Scholarlink

<p>Magnetotelluric (MT), as a natural-source electromagnetic exploration method, is highly susceptible to culture noise interferences, which can compromise the accuracy of transfer functions and limit the method’s application. This study proposes a novel noise suppression algorithm designed to address these challenges. The method combines Long Short-Term Memory (LSTM) networks with a Multi-Head Attention (MHA) mechanism to identify noise in MT data and employs a residual convolutional neural network (ResNet) to extract it, thereby achieving effective noise suppression. This approach is named LMR (LSTM-MHA &amp; ResNet Model for Noise Suppression). Leveraging its gating mechanism, LSTM effectively captures long- and short-term dependencies in time series, allowing precise noise localization. Multi-head attention mechanism dynamically adjusts the model’s focus on different noise types, greatly enhancing its robustness in multi-noise environments. ResNet utilizes convolutional operations and residual blocks to enable the model to extract features at different scales, capturing diverse noise types. When encountering noise, it can more robustly extract the corresponding noise while ignoring or mitigating the impact of non-noise. Through extensive testing on synthetic and field data, the method demonstrates significant advantages in handling various types of noise, such as pulse wave, triangular wave, and square wave noise. The proposed method significantly improves data quality and obtained smoother and more reliable response results.</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

Identify and Suppress Magnetotelluric Data using Long Short-Term Memory, Multi-Head Attention and ResNet

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Yi Tang,
  • Tianjun Cheng,
  • Lei Zhou,
  • Xingbing Xie,
  • Yurong Mao,
  • Liangjun Yan

摘要

Magnetotelluric (MT), as a natural-source electromagnetic exploration method, is highly susceptible to culture noise interferences, which can compromise the accuracy of transfer functions and limit the method’s application. This study proposes a novel noise suppression algorithm designed to address these challenges. The method combines Long Short-Term Memory (LSTM) networks with a Multi-Head Attention (MHA) mechanism to identify noise in MT data and employs a residual convolutional neural network (ResNet) to extract it, thereby achieving effective noise suppression. This approach is named LMR (LSTM-MHA & ResNet Model for Noise Suppression). Leveraging its gating mechanism, LSTM effectively captures long- and short-term dependencies in time series, allowing precise noise localization. Multi-head attention mechanism dynamically adjusts the model’s focus on different noise types, greatly enhancing its robustness in multi-noise environments. ResNet utilizes convolutional operations and residual blocks to enable the model to extract features at different scales, capturing diverse noise types. When encountering noise, it can more robustly extract the corresponding noise while ignoring or mitigating the impact of non-noise. Through extensive testing on synthetic and field data, the method demonstrates significant advantages in handling various types of noise, such as pulse wave, triangular wave, and square wave noise. The proposed method significantly improves data quality and obtained smoother and more reliable response results.