| 金相臣* **,吴子锐*,石敏*,朱登明**,周军***.基于BiLSTM的地质片段层位预测方法[J].高技术通讯(中文),2021,31(6):607~614 |
| 基于BiLSTM的地质片段层位预测方法 |
| Geological segment horizon prediction method based on BiLSTM |
| |
| DOI：10.3772/j.issn.1002-0470.2021.06.005 |
| 中文关键词: 双向长短期记忆神经网络(BiLSTM)； 测井曲线； 地质分层； 层位预测 |
| 英文关键词: bidirectional long-short memory neural network (BiLSTM), logging curve, geological stratification, horizon prediction |
| 基金项目: |
| 作者 | 单位 | | 金相臣* ** | | | 吴子锐* | | | 石敏* | | | 朱登明** | | | 周军*** | |
|
| 摘要点击次数: 1691 |
| 全文下载次数: 1102 |
| 中文摘要: |
| 地质分层是指对某一个地区的地层剖面中的岩层进行划分，可用于指导相应的地质找矿工作。传统的地质分层主要依靠专家根据经验进行人工判断，然而由于地质层位类别繁多，需要消耗大量的时间和人力成本。现有的地质层位自动识别方法，由于没能考虑到测井数据的序列关系以及地质层位分布的特点，导致识别效果较差。基于此，本文提出了一种改进的双向长短期记忆神经网络（BiLSTM）的地质片段层位预测方法，可以根据测井数据自动快速地进行地质分层预测。该方法首先对测井数据进行分段处理，然后基于片段式的数据对BiLSTM网络进行相应的修改，其充分利用了地质层位片段式分布的特点，且考虑到了测井数据两个方向上的序列相关性。实验结果表明，本文方法在某油田真实井位数据集上的识别准确率达到了93%，相较于其他网络有着显著的效果提升。 |
| 英文摘要: |
| Geological stratification aims to divide the rock layers in the stratum section of a certain area,which is of great significant to the problem of geological prospecting.However, considering that there exists a variety of geological horizons, and traditional geological stratification relies heavily on subjective judgment of the expert,performing geological stratification is time-consuming and knowledge-intensive work.Existing automatic geological stratification methods fail to consider the sequence relationship of well logging data and the characteristics of geological horizon distribution,thus making classification accuracy unable to reach the state-of-the-art level. Based on the above background, an improved bidirectional long-short memory neural network (BiLSTM) for geological horizon prediction is proposed. To fully make use of the characteristics of fragmented distribution of geological horizons and the sequence correlation in two directions of well logging data, the BiLSTM network following the structure of the segmented data is modified. The proposed method on a real well data set of an oil field is evaluated, and the experimental results demonstrate that the classification accuracy has reached 93%, which has a significant improvement compared with other methods. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
