD which can decompose a signal to a series of intrinsic mode functions (IMFs) fromhigh frequency to low frequency. These modes are components of the original signal, andwe judge and find out the modes need to be filtered by computing the correlation degree ofeach other, and then do TFPF for these modes by selecting different window length. Atlast, adding the filtered modes and the residual modes up to get the final filtering signal.This kind of method possesses great flexibility in the selection of window length for TFPF,that is to say, it could select window length for different frequency componentsdiscriminatingly: for the noise dominant components, it selects long window length tofilter, conversely, for the signal dominant components, it selects short window length tofilter, and for the pure signal mode, it doesn’t need to filter. By doing this, it not only cansuppress the random noise effectively, but also can preserve the amplitudes of validsignals. Through experiments on synthetic seismic signals and field seismic data, we haveobtained some comparatively ideal results thereby testified the superiority of this method.Next, this article develops the conventional1-D TFPF to be spatiotemporal2-D TFPF.Thus, it is a big improvement for suppressing seismic random noise, recovering reflectionevents and reducing filtering bias more effectively. The seismic data shows2-D features oftime and space, so we should adopt the filtering methods confirmed to the spatiotemporal2-D features of seismic data and this will bring very good results. In this article, the2-DTFPF method we studied utilizes the time-space correlation of seismic events andimplements spatiotemporal filtering by virtue of Radon transform. Radon transform is amethod which adds the original data up along predefined path, such as line, parabola orhyperbola et al. In view of Radon transform can identify the reflection events of theoriginal seismic records by presenting that these events are focused as energy waveletswith different locations in Radon domain, so it plays a role in representing the direction ofreflection events. Thus, it provides direction for TFPF processing and achieves thepurpose of filtering along the direction of valid reflection events. It breaks the limitation ofthe conventional TFPF which implements filtering along the channel direction and provides a new way for seismic random noise suppression. We demonstrate the validityand practicability of this method through experiments on synthetic seismic signals andfield seismic data, and extend the applicability to a wide range by proposing localspatiotemporal2-D TFPF method.At last, this article introduces the local spatiotemporal2-D TFPF method. Theapplicability of the local method is broader than the global method, so it is more effectivein processing the seismic data with relatively complicated situations of reflection events.The global Radon transform is very applicable to the seismic data with regular geometriesevents but incapable to the seismic data with irregular geometries events or interlacedevents with different slope or curvature parameter. At this time, we need to adopt localmethods to track the characteristic variation trend of the reflection events. The local Radontransform is the form of global Radon with spatiotemporal window actually. Thewindowed Radon transform could provide simplification for the situation of the seismicevents in local areas so that it is convenient to compute by adopting simple superpositionpaths. So, we adopt local Radon transform to identify the reflection events of the originalseismic record firstly, and then do TFPF in local Radon domain. This method is apopularizing way of the spatiotemporal TFPF and increases its universality and flexibilityin seismic random noise suppression. Therefore, it provides a new way for the processingof more complicated seismic data.更多还原