一种基于对抗神经网络的方法在钻井数据恢复中的应用

Application of a Method Based on Adversarial Neural Network in Drilling Data Recovery

张宁

ZHANG Ning

中国石化集团国际石油勘探开发有限公司，北京　100101

SINOPEC International Petroleum E & P Corporation, Beijing 100101, China

机器学习中数据缺失很普遍，导致数据缺失的因素通常有人为失误、数据处理软件的缺陷、获取数据的传感器错误等。数据缺失会导致机器学习的性能下降，因此缺失值的填补对机器学习任务变得格外重要。针对数据缺失问题，该文提出一种新颖的缺失数据填补方法，构建了一个生成对抗填补网络（简称GAIN）。GAIN主要包括生成器和判别器两个部分，其中生成器（G）用来观察真实数据的每一部分，然后根据观察的结果填补缺失数据的部分，输出一个填补后完整的向量；判别器（D）接受一个完整的向量，来判别哪一部分数据是真实的，哪一部分是被填补的。在4个UCI机器学习标准数据集和石油行业钻井液数据集间进行了实验，验证了GAIN方法的有效性，能提升机器学习任务的性能。

Data missing is a common phenomenon in machine learning, and the reasons are usually human errors, data processing software bugs, incorrect sensor readings, and so on. The performance downgrade of machine learning can be caused by data missing, and thus missing data imputation is of great importance for machine learning tasks. Aiming at this problem, a novel missing data imputation method is proposed, in which a generative adversarial imputation network (GAIN) is constructed. GAIN is mainly composed of two components, including generator and discriminator. The generator (G) is used to observe each part of the real data, then completes the missing part of the data according to the observation results, and finally outputs an imputed vector.The discriminator (D) accepts a complete vector to determine which part of the data is truly observed and which is imputed. Experimental results on four public UCI datasets and real drilling fluid datasets verify the GAIN is effective. It can improve the performance of machine learning tasks.

https://doi.org/10.3969/j.issn.1008-2336.2022.02.083