Raja Sattiraju

• The first four generations of mobile cellular networks were primarily utilized for personal communications and focused solely on increasing system capacity and user data rates. The fifth generation (5G) mobile communications standard revolutionized the cellular connectivity by adding support for new use cases such as Enhanced Mobile Broadband (eMBB), Ultra-Reliable Low-Latency Communications (URLLC) and massive Machine Type Communications (mMTC). This enabled the foray of cellular communication technologies into new verticals such as Vehicular-to-Everything (V2X) communication, Internet of Things (IoT), industrial communications etc. V2X allows vehicles to directly communicate with each other, roadside infrastructure, and other road users to deliver an array of services. It forms a compelling usecase for 5G due to its potential to provide benefits such as road safety, traffic efficiency, smart mobility, environmental sustainability, and driver convenience. Currently, two Radio Access Technologies (RATs) exist for V2X communication: Intelligent Transportation Systems (ITS)-G5, based on the IEEE 802.11 standard that has been well over 20 years in research and development, and Cellular Vehicle-to-Everything (C-V2X), based on the newer but more established and global 3rd Generation Partnership Project (3GPP) standard. Both technologies have their own pros and cons, with ITS-G5 offering ease of deployment due to its lightweight protocol and complete decentralized operation, whereas C-V2X offers better radio performance, tighter integration with cellular networks etc. Before deploying any RAT, it is important to extensively evaluate it by means of analytical, empirical and Monte-Carlo methods. This forms the core of the first part of this thesis, where a detailed link and system level performance comparison has been carried out for ITS-G5 and C-V2X. A new link-level simulation framework, namely pycv2x, has been developed from scratch for evaluating C-V2X, whereas for ITS-G5 the open-source implementation from ublox was used. An extensive library of digital signal processing blocks for coding, modulation, channel estimation & equalization, and frequency and timing offset correction was developed for C-V2X. The developed simulation framework is tested with the 3GPP reference channel models, and the simulation results match the expected values from the 3GPP Rel.14 specification. Once the link-level simulations are ready for both C-V2X and ITS-G5, the next step is to run Monte-Carlo simulations for a wide variety of channel models. The simulations considered varying channel models, ranging from simple Additive White Gaussian Noise (AWGN) models to multi-path fading models with varying delay profiles. After an extensive literature survey, a total of 8 different channel models (from the International Telecommunication Union (ITU) and Dedicated Short Range Communications (DSRC)) specifically designed for vehicular scenarios were selected and the simulations were carried out. The results show that, in a single transmission scheme, C-V2X outperforms ITS-G5 in almost all of the considered channel models, with some exceptions for 16-Quadrature Amplitude Modulation (QAM) and higher coding schemes. With one blind retransmission enabled, C-V2X exhibits a gain of at least 6 dB and in some cases reaches as high as 10 dB over ITS-G5. Link-level simulations capture the Physical Layer (PHY) performance with a single link only. In order to understand the system capacity, a system level analysis is necessary that considers the upper Media Access Control (MAC) layer schemes. The thresholds for calculating packet errors at the system level are derived from the link level simulation by means of link-system level mapping curves (SNR - BLER curves). In this regard, a realistic system level simulation framework using real-world maps and traffic was developed to evaluate the MAC schemes of ITS-G5 and C-V2X — namely Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) and Semi-Persistent Subchannel Selection (SPSS) respectively. The simulation approach is based on bidirectionally coupling the vehicular traffic simulator (Simulation of Urban Mobility (SUMO)) and the network simulator. The analysis was carried out for European Telecommunications Standards Institute (ETSI) defined scenarios such as highway and Manhattan grid. As an additional novelty, complex scenarios using real-world maps and traffic for the cities of Kaiserslautern and Merzig were also considered for simulation, capturing the traffic characteristics of both urban and rural areas. Overall, it can be seen that C-V2X has a higher range than ITS-G5 in line with the specification. This is reflected in the Packet Error Rate (PER), where ITS-G5 maxes out before C-V2X. In terms of distance, this translates to a range gain of almost 100 m for Quadrature Phase Shift Keying (QPSK) coding schemes and almost 200 m for 16-QAM schemes. C-V2X also makes more efficient use of the spectrum compared to ITS-G5, which can be seen in the form of higher Channel Busy Ratio (CBR). The retransmission gain with C-V2X is significant at higher Modulation & Coding Scheme (MCS) schemes, lower vehicular densities and higher speeds. Machine Learning (ML) / Artificial Intelligence (AI) is a topic that has generated huge interest in both academia and industry, due to the availability of data and the processing power necessary to analyze it. Consequently, it is being widely evaluated and used in almost every engineering domain. The wireless community has also started to embrace ML over the last 5 years, and many works have been carried out to assess its applicability across the entire protocol stack. This forms the core of the second part of this thesis, where different ML strategies were evaluated for signal processing operations such as channel coding, channel estimation, reliability prediction, and obstacle detection. Autoencoders are one of the first applications of ML conceived at the PHY layer, where a well-trained ML model can mimic the behavior of any signal processing block (eventually the entire PHY pipeline). The idea is that if we design a moderately complex ML model and train it with the input and output samples of any given signal processing block (such as coding/modulation), the autoencoder can learn the inherent patterns in the data and, after enough iterations, will start generating the same outputs as the considered signal processing block. In this regard, two ML models were developed for turbo decoding and channel estimation. The first is based on a Recurrent Neural Network (RNN) architecture, which is well suited to understanding time-series patterns and can therefore decode the inherent sequence dependency of turbo codes. The second is based on a Convolutional Neural Network (CNN) architecture, which is good at understanding spatial dependencies and is therefore a better candidate for channel estimation, which involves averaging operations across frequency subcarriers. The models were trained and compared with legacy signal processing blocks, and the results show that the ML-based models perform on par with or sometimes even outperform their legacy counterparts. CNN-based channel estimation in particular is seen to improve the system performance, especially in high-speed scenarios where there is a high Doppler. Proactive link adaptation and management is another key area in wireless communication where ML-based solutions can bring a significant performance gain and add value to the overall RF chain. In this regard, a novel reliability prediction scheme based on Long Short-Term Memory (LSTM) networks is proposed, which predicts the Signal to Interference plus Noise Ratio (SINR) for subsequent transmissions based on previous values obtained after channel equalization. This knowledge can help the transmitter adjust the MCS scheme before a Channel Quality Indicator (CQI) feedback from the receiver arrives. By combining the received CQI with the predicted value, the model can be tuned dynamically to make accurate predictions and subsequently optimize link adaptation. Another novel method for detecting the presence/absence of obstacles with ML algorithms on raw Ultra-Wide Band (UWB) waveforms is also investigated in this thesis. Raw UWB waveforms were collected in indoor/outdoor scenarios and labeled appropriately: 1 denotes the presence of an obstacle (human) and 0 denotes no obstacle. A suite of supervised ML models was trained on the training dataset and used to predict on the test set. The results show that just by using raw waveforms (without any need for further filtering), supervised ML-based methods can detect obstacles with accuracies close to 95%.