Neural network based decision fusion for abnormality detection via molecular communications

Abnormality detection is one of the most highly anticipated application areas of Molecular Communication (MC) based nanonetworks. This task entails sensing, detection, and reporting of abnormal changes in a fluid medium that may characterize a disease or disorder using a network of collaborating nanoscale sensors. Existing strategies for such distributed collaborative detection problems require a complete statistical characterization of the underlying communication channel between the sensors and the fusion centre (FC), with the assumption of perfectly-known or accurately estimated channel parameters. This assumption is usually impractical both due to mathematical intractability of the analytical channel models for MC except in a few ideal cases, and the slow and dispersive signal propagation characteristics that make the channel estimation a difficult task even in these ideal cases. This work, for the first time in the literature, proposes to employ a machine learning approach to this task and shows that this approach provides the robustness and flexibility required for practical implementation. We focus on detection based on deep learning, specifically on a feed-forward neural network and a recurrent neural network structure that learn the underlying model from data. This study shows that the proposed decision fusion strategy can perform well without any knowledge of the communication channel.