基于改进LSTM神经网络的测井曲线重构方法

Well Logging Curve Reconstruction Method Based on Improved LSTM Neural Network

地球物理测井过程中,由于仪器测量或者井眼原因等经常会造成部分测井曲线失真或缺失的情况,针对失真或缺失部分测井曲线的补全与生成问题,对测井领域知识和长短期记忆神经网络(LSTM)进行了研究,提出联合领域知识与深度学习的测井曲线重构模型(DK-LSTM)。利用测井领域知识中的地层岩性特征指数筛选数据得到高质量的训练样本,并将其作为深度学习重构测井曲线的依据;构建并训练带有领域知识约束层的长短期记忆神经网络模型;基于测井曲线间的强依赖关系在重构模型中引入注意力机制,进而生成并补全测井曲线中失真或缺失的信息。实验结果表明DK-LSTM测井曲线重构模型较标准长短期记忆神经网络和串级长短期记忆神经网络具有更准确的预测效果,为测井曲线重构提供了一种新思路。

In the process of geophysical logging,some well logging curves are often distorted or missing due to instrument measurement or borehole reasons.For the problem of completion and generation of the distorted or missing logging curve,the logging domain knowledge and LSTM(long and short term memory)are studied.A well logging curve reconstruction model(DK-LSTM)combining domain knowledge and deep learning is proposed.The high quality training samples are obtained by selecting the data of formation lithology characteristic index in logging domain knowledge,which are used as the basis for deep learning to reconstruct well logging curves.The neural network model of long and short term memory with domain knowledge constraint layer is constructed and trained.The attention mechanism is introduced into the reconstruction model based on the strong dependence between logging curves to generate and complete the distorted or missing information in well logging curves.The experiment shows that the DK-LSTM model has more accurate prediction effect than the standard LSTM and the cascade LSTM,which provides a new idea for well logging curve reconstruction.