RNN based abnormality detection with nanoscale sensor networks using molecular communications

Abnormality detection is expected to become one of the most crucial tasks of molecular communications (MC) based nanoscale networks. This task involves the sensing, detection, and reporting of abnormal changes taking place in a fluid medium, which may typify a disease and disorder, by employing a network formed by collaborating nanoscale sensors. By assuming that the channel parameters are perfectly known or accurately estimated, currently available methods for the solution of the distributed collaborative detection problems require the entire statistical characterization of the communication channel between sensors and fusion centre (FC). However, apart from some ideal cases, analytical channel models for MC are usually mathematically complex or, in many cases, analytical channel models don't exist at all. Furthermore, the accurate estimation of channel parameters is a difficult problem, even in ideal cases, because of the slow and dispersive signal propagation characteristics encountered in MC channels. Therefore, this fundamental assumption, which existing methodologies are based on, may be unsuitable in practical nanoscale sensor network implementations. For the first time in the literature, this paper proposes to employ a machine learning approach in this detection task. Specifically, we propose a deep learning-based recurrent neural network structure for decision fusion, which learns from data. Our results show that this approach leads to detectors that can perform well without any knowledge of the channel model and its properties, providing robustness and flexibility to the detection task, which is not present in existing approaches.