A CGAN Channel Model of Magnetic Induction Communications for Deep Learning End-to-end Optimization

Magnetic Induction (MI) Communications have been proposed as an effective method for cross-medium communications, However, MI communications still suffer from limited communication range, complex channel environment, and low dynamic adaptability. This paper proposes a deep leaning based method for the optimization of end-to-end MI communications. To allow the loss gradient propagation in deep learning joint optimization, a conditional generative adversarial network (CGAN) based MI communication channel model is presented in this paper. The proposed CGAN channel model is able to simulate the effect of a wireless MI channel on 20kHz Binary phase shift keying (BPSK) signals. The trained Generator in the CGAN model is able to generate simulated channel output signals with 2.89% and 9.85% deviations from the real signals. This channel model is expected to be applied in the future deep learning based MI communication system joint optimization.