基于深度卷积生成对抗网络的瑞雷波信号随机噪声去除

Deep convolutional generative adversarial network for random noise attenuation in Rayleigh wave signal

瑞雷波勘探是一种新兴的环境与工程地球物理勘探方法,野外采集的瑞雷波数据常含各种噪声干扰,对瑞雷波信号进行去噪处理至关重要.本文提出了一种基于深度卷积生成对抗网络(DCGAN)的瑞雷波信号随机噪声去除方法,其关键在于构建一个适用于瑞雷波信号去噪的DCGAN,包含生成器与判别器两部分.生成器由一个全卷积神经网络(FCN)构成,用于学习含噪数据到无噪数据的特征映射;判别器由一个卷积神经网络(CNN)构成,用于辅助生成器训练.基于DCGAN的去噪方法的实现分为三步:数据预处理、网络训练和噪声去除.一旦网络训练完成,去噪过程无需更多人为调整参数,减少了人力成本.对实际数据进行去噪试验,从地震数据和频散曲线两方面评价去噪效果,验证了本文方法的可行性,且相较于常用的小波变换、F-X反褶积算法,本文方法在不同比例噪声情况下均具有更好的去噪效果,对瑞雷波信号去噪方法选取有一定参考意义.

Rayleigh wave exploration is an emerging method of environmental and engineering geophysical exploration. Rayleigh wave data collected in the field often contains various noise. In order to denoise the Rayleigh wave signal, we introduce the deep convolution generating adversarial network in the field of deep learning into the Rayleigh wave exploration field in engineering exploration as a new method of Rayleigh wave signal denoising. We develop a Rayleigh wave signal denoising method based on the Deep Convolutional Generative Adversarial Network(DCGAN). The key requirement of the algorithm is to construct a DCGAN that is suitable for random noise attenuation in Rayleigh wave signal, which includes a generator and a discriminator. The discriminator is composed by a Convolutional Neural Network(CNN), which is used to aid the training of the generator. The generator composed by a Fully Convolutional Network(FCN), which is designed to learn a mapping from noisy data to non-noisy data. The application of the DCGAN includes three steps: data preprocessing, training, and noise attenuation. Convolutional layers are added with activation functions, using batch normalization optimization algorithms, and discarding the regularization layer to prevent overfitting, so that DCGAN training can guarantee higher resolution and accuracy, keeping more details of the original data makes the training process more stable. Once the network training is completed, the denoising process does not require more manual adjustment of parameters, reducing labor costs. The denoising test was performed on the actual data, for Rayleigh wave data with different amplitude noise, the methods in this paper have certain feasibility, and have achieved good results for Rayleigh wave signal denoising. The denoising effect was evaluated from the two aspects of seismic data and dispersion curve. Comparing with the common methods such as wavelet transform method and F-X deconvolution method, DCGAN based method shows better accuracy in the case of different proportions of noise, which has certain reference significance for the choice of Rayleigh wave signal denoising method.