Scholarlink

<p>To enhance the Exploration success rate of the Audio Magnetotelluric (AMT) method in strong noise environments, this paper proposed an AMT signal noise identification and suppression method based on an adaptive multi-stage U-Net. The specific designs were as follows: (1) During denoising, all four electromagnetic channels of AMT were considered simultaneously to strengthen their interrelationships and improve the denoising performance of the neural network; (2) A Data-Driven Tight Frame (DDTF) was employed for preliminary denoising to construct the sample set, ensuring that the samples incorporate as much noise information from the survey points as possible; (3) A cascaded U-Net architecture was designed, adopting a two-stage framework comprising coarse denoising followed by fine denoising. By utilizing the absolute value of residual noise to drive a gating network, adaptive denoising was achieved. Finally, through ablation studies, comparative experiments on simulated and field data, as well as evaluations including visualization analysis, apparent resistivity-phase curves, and Nyquist diagrams, it is demonstrated that the proposed method can more accurately locate and suppress noise regions, significantly improving the signal-to-noise ratio of AMT signals. This study provides an interpretable and modular solution for AMT signal denoising, though it is limited by its focus on a single denoising scenario. Future work will further explore adaptive threshold optimization and realtime processing applications in multi-noise coupling environments.</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

Residual-Gated Deep U-Net: An Adaptive Noise Suppression Method for AMT Signals

  • 期刊论文

  • 发表日期：

    2026年2月21日

查看全文
  • Lu-na Liang,
  • Liang Zhang,
  • Jun Xu,
  • Hao Zhang,
  • Jing-tian Tang

摘要

To enhance the Exploration success rate of the Audio Magnetotelluric (AMT) method in strong noise environments, this paper proposed an AMT signal noise identification and suppression method based on an adaptive multi-stage U-Net. The specific designs were as follows: (1) During denoising, all four electromagnetic channels of AMT were considered simultaneously to strengthen their interrelationships and improve the denoising performance of the neural network; (2) A Data-Driven Tight Frame (DDTF) was employed for preliminary denoising to construct the sample set, ensuring that the samples incorporate as much noise information from the survey points as possible; (3) A cascaded U-Net architecture was designed, adopting a two-stage framework comprising coarse denoising followed by fine denoising. By utilizing the absolute value of residual noise to drive a gating network, adaptive denoising was achieved. Finally, through ablation studies, comparative experiments on simulated and field data, as well as evaluations including visualization analysis, apparent resistivity-phase curves, and Nyquist diagrams, it is demonstrated that the proposed method can more accurately locate and suppress noise regions, significantly improving the signal-to-noise ratio of AMT signals. This study provides an interpretable and modular solution for AMT signal denoising, though it is limited by its focus on a single denoising scenario. Future work will further explore adaptive threshold optimization and realtime processing applications in multi-noise coupling environments.