A Variational Autoencoder Network for Heartbeat Signal Reconstruction and Breathing-State Classification Using 24 GHz Radar Sensors

Wireless radar sensors can remotely detect vital signs. This study investigates the wireless monitoring of physiological signals induced by respiration and heartbeat to classify breathing states. Two 24 GHz radar sensors are used. Noise, harmonics, and intermodulation products affect the measured physiological signals and can make heartbeat identification and analysis challenging. Deep-learning techniques have recently been used to denoise and classify signals in wireless communications. Variational autoencoders (VAEs) have also been used for signal and image processing. Multilayer VAEs support sequential learning by optimizing generative-model parameters and reducing reconstruction error. This study investigates the separation of measured heartbeat and respiratory signals and the reduction in noise and the effects of limited random body movements using VAEs under controlled conditions. Two radars are positioned in front of and behind a human subject to record vital signals simultaneously. The separated and denoised heartbeat signals are input to a ResNet network to classify normal and fast breathing states. The proposed method achieves 99.7% accuracy on the reported experimental dataset.