粒子群优化广义回归神经网络与HHT样本熵结合的地震辨识研究  

Research on seismic discrimination based on particle swarm optimization generalized regression neural network and HHT sample entropy

作　　者：庞聪[1,2,3] 丁炜 程诚[4] 吴涛 江勇 马武刚[1] 廖成旺 PANG Cong;DING Wei;CHENG Cheng;WU Tao;JIANG Yong;MA WuGang;LIAO ChengWang(Institute of Seismology,CEA,Wuhan 430071,China;Hubei Key Laboratory of Earthquake Early Warning,Wuhan 430071,China;Hubei Earthquake Administration,Wuhan 430071,China;School of Mathematics and Information Technology,Yuncheng University,Yuncheng 044031,China)

机构地区：[1]中国地震局地震研究所,武汉430071 [2]地震预警湖北省重点实验室,武汉430071 [3]湖北省地震局,武汉430071 [4]运城学院数学与信息技术学院,运城044031

出　　处：《地球物理学进展》2022年第4期1457-1463,共7页Progress in Geophysics

基　　金：湖北省自然科学基金(2019CFB768);中国地震局地震研究所和应急管理部国家自然灾害防治研究院基本科研业务费专项资助项目(IS201856290);中国综合地球物理场仪器研发专项(Y201707)联合资助。

摘　　要：天然地震与人工爆破信号具有许多相似的特性,为实现地震类型的准确辨识,提出一种基于粒子群优化广义回归神经网络和HHT样本熵的地震类型辨识新方法.该方法先通过经验模态分解(EMD)将初始信号分解成6个本征模态函数(IMF)及1个残余向量,对前三个IMF进行Hilbert变换得到瞬时频率和瞬时能量,进而提取出样本熵并构造神经网络训练集与测试集;利用粒子群算法和训练集优化广义回归神经网络的光滑因子参数值,建立PSO-GRNN分类模型.将该模型结果与BPNN模型、GRNN模型、PNN模型及RBFNN模型等神经网络模型进行性能对照,得到该模型单次识别的准确率、MAE、MAPE、RMSE R2及MSE分别为95%、0.1604、0.1204、0.2381、0.7123、0.0567,绝大多数性能评价指标优于其他4种神经网络模型.该研究建立的PSO-GRNN模型性能较稳健,在100次循环随机试验中辨识效果仍然较突出,计算得到的上述评价指标均值为97.42、0.04、0.04、0.12、0.89、0.02,其对应的标准差为3.53、0.05、0.04、0.08、0.11、0.02,将PSO-GRNN模型与HHT样本熵结合可作为天然地震与人工爆破事件辨识的有效方法.A new method of earthquake type identification based on particle swarm optimization generalized regression neural network and HHT sample entropy is proposed to realize the accurate identification of earthquake types since natural earthquakes and artificial blast signals have many similar characteristics.The method first decomposes the initial signal into six eigenmode functions(IMFs)and one residual vector by Empirical Modal Decomposition(EMD),and then performs Hilbert transform on the first three IMFs to obtain instantaneous frequency and instantaneous energy,and then extracts the sample entropy and constructs the training and test sets of the neural network;optimizes the smooth factor parameter values of the generalized regression neural network using the particle swarm algorithm and the training set.The PSO-GRNN classification model is established.The results of this model were compared with the neural network models of BPNN,GRNN,PNN and RBFNN,and the accuracy,RMSE,MAPE,MAE,MSE and R2 of this model were 95%,0.1604,0.1204,0.2381,0.7123 and 0.0567,respectively,and most of the performance evaluation metrics are better than the other four neural network models.The PSO-GRNN model established in this study has a more robust performance,and the discrimination effect is still outstanding in 100 cycles of randomized tests,and the mean values of the above evaluation indexes are calculated as 97.42%,0.04,0.04,0.12,0.89,0.02,and their corresponding standard deviations are 3.53,0.05,0.04,0.08,0.11 and 0.02.Combining the PSO-GRNN model with the HHT sample entropy can be an effective method for natural earthquake and artificial blast event identification.

关 键 词：地震辨识 广义回归神经网络 粒子群算法 光滑因子 样本熵 希尔伯特-黄变换 

分 类 号：P315[天文地球—地震学]