摘要
由于震源识别模型判据选取不当或关键参数设置不合理,现有地震辨识模型的性能受到一定制约。为了改善地震类型识别效果,优化反向传播(back propagation,BP)神经网络参数设置,提出了一种经验模态分解(empirical mode decomposition,EMD)多尺度近似熵(multi-scale approximate entropy)判据,并讨论训练函数、激活函数、隐藏层神经元数目与学习速率等网络参数对地震辨识的影响。首先,针对多组天然地震事件及人工爆破事件波形进行关键信息提取与归一化,利用EMD算法提取6个本征模态函数(intrinsic mode function,IMF)和1个剩余分量(residual),然后分别计算近似熵得到EMD多尺度近似熵向量;再次,设计一系列BP神经网络调参实验,研究BP神经网络参数对地震类型识别的影响。结果表明,训练函数、激活函数、隐藏层数目等参数对地震辨识产生了明显的影响,辨识效果较好的参数组合为:训练函数采用共轭梯度法,输出层激励函数为purelin函数,隐层激励函数为logsig函数,隐层节点数和学习率分别设置为10和0.01,此时的Accuracy、Speed、MSE分别为97.0897%、0.3059 s、0.0382。该研究弥补了神经网络调参和EMD多尺度近似熵判据在震源识别领域的空白,有助于提高与改善地震识别的准确率和稳定性。
Due to improper selection of the identification model criterion or unreasonable setting of key parameters,the performance of seismic discrimination models is restricted to a certain extent.In order to improve earthquake type recognition and optimize back propagation(BP)neural network parameter settings,an empirical mode decomposition(EMD)multi-scale approximate entropy criterion was proposed,and the effects of parameters such as training function,activation function,number of hidden layer neurons and learning rate were discussed on seismic recognition.Firstly,the key information was extracted and normalized for multiple sets of natural seismic events and artificial blast event waveforms.Secondly,six IMFs and one residual component were extracted from the processed waveform using the EMD algorithm.the approximate entropy was calculated separately to obtain the multi-scale approximate entropy vector.Finally,a series of BP neural network tuning experiments were designed and effects of BP neural network parameters for earthquake recognition was studied.Experimental results show that parameters such as training function,activation function,and number of hidden layer neurons have a significant impact on seismic identification.The parameter combinations with outstanding recognition effect are:the training function adopts the conjugate gradient method,the output layer excitation function is the purelin function,the hidden layer excitation function is the logsig function,the number of hidden layer nodes and the learning rate are set to 10 and 0.01,respectively.Meanwhile,the values of accuracy,speed,and MSE is 97.0897%,0.3059 s,and 0.0382,respectively,the study fills the gap of BP parameter adjustment in the field of seismic recognition,and helps to increase and improve the accuracy and stability of seismic recognition.
作者
庞聪
江勇
吴涛
廖成旺
马武刚
PANG Cong;JIANG Yong;WU Tao;LIAO Cheng-wang;MA Wu-gang(Institute of Seismology, China Earthquake Administration, Wuhan 430071, China;Hubei Key Laboratory of Earthquake Early Warning, Wuhan 430071, China;Hubei Earthquake Administration, Wuhan 430071, China)
出处
《科学技术与工程》
北大核心
2022年第18期7765-7772,共8页
Science Technology and Engineering
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(IS201856290)
中国大陆综合地球物理场仪器研发专项(Y201707)
中国地震局地震科技星火计划攻关项目(XH15030)。
关键词
天然地震
人工爆破
反向传播神经网络
影响因素
经验模态分解
多尺度近似熵
natural earthquake
artificial blasting
back propagation neural network
influencing factors
empirical mode decomposition
multi-scale approximate entropy