Efficient and Distortion-less Spectrum Multiplexer via Neural Network-based Filter Banks

TL;DR

This paper introduces a neural network-based filter bank spectrum multiplexer that significantly improves efficiency and reduces distortion in IoT signal transmission, outperforming traditional methods in various conditions.

Contribution

It presents a novel neural network-integrated filter bank design for spectrum multiplexing, achieving high efficiency and minimal distortion in IoT networks.

Findings

Achieves -39dB normalized mean squared error in distortion

Up to 35 times efficiency gain over conventional methods

High packet reception ratio up to 98% in field tests

Abstract

Spectrum multiplexer enables simultaneous transmission of multiple narrow-band IoT signals through gateway devices, thereby enhancing overall spectrum utilization. We propose a novel solution based on filter banks that offer increased efficiency and minimal distortion compared with conventional methods. We follow a model-driven approach to integrate the neural networks into the filter bank design by interpreting the neural network models as filter banks. The proposed NN-based filter banks can leverage advanced learning capabilities to achieve distortionless multiplexing and harness hardware acceleration for high efficiency. Then, we evaluate the performance of the spectrum multiplexer implemented by NN-based filter banks for various types of signals and environmental conditions. The results show that it can achieve a low distortion level down to $-39$dB normalized mean squared error. Furthermore, it achieves up to $35$ times execution efficiency gain and $10$dB SNR gain compared with the conventional methods. The field applications show that it can handle both the heterogeneous and homogeneous IoT networks, resulting in high packet reception ratio at the standard receivers up to $98\%$.