基于自注意力机制深度学习的重磁数据网格化和滤波方法

重磁数据网格化和滤波结果直接影响解释结果,为此设计了合理的深度学习网络结构以实现高精度重磁数据网格化和滤波处理。建立基于自注意力机制深度学习的网格化方法,使用自注意力机制层对二维位置编码进行处理,得到融合了全局与局部信息的位置编码向量,再将位置信息与异常信息融合输出节点异常,从而降低数据的失真性。针对重磁数据噪声具有随机性、条带状的特点,首先采用卷积神经网络进行噪声分类,针对条带状噪声和随机噪声分别采用自注意力机制神经网络和卷积自编码器进行去除,可获得质量较高的基础数据。模型试验表明,深度学习的网格化结构相对常规方法更接近真实结果,所开发的滤波方法能很好地实现不同类型噪声的去除,为后续反演提供更准确的基础数据。将基于深度学习的网格化和滤波方法用于实际磁场数据的处理,获得了较好的结果,证明该方法具有较强的实用性。

Gridding and filtering method of gravity and magnetic data based on self-attention deep learning

The gridding and filtering of gravity and magnetic data directly influence the result of data processing.This paper designs a more rational deep learning model to improve the accuracy of gridding and filtering.The gridding method based on selfattention deep learning is constructed,and the selfattention mechanism layer is utilized to process the two-dimensional position embeddings.In this way,the position embeddings vector is obtained with global and local information integrated.Then the position information and anomaly information are fused to output node anomaly,thus reducing the distortion.For the random and stripe noises of gravity and magnetic data,a convolutional neural network is first employed to classify noise.The stripe noise is filtered by self-attention convolutional neural network and the random noise by convolutional autoencoder to get high-quality basic data.Model experiment shows that the gridded structure of deep learning is closer to the real result than that of the traditional method.The proposed filtering method can remove various noises,providing more accurate basic data for the following inversion.The gridding and filtering method based on deep learning applied to practical magnetic data achieves good results,proving that it has strong feasibility and practicability.