基于卷积神经网络算法的自动地层对比实验

An experiment in automatic stratigraphic correlation using convolutional neural networks

深度学习善于从原始数据输入中挖掘其内在的抽象特征,十余年来,其在语音识别、语义分析、图像分析等领域取得了巨大成功,也大大推动了人工智能的发展。本文基于深度学习中广泛应用的卷积神经网络算法,以大庆油田某区块密井网数据为对象,开展自动地层对比试验。实验中,随机选取部分井作为训练样本,对另一部分井分层进行预测,并与原始分层数据比对进行误差分析。按照训练样本的井数据比例65%、40%、20%和10%,将实验分为4组,每组实验包括油层组、砂层组和小层级3个相互独立的实验。12个实验结果表明:训练量越大,地层级别越高(厚度越厚),自动对比效果越好;20%的训练量就可以较可靠地进行砂组及以上级别地层单元(厚度不小于10 m)的自动对比。该实验表明卷积神经网络算法能有效应用于依据测井曲线进行油藏规模地层自动对比,具有良好的发展前景。

Deep learning is good at extracting the inherent abstract features from input data. It has achieved great success in speech recognition, semantic analysis, image analysis and other fields in the past ten years, which has greatly promoted the development of artificial intelligence. Based on the convolutional neural networks algorithm widely used in deep learning, this paper carries out well auto-correlation experiments which take a block of Daqing Oilfield as the object. In the experiments, some wells were randomly selected as training samples and the other wells were used as tested samples to predict the welltops. The predicted welltops were compared with the original welltops for error analysis. The experiments were divided into 4 groups according to the proportion of training well data, which was 65%, 40%, 20%, and 10% respectively. Each group of experiments consisted of three independent experiments, including oil layer group, sand group, and single layers. The 12 experiment results show that the more training data and the higher stratigraphic unit(or the larger thickness) can get, the better the well auto-correlation result, and the 20% training data can reliably perform the well auto-correlation of sand group and above stratigraphic units(thickness is no less than 10 m). It also indicates that the convolutional neural networks algorithm can be effectively applied to reservoir-scale well auto-correlation based on well logs and has a promising future.