Tiny Neural Networks for ISM band demodulation
Marco Forleo, Danilo Pietro Pau, Nicolò Ivan Piazzese · 2022 IEEE 21st Mediterranean Electrotechnical Conference (MELECON) · 2022
The current industrial, scientific and medical band transceivers provide I (in-phase) Q (quadrature) samples which are generated at the output of the analog-to-digital converter or by the subsequent decimation and channel filter chain. These signals are further processed by a microcontroller, embodied in the transceiver, for different application purposes. By implementing the demodulation in a software manner, the objective is to deploy different and more accurate strategies in the transceiver. Unfortunately, the state-of-the-art solutions that consider a machine learning approach to the software defined demodulation did not provide measures of complexity nor if they are suitable for running on a low-cost low-power microcontroller. This work studied how feasible such a solution would be. Two case studies were considered: gaussian frequency-shift keying Bluetooth Low Energy and frequency-shift keying Wi-SUN demodulations. A comparison between artificial neural networks and classical algorithms was conducted. The deep learning performances proved to be comparable with those of the excellent Viterbi demodulator and they were finally validated with complexity measurements and analysis of processing time on microcontrollers to demonstrate the fulfilment with the memory requirements and the real time constraint for a demodulation with low data rate.