基于MFCC样本熵和灰狼算法优化支持向量机的天然地震与人工爆破自动识别  

Automatic recognition of natural earthquakes and artificial blasting based on the sample entropy of the Mel frequency cepstrum coefficient and support vector machine optimized by gray wolf optimization

作　　者：庞聪[1,2,3] 江勇 廖成旺 吴涛[1,2,3] 丁炜 PANG Cong;JIANG Yong;LIAO Chengwang;WU Tao;DING Wei(Institute of Seismology,CEA,Wuhan 430071,Hubei,China;Hubei Key Laboratory of Earthquake Early Warning,Wuhan 430071,Hubei,China;Hubei Earthquake Agency,Wuhan 430071,Hubei,China)

机构地区：[1]中国地震局地震研究所,湖北武汉430071 [2]地震预警湖北省重点实验室,湖北武汉430071 [3]湖北省地震局,湖北武汉430071

出　　处：《地震工程学报》2022年第5期1169-1175,共7页China Earthquake Engineering Journal

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,IS2018126178,IS201726156);中国大陆综合地球物理场仪器研发专项(Y201707)。

摘　　要：针对天然地震与人工爆破波形特征相似、难以区分的情况,结合灰狼优化算法和支持向量机,提出一种地震事件性质辨识新方法。通过梅尔频率倒谱系数法对2013年四川芦山地震地震事件信号和人工爆破信号进行分析,进过预加重、FFT、梅尔滤波及离散余弦变换等步骤,提取静态系数样本熵、一阶差分系数样本熵和二阶差分系数样本熵等作为样本特征集。使用灰狼算法优化支持向量机径向基核函数RBF中的惩罚系数和核函数半径形成新的GWO-SVM分类器,然后对事件进行辨识。结果表明:GWO-SVM分类器辨识效果明显优于SVM、RobustBoost集成学习、LDA、PLDA等分类器,其在1000次循环识别实验下的准确率均值相对SVM提高了9.2个百分点,标准差降低了3.2以上;t检验证明MFCC样本熵各特征具有可靠的地震事件分类效果;GWO-SVM与MFCC样本熵可作为天然地震事件与人工爆破事件的辨识方法与分类判据。The waveform characteristics of natural earthquakes and artificial blasting are similar but difficult to distinguish.Combined with gray wolf optimization(GWO)and support vector machine(SVM),a new method for identifying the nature of seismic events is proposed in this paper.The signals of seismic events during the Yushu M 7.0 earthquake in Qinghai Province and some artificial blasting events were analyzed by the Mel frequency cepstrum coefficient(MFCC)method.Through the pre-emphasis,fast Fourier transform,Mel filter,and discrete cosine transform,the sample entropies of the static coefficient and first-order and second-order differential coefficients were extracted as the sample feature set.GWO was used to optimize the penalty coefficient and kernel radius in the radial basis kernel function of SVM to form a new GWO-SVM classifier.Then,the GWO-SVM classifier was used to identify events.The results show that the recognition effect of the GWO-SVM classifier is obviously better than that of other classifiers,i.e.,SVM,RobustBoost ensemble learning,linear discriminant analysis(LDA),and probabilistic LDA.Under 1000 cycles of recognition experiments,the average accuracy of the GWO-SVM classifier increased by 9.2%compared with that of SVM,and the standard deviation was reduced by more than 3.2.The t-test proves that the MFCC sample entropy has a reliable earthquake event classification effect,and the GWO-SVM and MFCC sample entropy can be used as identification methods and classification criteria for natural earthquake events and artificial blasting events.

关 键 词：梅尔频率倒谱系数 样本熵 灰狼算法 支持向量机 径向基核函数 自动识别 

分 类 号：P315[天文地球—地震学]