Roger Green

Load pull correction for large antenna arrays

5G and future wireless communication systems will require disruptive technologies to be deployed to overcome the difficulties of transmitting signals in highly spectral-efficient way for sub-6GHz frequencies and reliably for mm-Wave frequencies. Massive MIMO systems offer these desired capabilities as it has been already reported in the literature showcasing unprecedented spectral-efficiency levels. This was achieved using industry based hardware for proof of concept. This project will investigate the most relevant massive MIMO scenarios in terms of digital and hybrid phased-array architectures. Secondly, the integration of the RF chain to the radiating elements will be investigated with a robust multi-physics modelling approach in order to come up with a theoretical framework of how to successfully design an active phased-array transceiver. Innovations in multi-parameters nonlinear behavioural modelling and antenna design will be targeted.

Peter James

Cheap & clever phased arrays

Phased arrays are becoming increasingly critical across a wide variety of applications ranging from 5G mmWave to satellite communications. The prospect of physically very small antennas at mmWave frequencies would make implementing electronically large antenna arrays at the terminal a possibility. However, these are accompanied by other challenges that would need to be tackled by intelligent phased array architectures such as the trade-offs of antenna array gain versus size, weight and power would dictate the phased array architectures for 5G mm-wave terminals. On the other hand, low elevation angle performance would be very critical for satellite phased array terminals.

This project will address issues such as interference mitigation, coverage, terminal power and low elevation angle performance; explore properties of metamaterials towards beamforming; develop prototypes of proposed architectures.

Sanat Nagaraju

An assessment of L-systems for media coding & decoding

Modern image & video encoding, and decoding systems tend to use pixel-based coding without much consideration of the image scene composition. Typical encoding standards like JPEG (still images) and H.264/H.265 (motion pictures) uses a transformation engine (e.g. DCT), quantisation of the residual blocks followed by entropy coding which provides a desired compression ratio for transmission over a bandwidth limited channel. The pixel-based coding is applied to the entire blocks & frames and very limited content composition of the image is exploited for compression. This project is an exploration of alternative methods of encoding based on the composition of image and further improve the amount of compression achieved. The exploration considers variations between the repetitive and scalable patterns available in most images and exploit the statistical redundancy available in the composition of images. For example, modelling and rendering of natural scenes have been studied over the last 50 years to describe models of terrain and plants. In this technique, the redundancy of plant growth and patterns occurring in nature can be represented as fractals and alphabet of symbols & collection of production rules for image reconstruction. The technique using Lindenmayer Systems is currently well suited for rendering scenes in computer graphics and animations, and considers the use of this technique for image coding. 

Chrys Paschou

Physical layer security and friendly jamming

The project deals with two problems in secure wireless communications. Firstly, to prevent eavesdroppers from listening in on secure transmission, one of the methods in physical layer security is to use jamming. In such a case, channel state information will need to be fed back from the receiver to transmitter. Using the CSI, the transmitter can transmit jamming signals (most of the time in the form of noise) in the null subspace of the receiver CSI. Alternatively, specialised jamming nodes can be used in place of the transmitter jamming. These nodes can operate in a distributed manner through game theory. Secondly, the receiver node may experience jamming attacks from malicious nodes, causing it to miss messages that may be critical.

Jonathan Thomas

Reinforcement learning for network management applications 

With increasing mobile network complexity as a result of the forthcoming introduction of 5G, it is likely that intelligent systems will be required in order to effectively facilitate network management. On current networks, this is performed by human operators, but their capacity is unlikely to be able to effectively scale or provide the agility necessary for 5G deployments, as such autonomous solutions are inherently attractive. Reinforcement Learning (RL) is a potentially highly effective method to provide solutions to problems of this nature due to its ability to develop policies which can map system states to actions based on reward acquisition. It achieves this without the need for explicit definition of the problem spaces behaviour which is often difficult to do for problems of significant scale. Deep Reinforcement Learning (DRL) which combines the powerful capabilities of Deep Learning and RL has recently achieved a number of notable successes where it has outperformed domain experts, for example in Go and a range of Atari Games. Through the formulation of network management problems in a manner which is similar to games, it is likely that the same level of expertise can be developed and a number of promising examples exist within the literature. The project aims to introduce RL with a particular focus on existing and future applications to communications networks. 

Robert Zakrzewski

Detection of unusual behaviour and actions in communication systems

Proliferation of malicious users and devices in the global network makes an automated, explainable and robust protection the paramount requirement. Methods for anomaly detection using artificial intelligence with multi-modal information processing, and the techniques for threat localization can be used to enable automated and robust countermeasures. Effectiveness of detection process is enhanced by distribution of agents and fusion of complementary information to support incident classification tasks. The distribution implies collaborative nature of the system and heterogeneity of the environment. Methods and algorithms enabling knowledge acquisition, micro-operations, co-learning and knowledge transfer are the building blocks for the integration of the global system with various classes of systems including constrained devices. Last but not least, models need validation which brings the key challenge with datasets accuracy and multi-modal information handling.