【作者】 田雅男；

【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】 随着全球对油气及矿藏资源需求量的不断增多，未知油气田及矿藏资源的开发成为目前地震勘探工作中的重点和难点。然而，地下构造及资源埋藏条件等不确定因素给勘探工作带来了很多困难。这就要求我们具备有效的地震资料处理手段，可以保证从地震勘探过程所采集到的地震数据中发现更多的有效信息，从而为新资源的开发提供有利依据。未知的埋藏条件及复杂的勘探环境使得采集到的地震数据中通常包含大量的噪声，具有较低的信噪比，有效信息淹没其中难以辨识。我们通常将地震噪声分为两大类：随机噪声和相关噪声。本文的研究对象是陆地地震勘探中的随机噪声，随机噪声是不规则的、没有规律的，且相邻道间彼此是互不相关的，它们没有固定的频率，几乎分布于整个频带，严重影响着地震记录的信噪比。因此随机噪声压制是地震数据处理中的重点和难点。我们针对复杂条件下地震资料中随机噪声的压制处理，分析了传统时频峰值滤波算法在地震勘探信号处理过程中存在的不足，基于径向轨线时频峰值滤波思想，提出了一种时频峰值滤波非线性二次轨线模型，并通过大量合成记录和实际地震资料，对新模型进行了性能验证。本文首先对地震勘探的基本原理、目前国内外研究现状以及地震波的性质等多方面地震勘探的相关知识进行了介绍，使我们可以更深的认识地震勘探，熟悉地震勘探信号处理的目的及研究意义。对时频峰值滤波方法的基本原理及现有改进时频峰值滤波方法的优缺点进行分析讨论，分别从轨线类型及特征、窗长的长短对去噪效果的影响等多个方面进行探讨，为后续研究奠定基础。针对新模型分别从模型的建立、滤波轨线的选取及样本重采样方案等方面进行详细说明。传统时频峰值滤波的输入信号是包含多个频率成分的一道地震记录，根据最优窗长与有效信号主频之间的关系可知，不同的频率具有各自不同的最优窗长。因此，传统时频峰值滤波中采用的固定窗长是无法实现对所有频率成分的有效估计的，部分频率成分会由于窗长不合适发生严重损失。其次，根据时频峰值滤波对湮没于高斯白噪声中线性信号的无偏估计性可知，含噪信号中有效信号的线性度越高，时频峰值滤波对其造成的估计偏差越小。而在最近被提出的径向轨线时频峰值滤波方法中，由于径向轨线的倾斜方向固定，而地震记录中的同相轴通常呈弯曲状态，且方向不固定，因此部分同相轴与轨线不能达到良好的匹配，从而造成有效信号的线性度得不到有效提高，可见径向轨线模型仍然存在缺陷，有待进一步改进。由于滤波轨线（即重采样轨迹）与同相轴的匹配程度决定了有效信号线性度提高的多少。于是，我们针对弯曲同相轴的情况，将平行的径向轨线发展为弯曲的二次轨线，分别提出了时频峰值滤波的抛物轨线模型及双曲轨线模型。对于弯曲的同相轴，二次轨线与轴的匹配程度明显高于径向轨线与同相轴的匹配程度，从而充分地提高了有效信号线性度。本文提出的二次轨线模型首先对含噪记录沿一些二次轨线进行重采样，将记录变换到一个新的作用域内，在新作用域内，有效信号的线性度大大提高，从而降低了时频峰值滤波过程中瞬时频率估计的偏差，更完整地保持了有效成分。在抛物轨线模型中，针对最优滤波轨线选取问题，我们采用了基于Canny算子的边缘检测法估计出含噪记录中同相轴的大致边缘，再利用曲线拟合技术通过这些边缘逻辑值拟合得到轴的包络线，并将包络线与不同弯曲程度的二次轨线进行相似度判定，相似度最大的轨线即为最优滤波轨线。在样本重采样过程中采用将时刻点与轨线交点作为样本点的提取方案。与按道与轨线交点的提取方案相比，该方案增加了采样序列中的样本数量，减少了提取数据序列中信号突变的现象，使采样结果更准确。同时为了保证提取样本序列的平滑性及有效成分的完整保持，我们还利用了插值技术对位移方向数据进行扩容，从而减少样本序列中的尖峰和毛刺，使滤波效果更理想，同时也更完整地保持了有效成分。在双曲轨线模型中，轨线的离心率可以随着记录中同相轴的弯曲程度进行调节，沿双曲轨线进行重采样的过程可以近似为将含噪地震记录由原来的位移-时间域变换到一个新的作用域。在新作用域内，有效信号主频降低，线性度大大提高。与抛物轨线模型相比，新模型在重采样过程中不是简单的采用就近原则，而是利用插值技术对数据进行扩容，提高了采样的准确度。此外，最优滤波轨线的选取方法适用性更强，原来的基于Canny算子的边缘检测法对同相轴的性质要求较高，且需要根据不同的含噪记录情况设定并调节阈值；而新模型中采用变离心率的轨线族，只要根据含噪记录中轴的分布趋势，选取适合的离心率范围即可，方便且实用性强。每一章节末尾给出了对整章内容的概括及创新点总结。对新模型利用不同合成地震记录进行性能检验，并分别与现有的几种常用去噪方法进行效果对比。与其他方法相比，二次轨线时频峰值滤波后，随机噪声得到了有效压制，同时有效同相轴变得更连续，有效成分保持完整。此外，新模型进一步被应用于实际地震资料随机噪声压制，并分别给出了整炮记录及部分记录的去噪结果。实验结果充分反映了新模型在低信噪比条件下同相轴恢复及随机噪声压制方面的优越性，处理后的同相轴变得更清晰，更连贯；且原本断裂的同相轴也连接起来，增加了可以从记录中获得的有效信息。更多还原

【Abstract】 With the global demand for oil, gas and mineral resources increasing, thedevelopment of the unknown oil/gas fields and mineral resources becomes important anddifficult in seismic exploration work. However, the subsurface structure and the uncertainresources burial conditions bring about many difficulties to the exploration work. Itrequires us to have an effective means for the seismic data processing to guarantee morevalid information can be found from the collected seismic data in the seismic explorationprocess so as to provide a favorable basis for the new resources. Due to the unknownburial conditions and complex exploration environment, the seismic data usually containsa lot of noise during the acquisition. It makes the signal-to-noise ratio (SNR) is so low thatthe effective information is difficult to identify. Seismic noise is usually divided into twocategories: random noise and correlated noise. In this paper, we mainly talk about therandom noise in land seismic exploration. Random noise is irregular, no laws, andunrelated to each other between the adjacent channels. It does not have a fixed frequencyand distributes in almost the entire frequency band. It has a serious impact on the seismicrecords in SNR. Therefore the random noise suppression is an important and difficult taskin seismic data processing. We analyze the shortcomings of the conventionaltime-frequency peak filtering (TFPF) algorithm in seismic signal processing and propose aquadratic-trace model of TFPF based upon the radial filtering trace ideas. Moreover, theperformance of the new model is verified on different synthetic records and applied in realdata processing.This paper analyzes the shortcomings of the conventional TFPF method in seismic dataprocessing. In the conventional TFPF, the input signal is a channel of seismic record thatcontains plurality of frequency components. According to the relationship between theoptimal window length (WL) and the dominant frequency of the effective signal, we knowthat different frequencies components have different optimal window lengths (WLs). Thus,the fixed window length (WL) used in the conventional TFPF can not be suitable for allfrequencies components. Some frequencies components will have serious damage due tothe unsuitable WL. According to the unbiased estimate property of time-frequency peakfiltering (TFPF) for the linear signal; we know that the higher of the linearity of theeffective signal in the input signal is, the smaller the deviation of TFPF brings about. Thus,improvement of the linearity of the effective signal through resampling the noisy recordsalong some filtering traces is the core idea of the principle of trace-based TFPF.Meanwhile, the matching degree of the filter trace (sample trajectory) and the reflectionevents determines the linearity enhancement of the effective signal. In2011, Wu et al. proposed a radial-trace time-frequency peak filtering (RT-TFPF) method using radialtraces. In this method, the noisy record is resampled along some radial traces to improvethe linearity of the effective signal. However, due to the fixed inclination direction oftraces, the reflection events are usually bent in shape, and the direction is not fixed.Sometimes, the reflection events and the trace can not achieve a good match. So, we focuson the case of bent events and develop the traces from parallel form to the quadratic form.In this paper, we propose a quadratic-trace time-frequency peak filtering includingparabolic-trace model and hyperbolic-trace model. For the case of bent events, thematching degree of the quadratic traces is clearly higher than the radial traces. Thus, thelinearity level of the effective signal is enhanced greater.In the selection of the optimal filtering traces, we have adopted edge detection methodto obtain the approximate edge of the events. Then, we get the envelope of the eventsthrough the curve fitting and compute the similarity of the envelope with quadratic curvesof different bending degree. The greatest similarity corresponds to the optimal filteringtrace. In the resampling process, we take the intersection point of time and the filteringtrace as the sample point to increase the number of samples in the extracted sequences. Inaddition, in order to ensure the smoothness of the extracted sample sequence, we alsomake use of interpolation techniques to reduce the spikes and glitches in the samplesequence. In the hyperbolic-trace model, the bending degree of the filtering trace isadjustable. Meanwhile, we deduce that along hyperbolic traces resampling process can beapproximated by the noised original seismic record x-t domain to the e-t domain. In thenew domain, the dominant frequency of the effective signal decreases and the linearity isimproved greatly. Different sample sequences have different eccentricity, which coincideswith the variable eccentricity mentioned originally. In the quadratic-trace model presentedin this paper, the noisy record is resampled along some quadratic traces. Then the noisyrecord is transformed into a new domain, where the linearity if the effective signal isgreatly improved and thereby the deviation of the instantaneous frequency estimation isreduced.Firstly, the basic principle of seismic exploration, the current research status and thenature of seismic wave and other aspects of seismic exploration related knowledge wereintroduced, so that we can get a deeper understanding of seismic exploration and seismicsignal processing. Moreover, the basic principles of TFPF and its existing improvedmethods are illustrated and the advantages and disadvantages are analyzed and discussed,respectively. The effects of the filtering trace and the window length (WL) on the noisereduction are discussed, respectively. Finally, we use different synthetic records to test theperformance of the quadratic-trace model, and compare it with several existing commonlyused de-noising methods. The experimental results verify that after the quadratic-tracetime-frequency peak filtering (TFPF), the random noise has been effectively suppressedand the effective events become more continuous. In addition, we further applied the novel model to the actual seismic data in random noise suppression. The entire shotrecords and some records denoising results are both given. Experimental results show thesuperiority of the novel model in the events recovery and random noise suppressionsufficiently under low SNR conditions. The reflection events become clearer, morecontinuous; meanwhile some originally broken events are linked to provide more validinformation.更多还原