基于去噪卷积神经网络的面波噪声压制方法

提高地震资料的信噪比是地震数据处理的重要任务之一,与依赖信号模型及其相应先验假设的传统地震噪声衰减算法相比,基于大型训练集的深度神经网络的去噪方法通过对大型数据集进行学习,训练完成后可以对面波进行自适应智能降噪。根据叠前高密度地震数据的特点,建立面波去噪训练库,通过去噪卷积神经网络来衰减地震数据的面波噪声。为了准确高效地提取地震数据面波噪声的特征,采用残差学习和批量标准化相结合的方式来加快训练过程并提高算法的面波去噪效果,去噪卷积神经网络能够有效处理未知噪声水平的面波降噪。模型数据和单点高密度地震数据测试结果表明,常规带通滤波及变分模态分解方法对有效信号损伤较大,而去噪卷积神经网络在高效去除面波噪声的同时能够较好地保护有效信号。

Seismic noise attenuation is an important step in seismic data processing.Denoising based on deep neural networks is different from that achieved via traditional seismic noise attenuation algorithms,which rely on signal models and their corresponding assumptions.Denoising based on deep neural networks achieves noise reduction by training on a large amount of data.After training,the deep neural network can quickly and adaptively remove noise.We used a denoising convolutional neural network to attenuate the surface wave noise in seismic data.To accurately and effectively understand the characteristics of such noise,we used residual learning and batch normalization to improve the training speed and denoising accuracy.A denoising convolutional neural network can efficiently remove surface waves.Through testing,it was observed that conventional band-pass filtering and variational modal decomposition methods can damage the effective signal,whereas the denoising convolutional neural network can effectively remove the surface wave while preserving the effective signal.

国家自然科学基金项目(41504097,418741533)资助。