地震信号自动截取算法与震源类型识别研究

发布时间：2018-01-22 03:16

本文关键词： 地震波震源识别 BP-Adaboost方法 EMD-TKEO算法波群识别　出处：《广西师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：地震事件是地球物理演变过程中的必然事件,地震活动具有潜在的巨大破坏性,在人口稠密地区发生的强烈地震事件会对人民生命财产造成巨大损失。人们已建立起分布广泛的地震观测台网监测地震活动,地震波形信号是监测地震活动的唯一手段。许多在地表附近发生的非天然地震事件如化学爆炸、地下核爆炸、泥石流、滑坡、矿塌、地面塌陷、甚至大型工程建设中的人工活动源等都会以地震波的形式向四周传播能量,随着观测台网布置密集程度的增加和观测仪器灵敏度的提高,这些地震波也会被地震观测仪器检测到。对于天然地震事件,地震三要素:震源、震级、发震时刻,可以从台网观测波形求得;当然,通过地震波反演波形传播路径的特性则是探求地球内部结构的唯一手段。如何从连续观测信号中自动截取出对应于有效事件活动源的地震波形信号、以及根据所截取的地震波形信号的波形特征分类识别出震源类型(本文只关注地震和爆炸分类)则是本研究的主要内容。本文就震源类型识别研究、地震信号P波和S波初至时刻检测及波形自动提取、波群(噪声、P波群和S波群)样本识别3个问题展开论述。提出了一种基于BP-Adaboost方法的震源类型识别算法:选取BP神经网络作为弱分类器,通过Adaboost方法组合成强分类器。通过与BP神经网络、PCA-SVM方法对所选用的地震、爆炸事件波形信号数据集的分类识别结果表明,本方法得到了 98%以上的正确识别率,且具有良好泛化能力。提出一种基于EMD-TKEO算法的P波和S波初至时刻检测及地震波形自动截取算法:首先,事件全过程波形整体滤波,滤波后波形64样本点为窗口长度分窗;接着,每一窗口信号5层EMD分解,得到IMF;然后,EMD分解之后的IMF1分量计算TKEO能量,并归一化;最后,选定合适能量阈值来推断P波和S波初至时刻。实验结果表明,与目前主流方法STA-LTA及AR_AIC方法相比,本方法计算时间较短,且更早能识别出P波初至时刻的到来。通过设定适当的阈值,还可以从连续观测信号中截取出涵盖一个事件全过程的整段有效地震波形,这样就可以去除连续观测信号中的非事件部分的信号,对实验地震数据的截取研究平均压缩比可达57%,这可减小观测地震信号的存贮空间。提出了波群检测的识别判据:平均谱能量值和均方根误差特征,在EMD-TKEO算法的基础上,准确提取出噪声、P波群和S波群样本,通过扩展的多分类SVM方法进行分类识别研究。通过网格搜索法得出(C,σ)的二元组在取值为(1,0.01)时对波群样本的识别率最好,识别率为99%,并且有较高的算法执行效率,同时实验也说明了提出波群特征的有效性,使用较少的特征就能达到理想的识别效果。
[Abstract]:Earthquake event is an inevitable event in the process of geophysical evolution, and seismic activity is potentially destructive. Strong earthquake events in densely populated areas will cause great losses to people's lives and property. A network of widely distributed seismic observatories has been established to monitor seismic activities. Seismic waveform signals are the only means of monitoring seismic activity. Many unnatural earthquake events occurring near the surface such as chemical explosion underground nuclear explosion debris flow landslide mine collapse ground collapse. Even the artificial sources in the construction of large-scale projects will propagate energy in the form of seismic waves, with the increase of the density of observation network arrangement and the increase of the sensitivity of observation instruments. These seismic waves will also be detected by seismic observation instruments. For natural earthquake events, the three elements of earthquakes are: source, magnitude, and time of occurrence, which can be obtained from the observation waveforms of the network; Of course, the characteristics of waveform propagation path by seismic wave inversion is the only means to explore the internal structure of the earth. How to automatically intercept the seismic waveform signal corresponding to the active source of effective events from the continuous observation signal. The main content of this study is to classify and identify the source type according to the waveform characteristics of the intercepted seismic waveform signal (only focus on earthquake and explosion classification in this paper). Seismic signal P wave and S wave first arrival time detection and waveform automatic extraction, wave group (noise). This paper discusses three problems in the identification of P-wave group and S-wave group. A new method of source type recognition based on BP-Adaboost method is presented. BP neural network is selected as the weak classifier. The Adaboost method is combined to form a strong classifier, and the classification and recognition results of selected seismic and explosion waveform data sets with BP neural network / PCA-SVM method show that. The correct recognition rate of above 98% is obtained by this method. A new algorithm based on EMD-TKEO algorithm is proposed to detect the first arrival time of P wave and S wave and to intercept the seismic waveform automatically. Firstly, the whole waveform filtering of the whole process of the event is proposed. The 64 sample points of the filtered waveform are window length window splitting; Then, each window signal is decomposed into five layers of EMD. Then the IMF1 component after EMD decomposition is used to calculate the TKEO energy and normalize it. Finally, the appropriate energy threshold is selected to infer the first arrival time of P wave and S wave. The experimental results show that the calculation time of this method is shorter than that of STA-LTA and AR_AIC methods. And the arrival of the first arrival of P wave can be recognized earlier. By setting the appropriate threshold, the whole effective seismic waveform covering the whole process of an event can be intercepted from the continuous observation signal. In this way, the non-event part of the continuous observation signal can be removed, and the average compression ratio of the experimental seismic data interception can reach 57%. This can reduce the storage space of observed seismic signals. The identification criteria of wave group detection are proposed: the average spectral energy value and the root mean square error feature. The noise is extracted accurately on the basis of EMD-TKEO algorithm. The sample of P wave group and S wave group is classified and identified by extended multi-classification SVM method, and the binary group of C, 蟽) is obtained by grid search method. The recognition rate of the wave group samples is the best, the recognition rate is 99%, and the efficiency of the algorithm is high. At the same time, the experiment also shows the effectiveness of the proposed wave group features. An ideal recognition effect can be achieved by using fewer features.
【学位授予单位】：广西师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P315

【相似文献】