基于改进Stacking方法的成像测井裂缝识别

Imaging logging fracture identification based on improved Stacking algorithm

摘要:

传统成像测井裂缝识别依赖专家经验，主观性强。深度学习方法可提取多维度特征构建裂缝识别模型，集成学习方法融合多个模型提升精度，但简单融合策略精度提升不明显，复杂融合策略易过拟合，需探索更有效融合方法提升裂缝识别准确性。针对上述问题，提出了一种基于Stacking方法改进的集成深度学习模型，将Deeplabv3+，YOLOv8，SegFormer三个模型进行融合，并设计了跳跃连接模块将原始图像骨干特征传递至元模型层，避免在模型融合时由于特征不足产生的错误拟合。经实验，该模型在构建的测井图像裂缝识别数据集上Dice系数可达89.6%，优于单一模型与少数模块构建的集成模型。将该方法用于柴达木盆地实际成像测井资料的裂缝识别，能够准确地提取出测井图像的裂缝信息,证明了该模型的有效性，为成像测井裂缝识别提供了新的思路。

Abstract:

Traditional imaging logging fracture identification relies on expert experience, which has strong subjectivity. Deep learning methods can extract multidimensional features to build fracture identification models. Integrating multiple models using ensemble learning improves accuracy, but simple fusion strategies show limited accuracy gains and complex strategies are prone to overfitting.Effective fusion methods need exploration to enhance fracture identification precision. To address this issue, a refined ensemble deep learning model based on the Stacking method is proposed.The model combines Deeplabv3+, YOLOv8, and SegFormer, incorporating a skip connection module to transfer original image backbone features to the meta-model layer, preventing errors due to insufficient features during fusion. In experiments, the model achieves a Dice coefficient of 89.6% on a constructed logging image fracture identification dataset, outperforming single and few-module ensembles. Applying this method to fracture identification in the Qaidam Basin's actual imaging logging data accurately extracts fracture information, proving the model's effectiveness and offering a new approach for imaging logging fracture identification.