Contactless Reconstruction of ECG and Respiration Signals With mmWave Radar Based on RSSRnet
Yingxiao Wu, Haocheng Ni, Changlin Mao, Jianping Han
Abstract
In response to the challenge of effectively rebuilding vital signals acquired from radar, we propose RSSRnet that works on the short-time Fourier transform (STFT) spectrum to separate radar-acquired chest mechanical signals and reconstruct electrocardiogram (ECG) signals and respiration (RSP) signals. RSSRnet is constructed based on an encoder–separator–decoder architecture, which combines multiple self-attention blocks to enhance the separated representation of bottleneck features. In addition, we incorporate channel attention (CA) and cross CA (CCA) blocks to refine the feature representation of the encoder and the decoder, respectively. RSSRnet is optimized using an improved loss function defined on both the time-domain waveform and STFT spectrum. The experimental results show that RSSRnet can not only separate chest mechanical signals with mixed respiratory and heartbeat components, but also effectively perform the transformation from mechanical signals to electrical signals. Compared to the ground-truth signals, the median-normalized root mean square error (NRMSE) for reconstructing RSP and ECG signals is 0.043 and 0.049, respectively. The median Pearson correlation coefficients (PCCs) for reconstructed RSP and ECG signals are 0.992 and 0.964, respectively. Furthermore, the median absolute error of the cardiac intervals between the reconstructed ECG signals and the ground-truth values is less than 8 ms.