基于循环一致性对抗网络的地震断层训练样本合成方法研究

为了获得真实的地震断层训练样本, 提出了基于循环一致性对抗网络的断层训练样本合成方法。使用随机生成的断层标签与实际断层数据作为输入, 利用无监督的对抗网络学习断层标签与断层数据之间的联系, 生成与断层标签特征相匹配的地震断层样本, 由此得到带有标签的断层训练样本集。该方法是一种获取断层训练样本集的方法, 一定程度上解决了深度学习地震断层解释缺少训练数据集的问题。对合成断层样本与真实断层进行平均主频与纹理差异的定量分析, 结果表明两者具有较高的相似性。使用合成的断层样本训练神经网络, 并将结果应用于实际数据测试并进行对比, 结果表明合成的断层训练样本具有真实可靠的特点, 所提方法可以针对不同工区生成具有目标导向性的断层, 能够灵活有效地应用于不同工区的地震断层智能识别。

In order to obtain realistic seismic fault training samples, a seismic fault training sample synthesis method based on cycle-consistent adversarial networks is proposed.This method takes randomly generated fault labels and real fault data as inputs, and employs an unsupervised adversarial network to learn the relationship between fault labels and fault data and generate seismic fault samples that match the characteristics of the fault labels, thereby obtaining a labeled fault training sample set.This new method alleviates the problem of training dataset shortage for deep learning in seismic fault interpretation.A quantitative analysis of the mean frequency and textural difference between the synthetic and real faults was performed, showing a high similarity between the two.The neural network trained with the fault samples generated by this method was applied to real data for testing and comparison.The results show that these samples are realistic and reliable.Furthermore, this method can generate targeted faults for different work areas, and be flexibly and effectively applied in intelligent seismic fault detection in work areas.