Scholarlink

This paper presents a lightweight neural network framework integrating variational mode decomposition (VMD) and CNN-BiGRU-Attention for accurate respiratory waveform reconstruction from millimeter-wave radar signals. The method first applies VMD to decompose preprocessed phase signals into intrinsic mode functions, isolating respiratory components while suppressing motion artifacts and noise. These components are then processed through a compact network architecture that combines multi-scale feature extraction, bidirectional temporal modeling, and attention-based feature weighting. Experimental results demonstrate excellent performance with a Pearson correlation of 0.952, respiratory rate MAE of 0.208 bpm, and precise waveform localization (peak timestamps error: 0.109 s, valley timestamps error: 0.210 s) while maintaining minimal computational complexity (0.054 M parameters). The proposed approach achieves an optimal accuracy-efficiency balance, enabling practical non-contact respiratory monitoring in complex environments.

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

Fine-Grained Respiratory Waveform Reconstruction with mmWave Radar by VMD and Lightweight CNN-BiGRU-AT Hybrid Network

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yubo Wang,
  • Zhongfei Ni,
  • Tianhao Guo

摘要

This paper presents a lightweight neural network framework integrating variational mode decomposition (VMD) and CNN-BiGRU-Attention for accurate respiratory waveform reconstruction from millimeter-wave radar signals. The method first applies VMD to decompose preprocessed phase signals into intrinsic mode functions, isolating respiratory components while suppressing motion artifacts and noise. These components are then processed through a compact network architecture that combines multi-scale feature extraction, bidirectional temporal modeling, and attention-based feature weighting. Experimental results demonstrate excellent performance with a Pearson correlation of 0.952, respiratory rate MAE of 0.208 bpm, and precise waveform localization (peak timestamps error: 0.109 s, valley timestamps error: 0.210 s) while maintaining minimal computational complexity (0.054 M parameters). The proposed approach achieves an optimal accuracy-efficiency balance, enabling practical non-contact respiratory monitoring in complex environments.