Generative Adversarial Network Enabled Sparse Signal Compression and Recovery for Internet of Medical Things

Abstract

Due to the limitation of energy supply and the requirements of high reliability in the mission-critical Internet of Medical Things (IoMT), the efficient and reliable transmission of the sensing siganl of implantable medical devices (IMDs) is still a challenge. In order to improve the spectrum efficiency and transmission reliability, in this paper, a Generative Adversarial Network-enabled Sparse Compression and Recovery (GAN-SCR) scheme is proposed by exploiting the physical knowledge of sparsity, which compressively measures the sparse IMD sensing signal in the transmitter, and recovers the sensing signal in the receiver. In the stage of sparse measurement in the proposed GAN-SCR scheme, a pre-trained measurement discriminative network (MDN) is used to conduct signal compression at the transmitter, which enhances the restricted isometry property via learning. In the stage of sparse recovery, exploiting the temporal correlation and inherent sparsity of physiological signals, a pre-trained representation generative network (RGN) is used to map the sensing signal to a low-dimensional latent vector for sparse representation learning. Subsequently, the projection from the latent vector onto the measurement vector is structured by jointly training an RGN and an MDN, by which accurate signal recovery can be implemented via online optimization. Simulation results verify that the proposed GAN-SCR scheme outperforms other state-of-art sparse reconstruction algorithms in the accuracy of sensing signal recovery.