基于神经网络的地磁观测数据重构研究

Reconstruction of geomagnetic data based on artificial neural network

在距离数据缺失台站一定范围内选取参考台作为输入,构建非线性BP神经网络并进行地磁观测数据重构研究.数据仿真结果显示,重构数据和原始记录数据吻合程度较高,重构残差较小,磁静日重构平均残差仅为0.11nT,磁扰日平均重构残差为0.23nT.重点对磁场活动最剧烈时段内的数据进行了短时重构,平均残差由0.4nT降低到0.2nT,重构效果得到较大改进.计算了原始数据与重构数据的功率谱密度,除部分高频信号外,二者变化特征基本相同,相关性高达1.0.从时域和频域验证了BP神经网络在地磁相对记录数据重构上的有效性,并将其运用于实际缺失数据重构,取得较好效果.

In this work,we used the back-propagation neural network to reconstruct missing geomagnetic data based on data of adjacent observatories.The simulation results show that reconstructed data are close to original data.The correlation of their power spectral density is about1.0.The average residual between reconstructed data and original data is about 0.11 nT in quiet days,and reaches 0.23 nT in disturbed days.When we use this method to reconstruct the data of a short time period with intense magnetic disturbance,the residual reduces to 0.2 nT from 0.4 nT.Based on the comparison of time and frequency domains,we suggest that the back-propagation network is an effective tool for geomagnetic data reconstruction.