CW Radar-based Non-Contact Respiration Monitoring using Complex-Valued Neural Networks

Machine learning-based respiration monitoring using continuous wave (CW) radar has been studied. However, most of the models are real-valued (RV), which only takes the magnitude of signals as input. Complex-valued (CV) signals can be formulated from radar baseband signals. In this paper, we compare CV-based and equivalent RV-based neural networks (NN) models to reconstruct the respiration signal with the 61 GHz CW radar. We demonstrate two model structures, a fully connected NN (FCNN) and a combination of the convolutional network and bidirectional gated recurrent units (ConVGRU), in CV-based and equivalent RV-based models, respectively. Ten hours of measurements from a clinical study are used to validate the performance. ConVGRU surpasses FCNN in both RV and CV. CV-based ConVGRU outperforms the RV-based ConVGRU with a smaller average breathing rate error of 0.55 breaths per minute (bpm), with more than 93% of the measurements achieving an error of less than 2 bpm.

Subjects

complex-valued neural networks | CW radar | machine learning | respiration monitoring

Options

CW Radar-based Non-Contact Respiration Monitoring using Complex-Valued Neural Networks