一种基于深度神经网络模型的测井曲线生成方法

A well log curve generation method based on deep neural network

老井复查是测井资料解释处理过程中一项重要的工作。鉴于有些老井测井曲线种类较少,有必要构建完整的测井曲线集,用于老井的解释处理。目前主要是采用BP神经网络这类浅机器学习方法,但存在生成曲线精度低,收敛速度慢,易陷入局部最优解等问题。因此,采用深度学习思想,提出一种测井曲线生成方法,构建多个隐含层的全连接深度神经网络模型,采用批处理方式的网络训练优化算法,ReLU激活函数和Dropout正则化方法,使网络快速、平稳收敛,并消除过拟合。实际测试中,从邻井选择训练数据集,生成目标井的缺失测井曲线,相比于传统的BP神经网络,生成曲线与实际曲线的相关系数提升了近20%,均方根误差降低11%,证实了所提方法的有效性。所提方法有效丰富了老井的测井曲线类型,使老井复查工作在油田的勘探开发中发挥了更加重要的作用。

Old wells logging data re-evaluation is an important work during logging data interpretation and processing.Because of smaller types of logging curves in some old wells,it is necessary to construct complete logging dataset applied in interpretation and processing for old wells.At present,shallow machine learning methods,such as BP are used,but problems such as lower accuracy of generated logs,slower convergence speed,and local optimum exist.So,a well log curve generation method based on full-connect deep neural network with multiple hidden layers is proposed by deep learning.Network is converged fast and smoothly and over-fitting is erased by adopting optimizing algorithm with mini-batch,ReLU activating function and Dropout regularization method.In actual experiments,absent well logging curves in target well are generated by training dataset selected from neighboring wells.Compared with traditional BP neural network,correlation coefficient is improved by 20%nearly and root mean squared error is decreased by 11%,improving the effectiveness of this method which effectively enlarges types of logs in old wells and make re-evaluation work play a more important role in exploration and development of oil fields.