基于隐半马尔可夫模型的微震信号分割方法

Microseismic signal segmentation method based on hidden semi-Markov model

微震监测系统采集到的连续微震信号中往往包含着多种微震事件,为了对各种事件做进一步的分析,如微震事件识别与分类、微震源定位等,对微震信号的分割进行研究是首要前提。针对此问题,提出了一种基于隐半马尔可夫模型(hidden semiMarkov model, HSMM)的微震信号分割方法,该方法将微震信号中有无微震事件发生视为HSMM中的状态转换过程,并考虑状态的持续时间。首先提取预处理后微震信号的香农能量包络作为特征,对应HSMM中的观测序列,然后对训练集信号使用Baum-Welch算法估计出HSMM最优参数,接着使用维特比算法预测待分割微震信号的状态序列,最后基于状态序列完成信号分割。使用来自实验室和隧道开挖工程中的微震数据验证了方法的有效性,对比方法为长短时窗比值(STA/LTA)算法和AIC拾取算法。实验结果表明,不论是初至时刻拾取还是结束时刻拾取,HSMM均取得了最好效果,平均拾取误差分别为5.44 ms和17.70 ms,且初至拾取误差在10 ms及20 ms内的占比分别为79.3%和100%。在对连续微震信号的分割实验中,HSMM的拾取效果也优于STA/LTA算法,初至时刻和结束时刻的平均拾取误差分别为3.55 ms和27.11 ms,优于STA/LTA算法的4.00 ms和167.88 ms。

The segmentation of microseismic signals is the primary prerequisite to analyze further various events,such as the identifi-cation and classification of microseismic events and the location of microseismic sources,since the continuous microseismic signals collected by the microseismic monitoring system often contain a variety of microseismic events.To solve this problem,a method of microseismic signal segmentation based on the hidden semi-Markov model(HSMM)was proposed,which took the occurrence of microseismic events as state transitions in the HSMM and considered the duration of each state.First,the Shannon energy envelope of the preprocessed microseismic signals was extracted as a feature,corresponding to the observation sequence in the HSMM.Then the Baum-Welch algorithm was used to estimate the optimal model parameters.The Viterbi algorithm was utilized to predict the state sequence of the microseismic signal to be segmented.Finally,the signal segmentation was performed based on the state sequence.The effectiveness of the proposed method was validated using microseismic data from laboratory experiments and tunnel excavation projects.STA/LTA algorithm and AIC pickup algorithm were employed as the methods of comparison.The experimen-tal results show that the HSMM has achieved the best results whether it is the first arrival pickup or the end time pickup.The aver-age pickup error is 5.44 ms and 17.70 ms,respectively,and the proportion of the first arrival pickup error within 10 ms and 20 ms is 79.3%and 100%,respectively.In the experiment of segmenting continuous microseismic signals,the picking performance of HSMM is superior to that of the STA/LTA algorithm.The average picking error at the first arrival and the end time are 3.55 ms and 27.11 ms,respectively,which are better than the STA/LTA algorithm’s 4.00 ms and 167.88 ms.