基于循环神经网络的自适应滤波方法及应用研究

针对目前地震工程研究领域在滤波方法上存在人为因素、峰值突刺、噪声干扰等方面的缺陷,结合递归最小二乘法(RLS)和循环神经网络(RNN)模型,提出了一种自适应滤波的新方法。研究分析表明,该方法通过设置自适应调节滤波器参数以及算法的自我迭代等方式进行滤波,对噪声识别能力和滤波速度上均优于美国地质调查局(United States Geological Survey,USGS)所推荐的传统滤波方法,并可有效降低滤波后对原始波形的失真损坏以及相位提前等问题。同时,运用所提自适应滤波方法将其应用于不同场地类型台站的含速度脉冲近场地震记录,进一步验证了自适应滤波方法的有效性和适用性。研究成果为地震工程领域的滤波分析提出了一种新思路和新方法,也可为地震记录处理及相关应用工作提供参考。

Robust adaptive polarization analysis method for eliminating ground roll in 3C land seismics

To improve the data quality of converted waves, and better identify and suppress the strong ground-roll interference in three-component （3C） seismic recordings on land, we present an adaptive polarization filtering method, which can effectively separate the ground- roll interference by combining complex polarization and instantaneous polarization analysis. The ground roll noise is characterized by elliptical plane polarization, strong energy, low apparent velocity, and low frequency. After low-pass filtering of the 3C data input within a given time-window of the ground roll, the complex covariance matrix is decomposed using the sliding time window with overlapping data and length that depends on the dominant ground-roll frequency. The ground-roll model is established using the main eigenvectors, and the ground roll is detected and identified using the instantaneous polarization area attributes and average energy constraints of the ground-roll zone. Finally, the ground roll is subtracted. The threshold of the method is stable and easy to select, and offers good ground- roll detection. The method is a robust polarization filtering method. Model calculations and actual data indicate that the method can effectively identify and attenuate ground roll while preserving the effective signals.

Improve Q estimates with spectrum correction based on seismic wavelet estimation

Characterization of seismic attenuation,quantified by Q,is desirable for seismic processing and interpretation.For seismic reflection data,the coupling between seismic wavelets and the reflectivity sequences hinders their usage for Q estimation.Removing the influence of the reflectivity sequences in reflection data is called spectrum correction. In this paper,we propose a spectrum correction method for Q estimation based on wavelet estimation and then design an inverse Q filter.The method uses higher-order statistics of reflection seismic data for wavelet estimation,the estimated wavelet is then used for spectral correction.Two Q estimation methods are used here,namely the spectral-ratio and centroid frequency shift methods.We test the characteristics of both Q estimation methods under different parameters through a synthetic data experiment.Synthetic and real data examples have shown that reliable Q estimates can be obtained after spectrum correction;moreover, high frequency components are effectively recovered after inverse Q filtering.

Ground roll separation method via threshold fi ltering and constraint of seismic wavelet support in curvelet domain

Most traditional ground roll separation methods utilize only the difference in geometric characteristics between the ground roll and the refl ection wave to separate them.When the geometric characteristics of data are complex,these methods often lead to damage of the reflection wave or incompletely suppress the ground roll.To solve this problem,we proposed a novel ground roll separation method via threshold filtering and constraint of seismic wavelet support in the curvelet domain;this method is called the TFWS method.First,curvelet threshold fi ltering(CTF)is performed by using the diff erence of the curvelet coeffi cient of the refl ection wave and the ground roll in the location,scale,and slope of their events to eliminate most of the ground roll.Second,the degree of the local damaged signal or the local residual noise is estimated as the local weighting coeffi cient.Under the constraints of seismic wavelet and local weighting coeffi cient,the L1 norm of the refl ection coeffi cient is minimized in the curvelet domain to recover the damaged refl ection wave and attenuate the residual noise.The local weighting coeffi cient in this paper is obtained by calculating the local correlation coeffi cient between the high-pass fi ltering result and the CFT result.We applied the TFWS method to simulate and fi eld data and compared its performance with that of frequency and wavenumber filtering and the CFT method.Results show that the TFWS method can attenuate not only linear ground roll,aliased ground roll,and nonlinear noise but also strong noise with a slope close to the refl ection events.

Analysis the First Arrival of P-Wave of Ina-TEWS and CTBT Stations to Support Earthquake Early Warning

The authors make the analysis of first arrivals of the P-wave from Ina-TEWS （Indonesian tsunami early warning system） and CTBT （comprehensive nuclear-test-band treaty） stations. These are used for earthquake early warning, magnitude determination and potential earthquake hazard mitigation based on seismogram acceleration. This research is focused on the study of energy duration of high frequency, and the maximum displacement of P-waves by observing broadband seismograms. The further analysis consists of deconvolution, integration or defferentiation, recursive filtering for data restitution, and applying a Butterworth filter of second order. The Butterworth filter uses high frequency 0.075 Hz to cut the effect of drift, and band-pass frequency 2-4 Hz for use in magnitude calculation. The authors choose potentially damaging earthquakes to be greater than Mw 〉 6.0. Based on the trigger on the three seconds the first arrival P-wave, the dominant period （Td） and amplitude displacement （Pd） was calculated by using data CISI （Indonesian CI Sompet） seismological station, Garut （west Java） and tested for data CTBT, LEM bang, Bandung （LEM station）. This research resulted determination of the P-wave arrival time accurately using integrated skewness and kurtosis. Performance data from the CTBT stations is very high. Signal to noise ratio 〉1,000 after passing through the filter. Such riset conducted to find out a rapid magnitude estimations from predominant frequency of displacement are： log Td = 0.2406 M- 1.3665 （R = 0.73） or M = 4.156 log Td ＋ 5.6797. Relationship of Pd, magnitude moment, Mw and hypocentre, R are log Pd = -4.684 ＋ 0.815 Mw - 1.36 log R. For relation of PGA （peak ground acceleration） and amplitude displacement are log PGA = 1.117 log Pd ＋ 0.728 （R = 0.91）. Furthermore, this formula can be used to support earthquake early warning in west of Java.