Scholarlink

<p>Precise identification of proppant particles in drill cuttings obtained from cored wells at pilot hydraulic fracturing test sites is essential for characterising fracture propagation behaviour and the factors governing fracture geometry within reservoirs. To overcome the drawbacks of manual identification, which include high labour demand, dependence on specialist expertise and susceptibility to subjective bias, we developed an enhanced ResNet-based model for recognising proppant particles. The network integrates a multi-scale linear deformable convolution (MDS_LDconv), an Adaptive Star operation block and a Mamba-inspired linear attention module (MLLA_G); moreover, the conventional fully connected output layer is replaced by a KAN layer, which significantly improves accuracy and computational efficiency when processing complex cutting images. Extensive experiments show that the improved model attains a test-set accuracy of 98.62%, representing an increment of 2.16% points over the standard ResNet-50. Field trials indicate that the machine-identified results agree with manual interpretation by 96.31%, while the inference time is only 4% of that required for human analysis, thereby greatly improving efficiency and reducing manpower expenditure.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Core-based recognition of well proppant particles using an enhanced ResNet model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月2日

查看全文
  • Shitan Yin,
  • Erlong Yang,
  • Xianjun Wang,
  • Chi Dong

摘要

Precise identification of proppant particles in drill cuttings obtained from cored wells at pilot hydraulic fracturing test sites is essential for characterising fracture propagation behaviour and the factors governing fracture geometry within reservoirs. To overcome the drawbacks of manual identification, which include high labour demand, dependence on specialist expertise and susceptibility to subjective bias, we developed an enhanced ResNet-based model for recognising proppant particles. The network integrates a multi-scale linear deformable convolution (MDS_LDconv), an Adaptive Star operation block and a Mamba-inspired linear attention module (MLLA_G); moreover, the conventional fully connected output layer is replaced by a KAN layer, which significantly improves accuracy and computational efficiency when processing complex cutting images. Extensive experiments show that the improved model attains a test-set accuracy of 98.62%, representing an increment of 2.16% points over the standard ResNet-50. Field trials indicate that the machine-identified results agree with manual interpretation by 96.31%, while the inference time is only 4% of that required for human analysis, thereby greatly improving efficiency and reducing manpower expenditure.