基于反向传播神经网络的闽台ML震级偏差分析与修正

选取我国台湾地区2012—2018年期间的浅源地震资料，将台湾气象局与福建地震台网中心测定的M_L震级进行对比分析，得出两机构测定的震级之间的差异主要是受地震震级大小、震源深度、震源地理位置等因素的影响，并采用线性回归方法对两机构测定的M_L之间的模型关系进行拟合。与此同时，引入反向传播神经网络技术对两机构测定震级之间的偏差进行预测训练，构建4?9?9?9?4的五层网络模型，利用2012—2017年台湾震例作为训练集进行学习训练，2018年数据作为测试集进行预测效果分析。测试结果表明，经过反向传播神经网络修正后，震级偏差较大改善，基本都控制在?0.4，0.3]之内，预测效果优于传统的线性回归方法，特别是对多震、少震区域震例的修正效果更为显著，进一步验证了反向传播神经网络技术具有较强的非线性拟合能力和泛化能力。 

Analysis and revision of magnitude ML deviation between Fujian and Taiwan based on BP neural network

In this paper, the shallow earthquake events in Taiwan island from 2012 to 2018 were used for analyses on magnitude deviation measured by Taiwan Weather Bureau （TWB） and Fujian Seismic Network Center （FSNC）. The result shows the differences in magnitude between Taiwan region and Fujian region were affected by magnitude value, focal depth and geographic location. And then we made linear regressions for magnitude M_L of the earthquakes measured by TWB and FSNC, respectively. In the meanwhile, a BP neural network with 4?9?9?9?4 five-layer model was introduced into the prediction training of these two regions. For the model, the Taiwan earthquakes in 2012?2017 were taken as training set, and the events in 2018 were used as the testing set. After the revision of BP neural network, the deviation values of magnitude were basically controlled within ?0.4 to 0.3, which is better than those from the traditional linear regression method, especially for multiple earthquakes and minority earthquakes. Furthermore, the testing results also validate the abilities of non-linear data fitting and generalization.