基于ST-CNN的脉冲型地震动与脉冲周期融合识别方法  

A hybrid method to identify pulse-like ground motions and pulse periods based on ST-CNN

作　　者：禹海涛[1,2] 朱晨阳 傅大宝 许乃星 卢哲超 蔡辉腾[4] YU Haitao;ZHU Chenyang;FU Dabao;XU Naixing;LU Zhechao;CAI Huiteng(State Key Laboratory of Disaster Reduction in Civil Engineering,Tongji University,Shanghai 200092,China;Department of Geotechnical Engineering College of Civil Engineering,Tongji University,Shanghai 200092,China;Fuzhou Research Institute of Sustainable Development in Cities Ltd.,Fuzhou 350100,China;Earthquake Administration of Fujian Province,Fuzhou 350003,China)

机构地区：[1]同济大学土木工程防灾减灾全国重点实验室,上海200092 [2]同济大学土木工程学院地下建筑与工程系,上海200092 [3]福州市可持续发展城市研究院有限公司,福建福州350100 [4]福建省地震局,福建福州350003

出　　处：《岩土工程学报》2024年第12期2675-2683,共9页Chinese Journal of Geotechnical Engineering

基　　金：国家重点研发计划项目(2022YFE0128400);国家自然科学基金面上项目(42177134);中央高校基本科研业务费专项资金项目。

摘　　要：如何快速准确地识别脉冲型地震动是困扰学术界和工程界的关键难题,定量识别方法虽然能够克服人工识别的经验性限制,但是传统定量识别方法存在识别结果不一致、适用范围不广泛、难以同时识别脉冲周期或识别的脉冲周期部分情况下差异明显等问题。为此建立了一种问题针对性融合学习规则并结合卷积神经网络(CNN),开发出了一种新的脉冲型地震动与脉冲周期同步识别方法。该学习规则通过对基于不同识别原理的多个传统典型识别方法进行融合学习并采用全球范围的30000条任意方向地震动数据进行训练和验证,摒弃了以往繁琐的人工标记过程并得到了3个问题针对性识别模型,分别命名为Strict识别模型、General识别模型以及TP识别模型。除此之外,为解决地震动时序输入信息不足从而导致模型泛化能力较弱的问题,对CNN的输入结构进行了优化增强,提出了ST-CNN模型。其引入了S变换层以将地震动时序变换至时频,从而增加了频域分布信息并进一步提高了识别精度。结果表明:Strict识别模型能严格区分脉冲型与非脉冲型地震动,识别结果得到已有方法的一致认可;General识别模型的识别能力更强,适用范围更加广泛;TP识别模型识别的脉冲周期更加准确,并可与前述识别模型并用以同步输出识别结果。提出的问题针对性融合学习规则还可推广至其他工程领域与其他机器学习模型,建立的识别方法可为脉冲型地震动研究提供科学指导。The rapid and precise identification of the pulse-like ground motions is a key challenge that perplexes both the academic and engineering communities.The quantitative identification methods can overcome the empirical limitations of manual identification.However,the traditional quantitative identification methods suffer from inconsistencies in the identified results,limited applicability,and difficulties in simultaneously determining the accurate pulse periods.In response,a problem-targeted fusion learning rule is established,combined with a convolutional neural network(CNN)model,to develop a novel method to synchronously identify pulse-like ground motions and their pulse periods.This learning rule integrates multiple traditional typical identification methods based on different identification principles,thereby eliminating the cumbersome manual labeling process.It employs 30000 ground motion data from arbitrary directions worldwide for training and validation,resulting in three problem-targeted CNN models named the Strict,General,and TP identification models.To address the issue of insufficient temporal input information for ground motions leading to weak model generalization capability,the input structure of the CNN model is optimized,and the ST-CNN model is proposed,incorporating the S-transform layer to convert ground motion time series to time frequency,thereby enhancing frequency domain distribution information and further improving the identification accuracy.The results indicate that the Strict model can strictly differentiate between the pulse-like and non-pulse-like ground motions,with the results consistent with those of other methods.The General model can identify more pulse-like ground motions and has broader applicability.The TP model accurately identifies pulse periods and can be used in conjunction with the aforementioned models to synchronously output the identified results.The proposed problem-targeted fusion learning rule can also be extended to other engineering fields and other machine learnin

关 键 词：脉冲型地震动 脉冲周期 识别方法 卷积神经网络 S变换 

分 类 号：TU411[建筑科学—岩土工程]