基于深度残差网络的重力数据去噪重构

Gravity data denoising and reconstruction based on deep residual network

为有效消除各种干扰噪声对实测重力数据的影响,提高重力数据处理的精度,将深度学习算法应用到重力数据处理中。构建了一种基于深度残差网络的含噪重力数据到去噪重构重力数据的非线性映射网络模型(GraResNet)。通过模型数据和实测数据对所提出的重力数据去噪重构方法进行了试验验证。数值实验表明:相较于传统的IIR滤波方法、FIR滤波方法和小波滤波方法,所提方法能够结合学习到的重力数据特征,有效区分重力信号和噪声信号,从而提高重力数据去噪重构精度。实验结果中,所提方法的信噪比相对IIR滤波和FIR滤波方法提高近一倍,其噪声衰减因子高达0.9387;另外,该方法还具有较好的泛化性和鲁棒性,当重力数据中噪声的幅度增大为原始数据标准差的80%和100%时,去噪结果的噪声标准差降低了78.3%和78.4%。

In order to effectively eliminate the influence of various noises on the measured gravity data and improve the accuracy of data processing,the deep learning algorithm to gravity data processing is introduced.A nonlinear mapping network model(GraResNet)based on deep residual network is constructed for gravity data denoising and reconstruction,which is verified by model data and measured data.Numerical experiments show that compared with the traditional IIR filtering method,FIR filtering method and wavelet filtering method,the proposed method can effectively distinguish the gravity signal and noise by combining the learned gravity data features,so as to improve the accuracy of gravity data denoising.The results demonstrate the SNR of the proposed method is nearly doubled compared with IIR filtering method and FIR filtering method,and the noise attenuation factor is up to 0.9387.Moreover,the proposed method owns good generalization and robustness.When the amplitude of noise in the gravity data increases to 80%and 100%of the standard deviation of the original data,respectively,the noise standard deviation of the denoised results decreases by 78.3%and 78.4%,correspondingly.