• 应用前沿、交叉与综合 • 下一篇
针对传统勘探地震波初至拾取工作量大、抗噪性差和精度低所导致的速度反演质量影响生产安全的问题，提出一种基于U形多层感知机（U-MLP）网络的地震波初至拾取与反演方法。首先，为解决传统U形网络（U-Net）中的交叉熵损失函数在数据类别不平衡时导致的性能变差问题，设计一种基于加权交叉熵Lovász归一化指数（WLS）的损失函数；然后，在特征融合阶段引入残差连接，缩小低级特征与高级特征间的差距，还原更多细节信息；最后，为使U-MLP网络更好学习图像局部特征，为高级语义引入标记化的多层感知机（MLP）模块，同时降低了参数量和计算复杂度。实验结果表明，较U-Net网络，U-MLP网络在训练中收敛性更强，初至拾取误差率降低了20%以上，交并比值提升了2%。可见，所提网络模型不仅显著提高了初至拾取精度，而且所获初至在仿真数据和实际数据中的速度反演均达到了理想效果，具有更好的性能且适应性更强。
A method for first-arrival picking and inversion of seismic waveforms based on U-shaped Multilayer Perceptron (U-MLP) network was proposed to solve the problems of safety in production affected by the quality of inverted velocity, which can be induced by heavy workload, poor noise immunity and low precision in traditional first-arrival picking methods. Firstly, a Weighted Cross-entropy Lovász-Softmax (WLS) loss function was designed to solve the problem of poor performance of U-shaped Network (U-Net) caused by the traditional cross-entropy loss function in processing unbalanced data categories. Then, residual connections were introduced in the feature fusion stage to reduce the discrepancies between low-level and high-level features, restoring more detailed information. Finally, a Multilayer Perceptron (MLP) module was introduced for high-level features so that the network can learn local image features better, and the number of hyperparameters and computational complexity are simultaneously decreased. Experimental results show that compared with the U-Net network, the U-MLP network converges faster in training, and its error rate of first-arrival picking decreases by 20% with its value of intersection over union increasing by 2%. It can be seen that the proposed network model not only improves the accuracy of first-arrival picking significantly, but also produces first arrivals for inverting ideal velocity models in both the synthetic and the real datasets, hence demonstrating its better performance and stronger adaptability.
摘要： 针对传统勘探地震波初至拾取工作量大、抗噪性差和精度低所导致的速度反演质量影响生产安全的问题，提出一种基于U形多层感知机（U-MLP）网络的地震波初至拾取与反演方法。首先，为解决传统U形网络（U-Net）中的交叉熵损失函数在数据类别不平衡时导致的性能变差问题，设计一种基于加权交叉熵Lovász归一化指数（WLS）的损失函数；然后，在特征融合阶段引入残差连接，缩小低级特征与高级特征间的差距，还原更多细节信息；最后，为使U-MLP网络更好学习图像局部特征，为高级语义引入标记化的多层感知机（MLP）模块，同时降低了参数量和计算复杂度。实验结果表明，较U-Net网络，U-MLP网络在训练中收敛性更强，初至拾取误差率降低了20%以上，交并比值提升了2%。可见，所提网络模型不仅显著提高了初至拾取精度，而且所获初至在仿真数据和实际数据中的速度反演均达到了理想效果，具有更好的性能且适应性更强。
中图分类号:
