An automatic recognition method of microseismic signals based on EEMD-SVD and ELM
An automatic recognition method of microseismic signals based on EEMD-SVD and ELM
复制标题
基于EEMD-SVD和ELM的微震信号自动识别方法
DOI：
10.1016/j.cageo.2019.104318
复制
发表时间：
2019-12-01
影响因子：
4.4
通讯作者：
Xu, Nuwen
中科院分区：
文献类型：
--
作者：
Zhang, Jinyong;Jiang, Ruochen;Xu, Nuwen
微震信号和爆破信号的识别对于地质灾害预测具有重要意义。用于模式识别的特征参数通常是人工设计的，采用起来不方便。提出了一种集成经验模式分解(EEMD)、奇异值分解(SVD)和极限学习机(ELM)相结合的方法。该方法利用EEMD将微震信号分解成多个本征模式函数(IMF)，并利用奇异值分解(SVD)从这些IMF分量组成的矩阵中提取特征值。在得到特征向量的基础上，引入ELM建立分类模型，对西南中国后子岩水电站的500个信号进行自动识别。实验结果表明，EEMD-SVD方法得到的奇异值能够有效地反映信号的关键特征。此外，去除噪声成分和虚假成分的识别效果要好于其他向量组合。与其他机器学习算法的输出相比，ELM算法的性能优于反向传播神经网络、遗传算法优化的神经网络和支持向量机。该模型在训练时间仅为0.152s的情况下，预测精度和马太相关系数分别高达93.85%和87.70%。
The recognition of microseismic and blasting signals is important for the prediction of geological disasters. Feature parameters for pattern recognition are usually designed manually, which is inconvenient to adopt. A new method combining ensemble empirical mode decomposition (EEMD), singular value decomposition (SVD) and extreme learning machine (ELM) was proposed. The method applied EEMD to decompose microseismic signals into multiple intrinsic mode functions (IMF) and used SVD to extract eigenvalues from matrices 'composed of these IMF components. After getting the feature vectors, ELM was introduced to establish a classification model to automatically identify 500 signals of Houziyan hydropower station in Southwest China. Experimental results suggested that the singular values obtained by the EEMD-SVD method could effectively reflect the key characteristics of the signals. Furthermore, the recognition result of eliminating noise components and false components was better than that of other vector combinations. Compared to the outputs of other machine learning algorithms, ELM performed better than back-propagation neural network, neural network optimized by genetic algorithm and support vector machine. The prediction accuracy and the Matthew's Correlation Coefficient of the model reached as high as 93.85%, and 87.70%, respectively, while the training time was only 0.152s.