Machine Learning Based Spatial Modulation for IoT Networks

  • Selina Shrestha

Student thesis: Master's Thesis


Spatial Modulation (SM) is an innovative modulation technique that transmits the information bits in 2 streams: the signal transmitted (data bits), and the transmit antenna selected (antenna selection bits). Thus, it overcomes challenges of traditional Multi-Input-Multi-Output (MIMO) schemes such as high complexity and inter-channel interference while maintaining high data rates. However, the SM receiver relies on the propagation conditions of the channel to detect the transmitting antenna. When the channel does not provide a rich scattering environment, which is a likely scenario in the context of Internet of Things (IoT), the signals from different antennas appear the same at the receiver, resulting in a high antenna correlation and poor detection. Conventional techniques to combat this problem result in added latency and computational complexity. Machine learning (ML) based low complexity enhancements either do not evaluate high antenna correlation scenarios or are limited to non-fading visible light communication systems. Moreover, existing solutions either apply ML at the transmitter or at the receiver and have not exploited the benefits of end-to-end optimization by employing machine learning at both sides and training it jointly to reach a global optimum. In this work, autoencoder (AE) based frameworks for SM under a quasi-static fading channel are proposed as low complexity solutions to overcome the problem of performance degradation in high antenna correlation scenarios of IoT networks. The transmitter and receiver are replaced by neural network based encoders and decoders that are jointly trained to reach a global optimum. Three autoencoder frameworks: Simple AE, AE with Antenna Signature, and Control bit encoded AE are proposed, each for an open-loop and a closed-loop (Channel State Information feedback at transmitter) system, among which the simple AE is a direct autoencoder implementation of SM and the other two frameworks are enhanced architectures to improve the system's robustness to high antenna correlation. While the AE with antenna signature adds a Phase Shift Keying (PSK) based signature of the selected antenna to the modulated signal generated by the learned encoder, the control bit encoded AE allows the encoder to learn the most optimum encoding of both data bits as well as antenna selection information into the modulated signal. The proposed architectures are evaluated and compared with a baseline SM scheme based on the Block Error Rate (BLER), number of bits transmitted and the power efficiency at various levels of antenna correlation. It is observed that the control bit encoded AE results in the least BLER followed by the AE with signature and then the simple AE and baseline scheme. While the simple AE and baseline SM scheme's performances degrade as the antenna correlation grows, the control bit encoded AE and AE with antenna signature are found to perform even better, owing to the fact that they encode the antenna information into the transmitted signal itself and therefore do not rely on the channel variations to detect it. However, the improved power efficiencies and robustness to high antenna correlation in the two enhanced autoencoders come with a cost of greater bandwidth requirement as they transmit both the data bits as well as the antenna selection bits through the channel while the simple AE and baseline SM scheme only transmit the data bits. Additionally, under a closed-loop system, the autoencoders also learn to perform some channel coding to overcome the effects of the fading channel without sacrificing spatial information needed to detect the transmit antenna.
Date of AwardJul 2022
Original languageAmerican English


  • Spatial Modulation
  • Machine Learning
  • Autoencoder
  • Antenna Correlation
  • Wireless Communication.

Cite this