Research on Discrimination Methods of Near-field and Far-field Long-period Ground Motions Based on Hybrid Machine Learning Models
-
摘要: 长周期地震动因具有丰富的低频成分而易使超高层建筑、大跨桥梁、大型生命线工程等长周期结构产生震害，根据形成机制与特征不同，可分为近场长周期地震动和远场长周期地震动两大类。目前对于长周期地震动的界定方法尚未有统一标准，现有方法多根据单一参数判定，无法表征长周期地震动的复杂特性，且缺少对不同类型长周期地震动的界定。鉴于此，本文从人工多角度提取特征和自动拾取时空联合特征两个角度，分别基于机器学习和深度学习搭建PCA-SVM模型和CNN-LSTM-Attention模型，实现了对近、远场长周期地震动的高效判别，在测试集上分别达到了96.57%和97.06%的准确率。结果表明，本文提出的两种分类模型对近、远场长周期地震动均有较好的判别效果，且相较于传统方法更具优越性，可为长周期地震动的识别与选取提供参考。Abstract: Long-period ground motions (LPGM), due to their abundant low-frequency components, are prone to causing earthquake damage to long-period structures such as super-high-rise buildings, long-span bridges, and large-scale lifeline engineerings, based on different formation mechanisms and characteristics, they can be divided into near-field long-period ground motions and far-field long-period ground motions. At present, there is no unified standard for the definition of LPGM, existing methods are mostly based on a single parameter, which cannot characterise the complex characteristics of LPGM and lack definition for different types of LPGM. In view of this, this study employs both manual multi-angle feature extraction and automated capturing of spatio-temporal features to establish a PCA-SVM model based on machine learning and a CNN-LSTM-Attention model based on deep learning, these approaches achieve effective classification of near-field and far-field long-period ground motions, attaining test accuracies of 96.57% and 97.06% respectively. The results show that the two classification models proposed in this study have good prediction effects on near-field and far-field long-period ground motions, and have superiority compared with the traditional methods, which can provide a reference for the identification and selection of LPGM.
-
Key words:
- Long-period ground motion recognition /
- SVM /
- CNN /
- LSTM /
- Attention mechanism
1)1 2http://peer.berkeley.edu./ -
图 12 不同特征输入组合下的混淆矩阵
Figure 12. Confusion matrix for different combinations of feature inputs
图 13 CNN-LSTM-Attention地震动识别基本框架
Figure 13. Basic framework of CNN-LSTM-Attention ground motion recognition
图 17 消融实验各模型在测试集上的混淆矩阵
Figure 17. Confusion matrix for each model of the ablation experiment on the test set
图 18 加速度时程曲线和频谱特征
Figure 18. Acceleration time history curve and spectral characteristics
表 1 特征参数及其定义
Table 1. Characteristic parameters and their definitions
参数性质 特征参数 特征定义 时域特征 PGA 地震动加速度时程中最大幅值的绝对值。 PGV 地震动速度时程中最大幅值的绝对值。 PGV/PGA 峰值速度PGV与峰值加速度PGA的比值。 90%能量持时T90% 地震动累积能量从5%达到95%所持续的时间。 相对速度脉冲能量Erp 脉冲时程能量与速度时程总能量的比值。 频域特征 Tavg 加速度反应谱平均周期。 TPA 加速度反应谱最大幅值对应的频率。 βl β谱2～10 s谱值的加权平均值。 fmax 傅里叶幅值谱最大幅值对应的频率。 Tmf 傅里叶谱平均周期。 fEmax Hilbert能量谱最大幅值对应的频率。 TmE Hilbert能量谱平均周期。 一般特征 震中距 震中与观测点之间的地球球面距离。 场地类型 依据等效剪切波速VS30将场地划分为A、B、C、D、E。 信号特征 功率 信号在单位时间内传递的能量。 过零率 信号在单位时间内跨越零的频率，体现地震动非平稳性。 
下载: 导出CSV
表 2 网格搜索信息
Table 2. Grid search information
核函数 linear核 rbf核 poly核 sigmoid核 搜索空间 C: 0.01, 0.1, 1,
10, 100,1000
class_weight: balancedC: 0.01, 0.1, 1, 10, 100, 1000
gamma:0.0001 , 0.001,
0.01, 0.1, 1, 10
class_weight: balancedC: 0.01, 0.1, 1, 10, 100, 1000
gamma：0.0001 , 0.001, 0.01, 0.1, 1, 10
degree: 2, 3, 4
coef0: -1, 0, 1
class_weight: balancedC: 0.01, 0.1, 1, 10, 100, 1000
gamma:0.0001 , 0.001, 0.01, 0.1, 1, 10
coef0: −1, 0, 1
class_weight: balanced最佳组合 rbf核：C=100，gamma=0.01，class_weight=balanced
下载: 导出CSV
表 3 五折交叉验证结果
Table 3. Five-fold cross-validation results
最佳参数组合 1折 2折 3折 4折 5折 均值 标准差 kernel=rbf核，
C=100，gamma=0.01
class_weight=balanced0.9451 0.9448 0.9451 0.9571 0.9632 0.9511 0.0086
下载: 导出CSV
表 4 不同特征输入组合下的模型评估指标
Table 4. Model evaluation metrics for different combinations of feature inputs
特征参数组合 精确率 召回率 FI分数 准确率 AUC值 卡帕系数 一般特征（2个） 0.7119 0.7059 0.6966 0.7059 0.8845 0.5605 信号特征（2个） 0.7421 0.7451 0.7432 0.7451 0.9243 0.6134 一般、信号特征（4个） 0.8280 0.8235 0.8243 0.8235 0.9504 0.7344 时域特征（5个） 0.8911 0.8824 0.8839 0.8824 0.9697 0.8234 频域特征（7个） 0.8616 0.8431 0.8471 0.8431 0.9577 0.7652 时域、频域特征（12个） 0.9510 0.9510 0.9510 0.9510 0.9905 0.9259 时、频域、一般特征（14个） 0.9609 0.9608 0.9608 0.9608 0.9915 0.9407 时、频域、信号特征（14个） 0.9410 0.9412 0.9411 0.9412 0.9905 0.9110 全部特征（16个） 0.9660 0.9660 0.9660 0.9657 0.9924 0.9481
下载: 导出CSV
表 5 超参数优化结果
Table 5. Hyperparameter optimization results
超参数 卷积层1通道数 卷积层2通道数 LSTM隐藏单元数 学习率 测试集准确率 搜索空间 8/16/32/64 8/16/32/64 128/256/512/ 1024 10−5 ～ 10−1 97% 取值 64 64 256 0.0001
下载: 导出CSV
表 6 五折交叉验证结果
Table 6. Five-fold cross-validation results
折数 训练准确率/% 测试准确率/% 训练损失值 测试损失值 F1分数 卡帕系数 1 99.1422 94.1176 0.0247 0.5451 0.9409 0.9112 2 98.6520 90.6863 0.0516 0.4421 0.9065 0.8591 3 98.2843 93.1373 0.0413 0.3565 0.9312 0.8962 4 99.5098 94.1176 0.0149 0.2887 0.9409 0.9102 5 99.5098 94.6078 0.0190 0.1723 0.9460 0.9187 均值 99.0196 93.3333 0.0303 0.3609 0.9331 0.8991 标准差 0.0054 0.0157 0.0156 0.1426 0.0158 0.0238
下载: 导出CSV
表 7 消融实验评估结果
Table 7. Evaluation results of ablation study
基准模型 精确率 召回率 F1分数 准确率 AUC值 卡帕系数 CNN 0.9152 0.9118 0.9118 0.9118 0.9344 0.8651 CNN-LSTM 0.9419 0.9412 0.9405 0.9412 0.9862 0.9103 CNN-LSTM-Attention 0.9705 0.9706 0.9704 0.9706 0.9928 0.9553
下载: 导出CSV
表 8 本文模型与传统方法识别结果
Table 8. Identification results of this paper's model and traditional methods
方法 SVM-PCA法 CNN-LSTM-Attention法 PGV/PGA法 βl值法 LPGI法 准确率 96.49% 95.61% 83.33% 81.58% 77.19%
下载: 导出CSV
表 9 示例地震动相关信息
Table 9. Relevant information on example ground motion
地震动记录 震级 震中距/km 场地类型 Tmf/s Tavg/s Tp/s PGV/PGA βl LPGI H-EMO270 6.53 0.07 D 1.88 1.83 3.08 0.34 0.24 9.99×10−1 JEN292 6.69 5.43 D 1.01 1.06 1.85 0.16 0.09 1.81×10−8 KAU086-EW 7.62 128.8 D 1.52 1.79 — 0.26 0.41 5.40×10−6
下载: 导出CSV
-
党育, 李连峰, 2018. 长周期地震动的界定及对应场地发生概率. 兰州理工大学学报, 44(6): 112−117.Dang Y., Li L. F., 2018. Definition of long-period ground motion and its occurrence probability on variously classified site. Journal of Lanzhou University of Technology, 44(6): 112−117. (in Chinese) 丁昊青, 2017. 基于机器学习算法的脉冲型地震动识别. 哈尔滨: 哈尔滨工业大学.Ding H. Q., 2017. Identification of pulselike ground motion based on machine learning algorithm. Harbin: Harbin Institute of Technology. (in Chinese) 窦世昌, 2020. 长周期地震动作用下高层结构地震反应及易损性分析. 哈尔滨: 哈尔滨工业大学.Dou S. C., 2020. Seismic response and fragility analysis of high-rise structures under long-period ground motion. Harbin: Harbin Institute of Technology. (in Chinese) 方登甲, 2022. 长周期地震下基于性能的高层斜交网格结构易损性评估. 成都: 西南交通大学.Fang D. J., 2022. Performance-based seismic fragility assessments of high-rise diagrid structures under long-period ground motion inputs. Chengdu: Southwest Jiaotong University. (in Chinese) 李梅, 宁德军, 郭佳程, 2019. 基于注意力机制的CNN-LSTM模型及其应用. 计算机工程与应用, 55(13): 20−27.Li M., Ning D. J., Guo J. C., 2019. Attention mechanism-based CNN-LSTM model and its application. Computer Engineering and Applications, 55(13): 20−27. (in Chinese) 李雪红, 王文科, 吴迪等, 2014. 长周期地震动的特性分析及界定方法研究. 振动工程学报, 27(5): 685−692.Li X. H., Wang W. K., Wu D., et al., 2014. The bounded method and characteristics analysis for long-period ground motions. Journal of Vibration Engineering, 27(5): 685−692. (in Chinese) 李英民, 赵晨晓, 谭潜, 2018. 基于HHT地震动分量分离的长周期地震动界定方法. 振动与冲击, 37(7): 164−171, 219.Li Y. M., Zhao C. X., Tan Q., 2018. Classification method for long period ground motions based on component decomposition with HHT. Journal of Vibration and Shock, 37(7): 164−171,219. (in Chinese) 陆继翔, 张琪培, 杨志宏等, 2019. 基于CNN-LSTM混合神经网络模型的短期负荷预测方法. 电力系统自动化, 43(8): 131−137.Lu J. X., Zhang Q. P., Yang Z. H., et al., 2019. Short-term load forecasting method based on CNN-LSTM hybrid neural network model. Automation of Electric Power Systems, 43(8): 131−137. (in Chinese) 牛志辉, 陈波, 卜春尧, 2021. 基于卷积神经网络的速度大脉冲识别方法研究. 震灾防御技术, 16(3): 485−491.Niu Z. H., Chen B., Bu C. Y., 2021. Identification of pulse-like strong ground motions based on convolution neural network. Technology for Earthquake Disaster Prevention, 16(3): 485−491. (in Chinese) 苏闻浩, 刘启方, 2024. 基于深度神经网络的地表地震动幅值预测研究. 震灾防御技术, 19(2): 387−396.Su W. H., Liu Q. F., 2024. Study on the prediction of ground motion amplitude based on deep neural network. Technology for Earthquake Disaster Prevention, 19(2): 387−396. (in Chinese) 王博, 2013. 多类型长周期地震动作用下高层结构的动力行为与设计方法研究. 西安: 西安建筑科技大学. Wang B., 2013. Study on dynamic behavior and design method of high-rise structures under multi-type long-period ground motions. Xi’an: Xi’an University of Architecture and Technology. (in Chinese). 王博, 刘伯权, 吴涛等, 2018. 远场长周期地震动频谱特征周期与强度指标研究. 振动与冲击, 37(7): 211−219.Wang B., Liu B. Q., Wu T., et al., 2018. Characteristic periods and intensity indexes for far-field long-period ground motion frequency spectra. Journal of Vibration and Shock, 37(7): 211−219. (in Chinese) 张宸嘉, 朱磊, 俞璐, 2021. 卷积神经网络中的注意力机制综述. 计算机工程与应用, 57(20): 64−72.Zhang C. J., Zhu L., Yu L., 2021. Review of attention mechanism in convolutional neural networks. Computer Engineering and Applications, 57(20): 64−72. (in Chinese) 张亮泉, 于建杰, 2019. 近、远场长周期地震动界定及分类研究. 自然灾害学报, 28(2): 29−36.Zhang L. Q., Yu J. J., 2019. Definition and classification of near and far field long period ground motions. Journal of Natural Disasters, 28(2): 29−36. (in Chinese) Bharadiya J. P., 2023. Convolutional neural networks for image classification. International Journal of Innovative Science and Research Technology, 8(5): 673−677. Boser B. E. , Guyon I. M. , Vapnik V. N. , 1992. A training algorithm for optimal margin classifiers. In: Proceedings of the Fifth Annual Workshop on Computational Learning Theory. Pittsburgh: ACM, 144−152. Cortes C., Vapnik V., 1995. Support-vector networks. Machine Learning, 20(3): 273−297. doi: 10.1023/A:1022627411411 Habib A., Youssefi I., Kunt M. M., 2022. Identification of pulse-like ground motions using artificial neural network. Earthquake Engineering and Engineering Vibration, 21(4): 899−912. doi: 10.1007/s11803-022-2127-y Hochreiter S., Schmidhuber J., 1997. Long short-term memory. Neural Computation, 9(8): 1735−1780. doi: 10.1162/neco.1997.9.8.1735 Jolliffe I. T. , Cadima J. , 2016. Principal component analysis: a review and recent developments. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 374(2065): 20150202. Krizhevsky A., Sutskever I., Hinton G. E., 2017. ImageNet classification with deep convolutional neural networks. Communications of the ACM, 60(6): 84−90. doi: 10.1145/3065386 Liao W. I., Loh C. H., Wan S., 2001. Earthquake responses of RC moment frames subjected to near-fault ground motions. The Structural Design of Tall Buildings, 10(3): 219−229. doi: 10.1002/tal.178 Wani F. M., Vemuri J., 2024. Shapelet-informed machine learning classifiers: a path towards precise identification of pulse-like ground motions. Journal of Earth System Science, 133(2): 96. doi: 10.1007/s12040-024-02314-2 -
点击查看大图
计量
- 文章访问数: 36
- HTML全文浏览量: 8
- PDF下载量: 6
- 被引次数: 0
