基于迭代启发网络算法的非平稳随机噪声压制

常规滤波方法常常放大了噪声的影响，同时噪声的存在也限制了分辨率的提升，并“平滑”了地震数据中的不连续信息。为此，提出了基于迭代启发网络（ⅡN）算法的非平稳随机噪声压制方法，利用迭代启发网络压制非平稳随机噪声，网络结构简单、紧凑。ⅡN由交替方向乘子算法的迭代过程推导而来，利用L₁范数优化变分模型。在训练阶段，通过增加一个新的辅助变量，将目标函数的极值转化为增广拉格朗日格式，使用L-BFGS（Large-Broyden Fletcher Goldforb Shanno）算法判别、训练所有网络参数，最终得到最优去噪模型。理论模型及实际资料的去噪结果表明：①由训练得到的去噪模型根据有效信号的特征，在去噪的同时可保留同相轴的形状特征；采用的迭代网络简单、紧凑，加快了网络的收敛速度，能够用相对较小的数据集和较短的训练时间快速训练去噪模型，达到预期的去噪效果。②所提方法具有较强的适应性，有效地压制了常规地震数据中的非平稳随机噪声。

Iterative scheme inspired network for non-stationary random denoising

Conventional filtering methods often magnify the influence of noise,which in return impedes the improvement of resolution and “smooths” discontinuous information in seismic data.We introduce a non-stationary random noise filtering method based on an iterative scheme-inspired network (ⅡN) which has a simple and tight structure and can be used to smooth non-stationary random noises.The L1 norm is used to optimize the objective function of the alternating directional multiplier algorithm which the ⅡN is derived from.A new auxiliary vari-able is added to transform the extreme value of the objective function into an augmented Lagrange form,and using the L-BFGS algorithm to distinguish and train all the network parameters.Finally an optimal denoising model is obtained.Applications to model and real data show that: ① the trained denoising model can effectively suppress noises while maintaining the characteristics of events according to the features of useful signals; and the simple and tight iterative network can speed up the rate of convergence and rapidly finish denoising and achieve expected results using a smaller database and shorter training time; ② the method proposed has a good adaptability and can suppress non-stationary random noises in conventional seismic data.