内蒙古地区爆破事件记录较多,准确识别天然地震和爆破事件成为该区地震监测工作的重要任务。选取内蒙古地震台网2010—2024年记录的200次天然地震和167次爆破事件观测资料,提取3 505条天然地震和3 034条爆破的单通道记录,截取包含主要震相的100 s时间窗口记录,利用短时傅里叶变换,将地震信号转换为时频灰度图像,作为训练和测试数据集,将事件分类问题转化为图像分类问题,并选择ResNet-18残差神经网络进行分类研究。实验结果表明,ResNet-18模型在识别天然地震与人工爆破中有优异表现,准确率达98.16%,精确率为99.56%,召回率为97.02%,F1分数为0.982 9,马修斯相关系数(MCC)为0.961 5,PR-AUC和ROC-AUC分别为0.997 87和0.997 72,证明本研究数据处理方法和选用的深度学习模型具有较好的分类效果,在地震监测预报领域具有应用潜力。
In the Nei Monggol region, there are many records of blasting events. Classification of natural earthquakes and blasts is an important task in earthquake monitoring in this region. This paper selects the observation data of 200 natural earthquakes and 167 blasts recorded by the Nei Monggol Seismic Network from 2010 to 2024, extracts 3 505 single-channel records of natural earthquakes and 3 034 single-channel records of blasts, intercepts a 100-second time window record containing the main seismic phases, and uses Short-Time Fourier Transform to convert seismic signals into grayscale images of time-frequency diagrams as training and testing data sets, converting the event classification problem into an image classification problem. The ResNet-18 residual neural network is selected for classification research. The experimental results show that the ResNet-18 model has excellent performance in identifying natural earthquakes and artificial blasts, with an accuracy of 98.16%, a precision of 99.56%, a recall of 97.02%, an F1 score of 0.982 9, a Matthews correlation coefficient (MCC) of 0.961 5, and PR-AUC and ROC-AUC of 0.997 87 and 0.997 72, respectively, proving that the data processing method and the selected deep learning model in this paper have good classification effects and have application potential in the field of earthquake monitoring and prediction.
2025,46(2): 50-58 收稿日期:2024-7-31
DOI:10.3969/j.issn.1003-3246.2025.02.005
基金项目:2023年度内蒙古自治区地震局局长基金(项目编号:2023QN18)
作者简介:堵伟鹏(1992—),女,本科,助理工程师,主要从事地震监测与预报工作。E-mail:649405278@qq.com
*通讯作者:王怡(1988—),女,本科,工程师,主要从事地震监测与预报工作。E-mail:120578747@qq.com
参考文献:
陈银燕. 基于HMM和GMM天然地震与人工爆破识别算法研究[D]. 桂林:广西师范大学,2011.
黄苇,周捷,高利君,等. 基于同步挤压改进短时傅里叶变换的分频蚂蚁追踪在断裂识别中的应用[J]. 物探与化探,2021,45(2):432-439.
刘莎,杨建思,田宝峰,等. 首都圈地区爆破、矿塌和天然地震的识别研究[J]. 地震学报,2012,34(2):202-214.
潘宇曜,陈焯辉,林佩欣,等. 基于ResNet-18网络的桥梁损坏图像分类研究[J]. 科学技术创新,2023,(16):93-96.
田宵,汪明军,张雄,等. 基于多输入卷积神经网络的天然地震和爆破事件识别[J]. 地球物理学报,2022,65(5):1 802-1 812.
王梦琪,黄汉明,吴业正,等. 基于多尺度注意残差网络的地震波形分类研究[J]. 地震工程学报,2024,46(3):724-733.
隗永刚,杨千里,王婷婷,等. 基于深度学习残差网络模型的地震和爆破识别[J]. 地震学报,2019,41(5):646-657.
杨千里,王婷婷,边银菊. 基于广义S变换的地震与爆炸识别[J]. 地震学报,2020,42(5):613-628.
曾融生,陈运泰,吴忠良. 探测地球内部的“雷达”——地震波(续)[J]. 城市防震减灾,2000,3(6):12-14.
张帆,杨晓忠,吴立飞,等.基于短时傅里叶变换和卷积神经网络的地震事件分类[J].地震学报,2021,43(4):463-473.
He K M, Zhang X Y, Ren S Q, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition Las Vegas: IEEE, 2016: 770-778.