基于深层神经网络压制多次波

有效压制多次波一直是地震勘探中的难点问题.尽管已发展了多种多次波压制方法，但仍存在多次波压制不全、计算耗时长等缺陷，使得应对复杂地质地震数据多次波压制具有挑战性.为了突破现有多次波压制方法的局限性，本文提出了一种基于深层神经网络的多次波压制方法，采用的深层神经网络是一种改进的具有卷积编码器和卷积解码器的U-net网络.不同于常规方法依赖于滤波或波动理论，该方法仅依赖于大量训练数据.训练数据以含多次波的原始地震数据作为输入，不含多次波的地震数据作为输出，通过最小化损失函数来优化神经网络参数.训练成功的网络模型具备较好地分离多次波和一次波的能力，可直接用来快速压制地震数据中的多次波，避免了常规方法涉及的大规模计算.工业界模型数据测试结果表明，本文提出的深层神经网络方法能有效压制复杂地质地震数据中的多次波，同时还具有较高的泛化能力和多次波压制效率.

Application of deep neural networks for multiples attenuation

Multiples attenuation is a persistent problem in seismic exploration. Although a variety of multiples-attenuation methods have been proposed, each of which is still limited, such as incomplete multiples attenuation and a large amount of calculation, which makes it challenging to cope with the massive complex geological seismic data. To break through the limitations of the existing multiples-attenuation methods, we design a deep neural network (DNN) to suppress multiples from seismic data. Unlike conventional methods that rely on filtering or wave theory, the DNN-based method only relies on a large amount of training data. The training data takes the original seismic data with multiples as input and the data without multiples as output. The parameters of the designed DNN can be optimized through minimizing the loss function. The well trained DNN has the ability to separate multiples and primaries, which can be directly used to attenuate multiples in the seismic data, and, as a result, avoids the large-scale calculations involved in conventional methods. The testing results of industrial model data show that the DNN-based method can effectively suppress the multiples in complex geological seismic data, and has high generalization ability and multiples-attenuation efficiency.