Sujet de thèse

The new era of connected (autonomous) vehicles has the potential for realizing unprecedented enhanced road safety with a lower environmental impact. Hence, a significant improvement in vehicular-to-infrastructure (V2I) communication is demanded. Solutions must support very fast information exchanges, with high data rate, ultra- reliability and low latency. Massive MIMO (mMIMO) is one of the 5G key enabling technologies to achieve this performance combined with machine learning (ML) algorithms suitable for this high dynamic context. Two important factors hindering the straightforward ubiquitous deployment of mMIMO 5G technology systems are the huge energy consumption and complexity required for efficient PHY and MAC algorithms handling the radio resources in V2I dynamic and possibly crowded scenarios. However, vehicular communication based on ML have been recently proposed in literature, but to the best of our knowledge, there is no ML solution yet from a real mMIMO testbed.

Résumé

The PhD objective is to fill this gap. A V2I realistic testbed can be the new foundation for precise estimate of the performance indicators (incl. energy consumption), design of energy-aware V2I mMIMO algorithms (with focus on PHY and MAC layers), and energy-aware deployments. A large and trusted radio channel dataset will be produced from both a channel sounder and a ray-tracing suitable for vehicular mMIMO scenarios. This will permit for consistent analysis and development of beyond-state-of-the-art ML or heuristic approaches to define and optimize intelligent PHY+MAC algorithms and network topologies, based on accurate PHY layer modelling, and focusing on ITS V2I wireless communications at 5.9 GHz.

The PhD student will contribute to the following sub-objectives and will be assisted by post-docs and engineers from the advising team:
-OBJ1: Production of a rich and accurate set of data to specifically characterize the massive MIMO radio channel in suburban and urban vehicular (V2I) scenarios. Real data will be measured by a cutting-edge channel sounder with a massive antenna array at the transmission and a mobile reception. This will be complemented by extensive simulations from a ray-tracing tool that will have been adjusted and validated on relevant massive MIMO metrics e.g. covariance matrix. The cross-analysis of the measured and predicted V2I channels will permit a better identification of all contributions, incl. scattering by urban furniture and vehicles, leading to a precise characterization of their properties like space correlations.
-OBJ2: Elaboration of new energy-aware PHY and MAC algorithms to minimize the power consumption while guaranteeing the real-time responsiveness and reliability required by critical V2I communications. First, we will study the impact of PHY and MAC technologies (scheduler, power allocation, precoder) on the used downlink resources and the energy consumed by the massive MIMO system. Key technical levers for power savings will be deduced. Then, new optimized energy-aware techniques will be designed from AI-based resource allocation algorithms. The large amount of consistent radio channel data available from OBJ1 is a major asset for accurate evaluation and training of those new ML approaches.
-OBJ3: Evaluation of algorithms and topologies to find solutions that adequately reduce the consumed energy statistics. A realistic virtual V2I testbed will be developed. It will be derived from existing software, in which the power consumption model, new V2I ray-tracing and proposed ML resource allocation techniques are integrated. The simulation of large-scale scenarios with different network topologies will provide accurate V2I performance indicators, to permit the identification of suitable low-power deployment strategies. For fair evaluation, the power consumed by the backhaul layer in case of a dense small-cell topology will be accounted for.

The advising team is highly complementary on this topic. At IEMN, there is a recognized expertise on vehicular MIMO signal processing and radio channel characterization. The recent results achieved on the characterization of the environment by the means of real time features and parameters based on a MIMO sounder will be the basis for identifying the key parameters to be accounted for in the ML approaches. In particular, the expertise they have gained in collecting data in the harsh context together with their expertise in V2X radio channel in suburban environment, will be exploited to correctly characterize the outdoor scenarios and identify the key features pertinent for the communication system.

At Inria, the FUN team has expertise on wireless communication paradigms in constrained and dynamic contexts. In particular, constraints in terms of resources and energy are at the core of the activities of the team and most of the solutions developed are based on the integration of ML approaches for wireless communications.

Our thesis topic is at the heart of AI, since it aims to collect a database in order to learn and test machine learning solutions, the application of which relates to the communications of connected vehicular networks. A point of attention of the thesis will be to offer machine learning solutions that consume less energy as possible. We can consider the following breakdown for the tasks to be carried out in order to fulfill the objectives:
WP 1: Scenarios Definition → OBJ1, OBJ2, OBJ3
The considered system will be characterized by PHY and MAC massive MIMO functions, network topologies, vehicle mobility patterns (fluid or crowded), data traffic, performance metrics and ITS use case requirements. We will identify the system functions likely to impact the energy consumption in a significant way, and which will be further studied in WP 3 for optimization. We will rely on the expertise of the whole consortium for definition of the vehicular scenarios that will be measured/emulated in WP 2
WP 2: Radio Channel Data Collection → OBJ1
We will rely on the expertise of IEMN to produce and analyze real and simulated radio channel data that will feed the design of intelligent algorithms in WP 3. Real data will be collected by the MaMIMOSA radio channel sounder by IEMN, which was developed to realize complex measurement campaigns in harsh propagation scenarios, possibly with mobility. Measurements will be acquired in urban and suburban environments. The radio channel will be emulated in various scenarios, with different infrastructure topologies, antenna sizes and vehicle densities.
WP 3: Design of Energy Aware Intelligent Algorithms → OBJ2
Based on the accurate data collected in WP 2, we will design ML algorithms for managing the 5G massive MIMO resources in a reliable and efficient way for V2I dynamic and possibly crowded situations. The objective is to minimize the energy consumption while still complying with critical-service requirements: size of the antenna array, scheduling, multi-user precoding, and channel access policy are part of the problem to be considered. First, we will implement accurate energy models to account how the different parts of the system impact on the energy consumption. This work will permit the identification and characterization of the system parameters and system blocks having more impact; this will be the basis for the design of energy-aware and ML wireless communication approaches at the PHY, MAC or cross-layer, trained from available channel data.

We will share relevant measured and simulated channel data and open code with the research community for re-applicability purposes or further studies. From an exploitation point of view, we intend promoting our conclusions towards territories, regulators and industry. We already had promising contacts with the Metropole Europeenne de Lille (MEL) and i-Trans cluster that have manifested their interest for the potentiality of this topic regarding optimization of the V2I infrastructure.