Applications of Non-Orthogonal Waveforms and Artificial Neural Networks in Wireless Vehicular Communications

Abstract

We live in an ever increasing world of connectivity. The need for highly robust, highly efficient wireless communication has never been greater. As we seek to squeeze better and better performance from our systems, we must remember; even though our computing devices are increasing in power and efficiency, our wireless spectrum remains limited. Recently there has been an increasing trend towards the implementation of machine learning based systems in wireless communications. By taking advantage of a neural networks powerful non-linear computational capability, communication systems have been shown to achieve reliable error free transmission over even the most dispersive of channels. Furthermore, in an attempt to make better use of the available spectrum, more spectrally efficient physical layer waveforms are gathering attention that trade increased interference for lower bandwidth requirements. In this thesis, the performance of neural networks that utilise spectrally efficient waveforms within harsh transmission environments are assessed. Firstly, we investigate and generate a novel neural network for use within a standards compliant vehicular network for vehicle-to-vehicle communication, and assess its performance practically in several of the harshest recorded empirical channel models using a hardware-in-the-loop testing methodology. The results demonstrate the strength of the proposed receiver, achieving a bit-error rate below 10−3 at a signal-to-noise ratio (SNR) of 6dB. Secondly, this is then further extended to utilise spectrally efficient frequency division multiplexing (SEFDM), where we note a break away from the 802.11p vehicular communication standard in exchange for a more efficient use of the available spectrum that can then be utilised to service more users or achieve a higher data throughput. It is demonstrated that the proposed neural network system is able to act as a joint channel equaliser and symbol receiver with bandwidth compression of up to 60% when compared to orthogonal frequency division multiplexing (OFDM). The effect of overfitting to the training environment is also tested, and the proposed system is shown to generalise well to unseen vehicular environments with no notable impact on the bit-error rate performance. Thirdly, methods for generating inputs and outputs of neural networks from complex constellation points are investigated, and it is reasoned that creating ‘split complex’ neural networks should not be preferred over ‘contatenated complex’ neural networks in most settings. A new and novel loss function, namely error vector magnitude (EVM) loss, is then created for the purposes of training neural networks in a communications setting that tightly couples the objective function of a neural network during training to the performance metrics of transmission when deployed practically. This loss function is used to train neural networks in complex environments and is then compared to popular methods from the literature where it is demonstrated that EVM loss translates better into practical applications. It achieved the lowest EVM error, thus bit-error rate, across all experiments by a margin of 3dB when compared to its closest achieving alternative. The results continue and show how in the experiment EVM loss was able to improve spectral efficiency by 67% over the baseline without affecting performance. Finally, neural networks combined with the new EVM loss function are further tested in wider communication settings such as visible light communication (VLC) to validate the efficacy and flexibility of the proposed system. The results show that neural networks are capable of overcoming significant challenges in wireless environments, and when paired with efficient physical layer waveforms like SEFDM and an appropriate loss function such as EVM loss are able to make good use of a congested spectrum. The authors demonstrated for the first time in practical experimentation with SEFDM that spectral efficiency gains of up to 50% are achievable, and that previous SEFDM limitations from the literature with regards to number of subcarriers and size of the transmit constellation are alleviated via the use of neural networks