Demodulation with Deep Neural Networks for Time-Dependent Flow-Based Molecular Communication Channels

A major targeted application scenario of molecular communication (MC) is information transmission in biological systems and in the human body in particular. These systems, however, are strongly time-dependent due to their interaction with a changing environment, and so are the corresponding transmission channels. To address this issue for flow-based channels, a deep neural network consisting of a heterogeneous ensemble containing a convolutional neural network (CNN) block architecture and a bi-directional long short-term memory (LSTM) model is proposed to improve demodulation for a straight tube with time-varying length, a scenario oriented towards MC in the cardiovascular system. For evaluation, simulated data is generated with a receiver position oscillating in axial direction. The proposed neural network architecture achieves an accuracy of 97 % and a bit error rate (BER) of 0.0028, which demonstrates its capability of dealing with data transmission setups with varying channel parameters. If trained with data from a static receiver, however, accuracy and BER worsen to 10 % and 0.3, respectively, which demonstrates the need for demodulation techniques tailored for dynamically changing MC setups, as particularly present in biological systems.