地震震相初至自动检测技术综述

Review of automatic onset time picking for seismic arrivals

地震学研究中地震震相初至拾取处于基础而又关键的环节,其拾取速度和精度直接影响其在地震精确定位、震相识别、震源机制及破裂过程、地震勘探以及地震层析成像中的应用效率和精度.早期的震相初至拾取是人工的、非实时分析;随着计算机技术、数据采集与处理技术以及定量或数字地震学的发展,震相初至拾取也由早期的人工分析过渡到人机互动的半自动分析以及后来的自动实时检测.目前,可以进行震相初至自动检测的方法有很多,但没有一种单独的方法能在所有不同类型震源、传播路径、接收方式以及噪声背景下对初至进行一致的拾取,且对于信噪比低、初动不明显或后期弱震相埋在早期震相尾波中、噪声与地震信号频率相近时的地震记录,初至自动拾取的效果通常都不理想.鉴于此,有必要对目前流行的各种震相初至自动识别检测方法进行归纳总结,以期对该领域的发展有所裨益.本文就目前常用的检测方法技术(如能量法、分形维数法、频率法、偏振法、自回归模型、相关法、小波变换、人工神经网络法等)按时间域、频率域、时频域、综合方法四大类进行了回顾、分析及综述.结果表明:寻找一种综合信号和噪声多特性差异及多震相特征量的方法可能是目前初至自动检测技术的发展方向,即充分利用信号与噪声在运动学、动力学、频谱特征、偏振属性等方面的显著差异性,形成一套同时具有算法简单、检测精度高、多道处理功能、可用于实时处理特征的综合识别检测方法技术.

The onset time picking is a fundamental and key role in seismology and the picking speed and accuracy is directly related to the efficiency and precise in earthquake relocation, seismic phase recognition, focus dynamic process, seismic exploration and seismic tomography. In the early day, the onset time picking was done by manually. With the fast development of computational technique, data acquisition and data processing technologies, as well as the coming quantitative or digital seismology, the onset time picking has been transformed from the early- stage manually analysis to the middle-stage human-computer interaction, and then to the current-stage automatic picking. At present, there are many algorithms for onset time picking, but none of them is able to make a consistent picking for different source environments, travel paths, receiver locations and signal/noise levels, especially when the noise and signal has nearly the same frequency content, or the later arrivals are buried in the coda of early arrivals. For these reasons, it is constructive to make an exhaustive review on the current methods so as to help to improve and advance the onset time picking methodology. In this paper we give a briefly review, analyze and summarize the current methods （such as energy aaalysis, the fractal dimension method, frequency method, polarization method, regression model, correlation method, wavelet transform, artificial neural network and etc） concentrated ourselves in time domain, frequency domain, time-frequency domain and comprehensive methods in the hope that this will do some helps in the development and application of this technique. The results show that it is constructive to look for a comprehensive method, which were based on the differences between the signal and noise in terms of kinematic, dynamic, polarization content and frequency spectrum to form a simple algorithm, high picking accuracy, multi-channel functional and suitable for real time comprehensive picking method and technology.