【作者】 章涛；

【导师】 吴小平；

【作者基本信息】 中国科学技术大学 ， 地球物理学， 2024， 硕士

【摘要】 中国地震灾害频发,往往造成严重的人员伤亡和巨大的财产损失。地震是全球面临的一项长期而广泛的科学挑战,地震预测是全球研究人员和专家面临的共同问题,因此研究地震前兆,发展地震预测方法的研究显得十分重要。磁场作为一个重要的前兆,成为近年来的研究热点,研究表明地震的发生会导致电离层出现扰动异常,而随着卫星技术的发展,利用卫星搭载高精度载荷来采集磁场数据成为重要手段。本文应用深度学习方法技术分析卫星磁测数据,识别震前磁场异常,并基于卫星数据对地震三要素预测进行了探索,旨在为当前中强地震短临预报系统提供新的思路。对于卫星观测数据,本文选取了德国地球科学研究中心研制发射的CHAM P。地震的孕育往往发生于岩石圈,为了研究岩石圈孕震区的信号对电离层造成的影响,本文基于 IGRF(International Geomagnetic Reference Field,即国际地磁参考场)第12代模型进行了主磁场的去除,随后根据太阳活动指数对受到影响的数据作了进一步处理,为了进一步防止白天电离层活动的影响,我们提取了地方时为夜间的磁场序列,最终得到了 NEC(North East Center)坐标系下去除干扰后的卫星磁场观测三分量值,是后续深度学习模型建立的基础。本文选择全球范围内卫星飞行期间的地震事件在全球范围内展开分析,首先利用循环神经网络对震前电离层异常进行识别,将异常识别转化为了深度学习分类任务。数据集建立上,我们针对每个地震事件进行处理,空间范围筛选了经纬度10°的磁场序列,时间范围分别筛选了震前15天、震前7天、震前24h,分别采用了 RNN、GRU、LSTM网络结构进行了深度学习训练任务,并利用ROC曲线对预测结果进行了评估。此外,本文还对震级和地震位置进行了分析研究,利用卫星观测数据通过神经网络对其进行预测,所训练的模型在测试集上均有不错的表现。最后,本文对利用卫星电磁数据进行地震位置预测的方法进行了改进,通过将地球按照地震的发生情况划分为众多非均匀网格,利用残差网络及Transfor mer网络对震前磁场序列数据集进行训练并在测试集上进行了评估,最高能够达到80%左右的预测准确率。更多还原

【Abstract】 China is notably prone to significant earthquake disasters,often resulting in severe casualties and substantial property losses.Earthquakes pose a persistent and widespread scientific challenge globally,with earthquake prediction being a common issue faced by researchers and experts worldwide,and a challeng-ing task for the earth science community,so the study of precursors of earthquakes and the development of earthquake prediction methods of the research is very important.Magnetic field,as an important precursor,has become a research hotspot in rece-nt years,and studies have s hown that the occurrence of earthquakes will lead to dist-urbing anomalies in the ionosphere,and with the development of satellite technology,the use of satellites carrying high-precision payloads to collect magnetic field data has become an important means.In this paper,we apply a deep learning method to analyse the relationship between the magnetic field data observed by satellites and the precursors of earthquakes,and based on the satellite data,the three-factor earthq-uake prediction has been This paper applies deep learning methods to analyse the re-lationship between satellite-observed magnetic field data and earthquake precursors,and makes a prediction of the three elements of earthquakes based on satellite data,aiming to provide new ideas for the current short-range prediction system of medium-strength earthquakes.For the satellite observation data,this paper selects the CHAMP developed and launched by the German Research Centre for Geosciences(GFZ),the breeding of earthquakes often occurs in the lithosphere,in order to study the impact of the signals of the lithosphere breeding seismic region on the ionosphere,this paper carries out the removal of the main magnetic field based on the 12th generation of the IGRF m-odel,and then according to the solar activity index of the data affected by the further process-ing,and selects the data of the night of the local time in order to remove the interference of the ionosphere during the daytime,and ultimately processed to get the triangular value of the satellite magnetic field observation in the NEC coordinate system after re-moving the interference,which lays a foundation for the subsequent input of the deep lea-rning model.In this paper,we select the seismic events during the satellite flights on a global scale to be analysed globally,and firstly,we use recurrent neural networks to identify the pre-seismic ionospheric anomalies,and transform the anomaly identification into a deep learning classification task.For dataset establishment,we processed for each seismic event,the spatial scope screened the magnetic field sequence with 10° latitude and longitude,and the temporal scope screened 15 days before the earthquake,7 days before the earthquake,and 24h before the earthquake,respectively,and adopted RNN,GRU,and LSTM network structures for the deep learning training task,and evaluated the prediction results using ROC curves.In addition,this paper also analyses and studies the magnitude and the location of the earthquake and predicts them by neural network using satellite observation data,and the trained models have good performance on the test set.Finally,this paper improves the method of earthquake location prediction using satellite electromagnetic data by dividing the earth into many non-uniform grids according to the occurrence of earthquakes,and using residual networks and Transformer networks to train the pre-seismic magnetic field sequence dataset and evaluate it on a test set,and the highest prediction accuracy can be achieved around.更多还原