基于自注意力机制的卷积自编码器多次波压制方法

A multiple suppression method based on self-attention convolutional auto-encoder

地震数据的智能化处理可以降低人工成本,减少对未知先验信息的依赖,提升数据处理效率。在地震勘探数据中多次波通常被视作噪声,需要基于一定的数学物理模型对其进行压制或分离。研究利用与多次波全局时空高度相关的自注意力卷积自编码器神经网络压制多次波,可以避免实际计算中的超参数选取,大幅提高计算效率。其中,自注意力机制可以提升网络性能。将实测地震数据成像道集作为神经网络输入,使用商业软件将多次波压制后的结果作为标签数据,利用10%的工区地震数据训练神经网络以及90%的工区地震数据测试神经网络。神经网络测试的输出结果与标签数据的残差均值为0.0014,两者差距极小,说明使用该神经网络压制多次波的结果是正确的。与传统方法相比,基于自注意力机制的卷积自编码器多次波压制方法只需人工处理小样本量数据,再进行神经网络训练便可处理工区的大体量地震数据,为实际地震数据的多次波压制提供了一种有效且高效率的智能化处理方法。

Research on the intelligent processing of seismic data can reduce labor costs and dependence on unknown prior information,as well as improve data processing efficiency.Multiple waves are traditionally regarded as a type of noise in seismic exploration data and must be suppressed or separated based on certain mathematical and physical models.This study used a self-attention convolutional auto-encoder neural network that focuses on the global spatiotemporal correlation of multiples to suppress multiples,avoid the selection of hyperparameters in actual calculations,and significantly improve the calculation efficiency.Images of the measured seismic data were used as the input of the neural network,and the results of multiple suppression using a commercial software were used as the label data.Of the field datasets,10%were used to train the network in this study,and the remaining 90%were used for testing.The tested average residual between the output and label data is 0.0014.This result demonstrates that the gap between the network output and label data is extremely small.In other words,compared with the traditional method,the result of using this network to suppress multiple waves in seismic data is correct,and only a small amount of data needs to be processed manually.A suitable network can be trained to process a large amount of seismic data in a work area.This study affords an effective,intelligent,and efficient processing method for multiple suppression of actual data.