基于半监督对比学习的地震相智能识别方法研究

Intelligent recognition of seismic facies based on contrastive semi-supervised learning

近年来,基于无监督学习和有监督学习的深度神经网络在地震相智能识别中发挥了重要的作用。但在实际应用中,无监督学习因缺少先验知识的引导,而识别精度相对较低。有监督学习需依赖大量标签信息,而实际情况难以满足该要求。提出一种半监督对比学习地震相识别方法,采用无标签数据和有标签数据共同学习以优化模型的性能和学习能力,从全部数据中学习相似样本之间的特征和不相似样本的差异,尽可能缩小同类地震相的类内距离并尽可能扩大类间距离;然后,利用少量的标签学习,将相类型和学习到的特征对应起来;最终实现全区的地震相高精度识别。将该方法应用于SEAM AI地震数据和南海某工区地震数据,获得了地震相识别结果,并将其与常规有监督地震相识别方法得到的结果进行对比,发现在少量标签的情况下,半监督对比学习地震相识别方法能有效识别不同类别的地震相,提高了地震相识别精度,具有良好的应用前景。

Traditional manual interpretation of seismic facies relies heavily on the experience of interpreters,resulting in low efficiency and strong subjectivity.In recent years,deep learning techniques,represented by deep neural networks based on unsupervised learning and supervised learning,have played an important role in the intelligent identification of seismic facies.However,the problem of low accuracy of identification has vexed unsupervised learning with no guidance of prior knowledge in practical application,while supervised learning relies on a large amount of labelled information,which is often unavailable in practice.We propose a method of contrastive semi-supervised learning to optimize the model for seismic facies identification,which uses unlabelled and labelled data to learn common characteristics of similar samples and the discrepancies among dissimilar samples so as to minimize intra-class distance for similar facies and maximize inter-class distance for different facies as much as possible.The facies and learned features are correlated using a small number of labels to achieve high-precision identification of seismic facies in the region of interest.The proposed method was successfully applied to the public SEAM AI dataset and a filed dataset from the South China Sea.Compared with conventional supervised methods for seismic facies recognition,the proposed method can identify different types of seismic facies more accurately with a small number of labels.