形变观测数据的多异常形态统一识别

地震前兆数据中的形变观测数据变化复杂,地球物理场变化和环境干扰等信息识别与剔除是与地震相关现象分析的关键.传统的信号识别主要采用回归分析、经验模态分解、频域信号分解等方法,但它们难以统一识别高幅值变化(尖峰、阶跃)与高频变化波形.本文利用信息熵参与形变时序数据的自动化分段构造子序列,一定程度上避免了这两种波形被分割的弊端,然后以统计描述方式表达子序列,最后利用角度异常因子(Angle-Based Outlier Factor,ABOF)和局部异常因子(Local Outlier Factor,LOF)构建对数函数定义离群点,以解决统一识别高幅度变化与高频率变化的问题.实验表明,对于特征向量维度变化的情况,LOF-ABOF算法的计算效率呈线性变化关系;在特征表达策略改变的情况下,该算法对高幅值变化和高频变化的异常识别效果良好.本文所提供方法可以检测出高幅值变化与高频率变化的异常形态,为地震前兆数据中形变观测数据"前兆信号"的识别提供指导与参考,为深入认识地震现象及其产生机理奠定基础.

Identification of multi-anomalies of precursory deformation data

Seismic precursory data are generally very complicated. Thus identification of geophysical field changes and elimination of ambient noise is crucial to analysis of the phenomenon associated with earthquakes. Traditional methods, including regression analysis, empirical mode decomposition and signal frequency domain analysis, are not feasible to recognize the changes of high amplitude (peaks and steps) and frequency simultaneously. In this study, the entropy of time series was used to automatically segment deformation time-series data into subsequences, partly avoiding the dissection of waveforms with high amplitude and frequency. Then the subsequences were featured by statistics, and outliers log function through Angle-Based Outlier Factor and Local Density Outlier Factor were defined to solve the problem of recognition of those waveforms. Experimental results show that there is a linear relationship between the efficiency of the algorithm and the changes of feature dimensions. When the dimension of feature changes, the LOF-ABOF algorithm takes effect on the recognition of high amplitude and high frequency. Our method can be used to detect the types of anomalies, which provide the guideline for precursor recognition and lay a foundation for further understanding seismic phenomena and their mechanism.