Scholarlink

<p>Geophysical inversion translates seismic data into underground physical parameters, presenting a nonlinear regression challenge. However, owing to geological complexities and noise interference, density inversion often struggles to accurately represent underground physical properties. Issues such as insufficient extraction of data features, poor mapping relationships, and strong dependence on the initial model further complicate the process. To solve these problems, a joint dense residual and bidirectional long short-term memory network (Bi-LSTM) inversion method is proposed. The Bi-LSTM module enhances data interpretation by leveraging forward and backward hidden states, dynamically adjusting attention to the data, and improving the network’s ability to learn information. At the same time, the dense residual structure integrates different levels of feature information, promotes the reuse and fusion of information features, reduces gradient disappearance during backpropagation, and enables the network to efficiently extract meaningful features during inversion, thereby improving prediction accuracy and continuity. Under consistent network parameters, comparisons were made with fully convolutional networks and fully convolutional residual networks, and network noise resistance experiments were conducted. Experimental results demonstrate that the proposed method significantly improves prediction accuracy and is highly stable.</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

Density inversion method combining dense residuals and Bi-LSTM structure

  • 期刊论文

  • 发表日期：

    2025年3月12日

查看全文
  • Wei-jia Zong,
  • Xing-jian Wang,
  • Rong Qi,
  • Yang Liu,
  • Zhong-yue Luo,
  • Zhan-bo Liu

摘要

Geophysical inversion translates seismic data into underground physical parameters, presenting a nonlinear regression challenge. However, owing to geological complexities and noise interference, density inversion often struggles to accurately represent underground physical properties. Issues such as insufficient extraction of data features, poor mapping relationships, and strong dependence on the initial model further complicate the process. To solve these problems, a joint dense residual and bidirectional long short-term memory network (Bi-LSTM) inversion method is proposed. The Bi-LSTM module enhances data interpretation by leveraging forward and backward hidden states, dynamically adjusting attention to the data, and improving the network’s ability to learn information. At the same time, the dense residual structure integrates different levels of feature information, promotes the reuse and fusion of information features, reduces gradient disappearance during backpropagation, and enables the network to efficiently extract meaningful features during inversion, thereby improving prediction accuracy and continuity. Under consistent network parameters, comparisons were made with fully convolutional networks and fully convolutional residual networks, and network noise resistance experiments were conducted. Experimental results demonstrate that the proposed method significantly improves prediction accuracy and is highly stable.