基于改进一维U-Net-BiLSTM的测井岩性识别方法研究

A Well-Log Lithology Identification Method Based on an Improved 1D U-Net-BiLSTM

摘要: 针对测井岩性识别中深度上下文利用不足、类别分布不均衡以及随机样本划分易造成评价结果偏高等问题，本文提出一种改进一维U-Net-BiLSTM岩性识别方法。以12口井7470条深度采样记录为研究对象，选取CAL、GR、SP、LLD、LLS、AC、DEN和PEF等常规测井曲线作为基础输入，并构造深浅电阻率比值、声波密度比值和伽马声波比值等辅助特征；在数据清洗、缺失值插补、对数变换和归一化处理基础上，采用按井划分方式构建训练集、验证集和独立测试集。所提模型以长度为64的多维测井序列窗口为输入，实现窗口内各深度采样点的逐点岩性分类；通过特征增强模块补充局部变化和统计特征，利用多尺度一维卷积提取不同深度尺度下的测井响应模式，并在瓶颈端结合BiLSTM和Transformer建模井深方向上下文关系，同时通过注意力跳跃连接和DRSN解码模块融合多层级特征、抑制弱扰动。以J393和T251井作为独立测试井的实验结果表明，本文模型在总体测试集上的Accuracy和Macro F1分别达到0.7565和0.6806，优于SVM、MLP/BP、1D-CNN、LSTM、CNN-BiLSTM、Transformer、Vanilla 1D U-Net和Attention-TCN等对比模型。在分井测试结果中，本文模型在J393井上的Accuracy和Macro F1分别为0.7838和0.7646，在T251井上分别为0.7394和0.6495，表明模型在不同测试井上均保持了较好的识别性能。岩性深度剖面对比进一步显示，所提方法能够较好保持主要岩性层段的连续性，在跨井岩性识别中具有较好的泛化能力和应用潜力。

Abstract: To address the problems of insufficient utilization of depth-wise contextual information, imbalanced lithology class distribution, and overestimated evaluation results caused by random sample splitting in well-log lithology identification, this study proposes an improved one-dimensional U-Net-BiLSTM lithology identification method. A dataset containing 7470 depth sampling records from 12 wells was used in this study. Conventional well-log curves, including CAL, GR, SP, LLD, LLS, AC, DEN, and PEF, were selected as the basic inputs, and auxiliary features such as the deep-to-shallow resistivity ratio, acoustic-density ratio, and gamma-ray acoustic ratio were constructed. After data cleaning, missing-value imputation, logarithmic transformation, and normalization, a well-wise splitting strategy was adopted to construct the training set, validation set, and independent test set. The proposed model takes a multidimensional well-log sequence window with a length of 64 as input and performs point-wise lithology classification for each depth sampling point within the window. Specifically, the feature enhancement module supplements local variation and statistical features, multi-scale one-dimensional convolutions are used to extract well-log response patterns at different depth scales, and BiLSTM and Transformer modules are combined at the bottleneck to model depth-wise contextual relationships. Meanwhile, attention-based skip connections and a DRSN decoding module are introduced to fuse multi-level features and suppress weak disturbances. Experimental results using wells J393 and T251 as independent test wells show that the proposed model achieves an overall Accuracy of 0.7565 and a Macro F1 of 0.6806 on the overall test set, outperforming comparison models including SVM, MLP/BP, 1D-CNN, LSTM, CNN-BiLSTM, Transformer, Vanilla 1D U-Net, and Attention-TCN. In the single-well test results, the proposed model achieves an Accuracy and Macro F1 of 0.7838 and 0.7646 on well J393, and 0.7394 and 0.6495 on well T251, respectively, indicating that the model maintains good recognition performance across different test wells. Lithology depth-profile comparisons further demonstrate that the proposed method can better preserve the continuity of major lithological intervals and has good generalization ability and application potential for cross-well lithology identification.