基于迭代启发网络算法的非平稳随机噪声压制

Iterative scheme inspired network for non-stationary random denoising

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。

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 variable 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.