基于Triplet-CNN的强弱地震预判研究

Study on Prediction of Strong and Weak Earthquakes Based on Triplet-CNN

及早判定地震是否为强震,可为快速开展救援行动、减少生命和经济损失争取时间。提出三元组卷积神经网络(Triplet-CNN)模型,对地震强弱进行预判研究。以日本宫城地区为例,通过该地区地震目录获取2000-01-01至2008-12-31间历史地震事件记录,包括震级以及对应的KiK-net和K-net强震仪时间序列数据,利用设计的TripletCNN结构对这些强震仪数据进行训练,实现对强弱地震的快速判定。通过超参数优化,该模型的准确率达到96.85%,精确率96.83%,召回率96.82%,F_(1)值96.82%。将Triplet-CNN分类模型与CNN分类模型、随机森林分类模型和支持向量机(SVM)分类模型进行比较,结果表明CNN震级分类模型具有更高的准确率、精确率、召回率和F_(1)值。基于Triplet-CNN的震级分类模型能有效、可靠地对强弱地震进行预判,从而辅助应急决策,为地震预警工作提供科学依据。

In order to determine early that the occurred earthquake is a strong earthquake,so that rescue operations can be carried out quickly and human lives and economic losses can be reduced,a model based on triplet convolutional neural network(Triplet-CNN)is proposed to in⁃vestigate the prediction of strong and weak earthquakes.Taking the Miyagi region of Japan as an example,the historical seismic events in the region between 2000-01-01 and 2008-12-31,including the specific magnitude and the corresponding KiK-net and K-net strong-motor time series data,are obtained from the earthquake catalog of the region,and the designed Triplet-CNN structure is used to train the magnitude pre⁃diction model by combining these strong-motor data.The model is trained with the designed Triplet-CNN structure and these strong seismome⁃ter data to achieve fast determination of strong and weak earthquakes.Through the optimization of hyperparameters,the model achieves 96.85%accuracy,96.83%precision,96.82%recall,and 96.82%F1 value.Meanwhile,the Triplet-CNN classification model is compared with CNN classification model,random forest classification model and support vector machine(SVM)classification model,and the results show that the CNN seismic classification model has higher accuracy,precision,recall and F1 value,and the proposed Triplet-CNN-based seismic classification model can effectively and reliably make strong and weak earthquake prediction,thus aiding emergency decision-making and providing a favorable complement to earthquake early warning work.