| 贾澎涛*,成宇超*,蒋永杰**,李娜*.基于改进Transformer网络的测井曲线生成方法[J].高技术通讯(中文),2026,36(3):279~288 |
| 基于改进Transformer网络的测井曲线生成方法 |
| Logging curve generation method based on improved Transformer network |
| |
| DOI：10. 3772 / j. issn. 1002 - 0470. 2026. 03. 006 |
| 中文关键词: 测井曲线生成模型； Transformer神经网络； 局部特征； 位置编码； 注意力机制 |
| 英文关键词: well log generation model, transformer neural network, local features, position coding, attention mechanism |
| 基金项目: |
| 作者 | 单位 | | 贾澎涛* | （*西安科技大学计算机科学与技术学院西安 710054）
（**陕西煤业集团黄陵建庄矿业有限公司延安 727300） | | 成宇超* | | | 蒋永杰** | | | 李娜* | |
|
| 摘要点击次数: 44 |
| 全文下载次数: 40 |
| 中文摘要: |
| 为了解决测井曲线生成模型精度低、训练时间长的问题，提出了一种基于改进Transformer神经网络的测井曲线生成模型（well log prediction Transformer, WLP-T）。首先，该方法改进了Transformer的输入嵌入模块，使得网络能够捕捉输入序列中的局部特征，提升了模型的局部空间感知能力。其次，采用了一种可学习的位置编码方案，解决了原Transformer中位置编码无法很好地捕获时序数据位置特征的问题。最后，设计了一种更加轻量高效的解码器模块，替换了Transformer原有的解码器模块，在保证模型性能的前提下极大提升了模型的训练速度。在真实测井数据上分别进行了未钻地层曲线预测实验、缺失曲线补全实验以及曲线校正实验。结果表明，与长短期记忆网络（long short-term memory, LSTM）、门控循环单元（gated recurrent unit, GRU）及原始Transformer网络模型相比，WLP-T模型取得了更好的效果，为测井曲线生成工作提供了一种新思路。 |
| 英文摘要: |
| To address issues of diminished accuracy and prolonged training periods in well logging curve generation models, a method based on an improved Transformer neural network, termed WLP-T (well log prediction Transformer), has been proposed. Firstly, the model enhances the input embedding module of Transformer, so that the network can capture the local features in the input sequence, and improves the local spatial awareness ability of the model. Secondly, we employ a learnable position coding to rectify the inadequacy of the Transformer’s position coding in capturing the temporal characteristics of time series data. Finally, a more streamlined and efficient decoder module is introduced to replace the original Transformer decoder, significantly boosting the model’s training speed while upholding performance standards. Experiments on predicting curves in un-drilled formations, completing missing curves, and correcting curves are conducted on real logging data. The results show that compared to LSTM (long short-term memory), GRU (gated recurrent unit), and the original Transformer network models, the WLP-T model achieved better results, offering a new approach to generate logging curves. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
