基于深度学习的位场边界识别方法

边界识别是位场数据处理中极为重要的一种技术, 现有的边界识别方法属于无监督式机器运算, 其识别精度与地质体的空间分布存在很大关系, 尤其是对深部复杂异常体的识别存在边界模糊的特点.为了进一步提高边界识别的精度, 受深度学习卓越非线性映射能力和监督式学习优点的启发, 本文提出了基于深度学习的位场边界识别方法, 深度学习网络结构是一种融合了多尺度特征和全局注意力机制的密集跳跃连接网络(PFD-Net).该网络结构首先以改进的U-net为骨干网络获取位场边界特征信息, 然后在嵌套的标准卷积模块之间进行密集跳跃连接来缩减编码阶段到解码阶段的语义鸿沟, 以及减少训练阶段梯度消失等问题, 随后再采用全局注意力机制模块将多尺度的高低层特征信息进行融合, 以此进一步加强边界的全局及细节定位.模型试验表明, PFD-Net网络能够准确识别出异常体的边界信息, 且对于含噪声数据, 其预测结果的质量不会降低, 该网络表现出较强的泛化性和鲁棒性.最后将本文方法应用于藏东南某铁路隧道西段的航空磁测数据, 取得了良好的边界识别结果并能够获得更多的构造信息.

Deep learning for potential field edge detection

Edge detection is a fundamental technique in the potential field data processing. The current methodology for edge detection belongs to unsupervised machine operation, whose accuracy largely depends on the spatial distribution of geological bodies, especially for deep complex geological anomaly. In order to further improve the accuracy of edge detection, a dense skip connection network (PFD-Net) with multiple scale features and a global attention mechanism is proposed for potential field edge detection, which is inspired by the excellent nonlinear mapping ability of deep learning and the advantages of supervised learning. First, the improved U-net is employed as the backbone network to obtain the featured information of the potential field edge, followed by the dense skip connection that is applied among the nested standard convolution modules, reducing the semantic gap from the encoding stage to the decoding stage and preventing phenomenon such as gradient disappearance during training. The global attention mechanism module is then applied to integrate multiple scale high- and low-level feature information to further strengthen the global and detailed position of the edge. In this paper, a fast forward algorithm of potential field anomaly spatial domain based on grid point geometrical framework is proposed, thus constructing a sufficient and diverse sample data set. The network structure of PFD-Net is built, where the input layer is the potential field anomaly, and the output layer is the projection edge of the geological anomaly body on the horizontal plane, thereby allowing the training and parameter optimization to perform based on PFD-Net, as well as the edge detection. As shown in the tests of the four sets of models, the edge information of anomalies can be accurately distinguished via PFD-Net while the quality of prediction results for the noise-contaminated data is not influenced, demonstrating the strong generalization and robustness of the network. Finally, the method in this paper is applied to the aerial magnetic survey data of the western section of a railway tunnel in southeastern Tibet, and a good edge detection result and more interpretable information are obtained.