UWB Respiratory Pattern Recognition Method Based on RTF-CGMambaNet

Respiration is a vital physiological process, and respiratory dysfunction has become a significant public health concern. To address the challenge of accurate recognition of complex respiratory disorder patterns using ultra-wideband (UWB) radar, this study proposes a Radar Time–Frequency Cross-Gated Mamba Network (RTF-CGMambaNet) for multi-class respiration classification, including normal breathing, bradypnea, tachypnea, Cheyne–Stokes respiration, Biot respiration, and Kussmaul respiration. First, an Adaptive Weighted Grey Wolf Optimization-based Variational Mode Decomposition (AWGWO-VMD) is employed to decompose and reconstruct radar echo signals, extracting a high signal-to-noise ratio one-dimensional respiratory component. Subsequently, a dual-branch time–frequency feature learning framework is constructed. The temporal branch integrates multi-scale convolution, Temporal Convolutional Networks (TCN), and bidirectional Mamba modules to capture both local and long-range temporal dependencies. The frequency branch utilizes Continuous Wavelet Transform (CWT) to generate time–frequency representations, combined with band embedding and Spectral Mamba to extract spectral structural features. Furthermore, a time–frequency interactive gated fusion module is designed to adaptively enhance and select cross-modal features. Finally, global Mamba modeling and attention pooling are adopted for classification. Experimental results on a self-built UWB radar dataset demonstrate that the proposed method achieves an accuracy of 97.32% and an F1-score of 97.35%, outperforming several state-of-the-art approaches and confirming its effectiveness and robustness.