应用平稳小波变换与深度残差网络压制地震随机噪声

常规去噪方法众多,但每种方法都受某种假设或条件限制。另外,常规去噪方法中一些优化问题具有多个局部极值,导致算法可能收敛到局部最优解而非全局最优解。为此,提出了一种基于平稳小波变换与深度残差网络的地震随机噪声压制方法。采用残差网络(ResNet)的拓扑结构,结合平稳小波变换压制地震数据噪声。残差模块有效避免了网络过深引起的梯度消失或计算消耗但损失函数趋于饱和的问题。另外,小波变换是一种高效的特征提取方法,可获得信号低频和不同方向高频特征信息,分区域学习信号或噪声的特征。首先,对Train400数据集中的每幅图片旋转不同角度以增加训练集数据量,经过旋转变换后加入高斯噪声。然后,对每幅图片进行1级平稳Haar小波分解,得到训练数据集;通过训练提取信号中噪声的小波变换高、低频信息,在此基础上通过直连通道,从含噪数据的小波分解中减去学习到的噪声的小波分解,得到去噪信号的小波分解。最后,通过逆平稳小波变换得到去噪信号。合成信号和实际地震数据去噪试验表明,所提方法能较好地压制地震随机噪声,去噪信号的信噪比、峰值信噪比均较高。

Seismic random noise attenuation based on stationary wavelet transform and deep residual neural network

There are many conventional denoising methods,but each method is limited by certain assumptions or conditions.In addition,multiple local extrema may cause the denoising algorithm to converge to a local optimal solution instead of the global one.For this reason,a random noise suppression method based on the stationary wavelet transform and deep residual neural network(WaveResNet)is proposed.It combined the topology structure of the residual neural network(ResNet)with the stationary wavelet transform.The residual module effectively avoids the vanishing gradient or computational consumption caused by the deep network but loss function saturation.In addition,the wavelet transform is an efficient feature extraction method.It can obtain the low-frequency and high-frequency feature information in different directions and learn the characteristics of signal or noise in different regions.First,each picture in the Train400 dataset is rotated by different angles to increase the amount of data in the training set,after which Gaussian noise is added.Then,the one-level stationary Haar wavelet decomposition is performed on each picture to gain a training dataset.The high-and low-frequency information in the wavelet transform domain is extracted through training.On this basis,the wavelet decomposition of the learned noise is subtracted from that of the noisy data,thus achieving the wavelet decomposition of the denoised signal through the direct channel. Finally,the denoised signal is obtained through the inverse stationary wavelet transform.Experiments of synthetic signals and field seismic data show that the proposed method can suppress seismic random noise well,and the signal-to-noise ratio and its peak of the denoised signal are higher than those of conventional methods.