基于标签精炼方法的地震相深度学习预测  

Seismic facies prediction using deep learning based on label refinery

作　　者：赵军才 ZHAO Juncai(Huabei Branch of Sinopec Geophysical Corporation,Zhengzhou 450016,China)

机构地区：[1]中石化石油工程地球物理有限公司华北分公司,河南郑州450016

出　　处：《石油物探》2023年第3期431-441,共11页Geophysical Prospecting For Petroleum

基　　金：中石化石油工程科技项目(SG-22-47K)资助。

摘　　要：在监督学习的三要素(数据、标签和模型)中,对标签的结构及地球物理意义的研究较少受到关注。当训练地震相分类这种多类分类模型时,通常将分类值转换为独热编码。由于手动解释的标签含有噪声,因此信息单一的独热编码在用于训练时很容易引起深度学习模型的过拟合。为此,在分析现有深度学习地震相分类存在的问题的基础上,引入了一种基于标签精炼的地震相标签动态生成方法,可以在不修改深度学习模型的基础上,提高预测准确率。首先制作包含空间信息的“先验标签”,替代独热标签送入深度学习模型进行训练;然后将前一次模型的输出结果作为下次训练的数据标签,依次迭代,不断更新该深度学习模型的网络权重与输出结果,从而得到具有更多信息量、更符合实际概率分布的地震相标签。F3工区与Parihaka工区数据的应用结果表明,该方法可以在不改变模型结构等其它超参数的条件下提高地震相预测的准确率。In supervised learning systems,the three main components are data,labels,and models.However,labels and their properties have received little attention.Categorical values are often converted into one-hot encoding when training a multiclass classification model like facies classification.However,this can lead to overfitting and overconfidence because of noisy labels and the geological sense.In this study,we first observed common shortcomings in facies classification labels and introduced a label refinery scheme.We created a prior label and then iterated through several successions of training a new model using the previously trained model as a label refiner.The subsequent network models were trained using more accurate labels.Refining the labels during training enabled us to obtain soft,smooth,dynamic,and informative labels that can reduce overfitting and overconfidence in models by using dynamically generated labels.By applying this method to the F3 and Parihaka datasets,we observed improvements in validation accuracy at successive stages of refinement.

关 键 词：深度学习 地震相预测 先验信息 标签精炼 监督学习 泛化性 

分 类 号：P631[天文地球—地质矿产勘探]