Adaptive equalisation for fading digital communication channels

Abstract

This thesis considers the design of new adaptive equalisers for fading digital communication channels. The role of equalisation is discussed in the context of the functions of a digital radio communication system and both conventional and more recent novel equaliser designs are described. The application of recurrent neural networks to the problem of equalisation is developed from a theoretical study of a single node structure to the design of multinode structures. These neural networks are shown to cancel intersymbol interference in a manner mimicking conventional techniques and simulations demonstrate their sensitivity to symbol estimation errors. In addition the error mechanisms of conventional maximum likelihood equalisers operating on rapidly time-varying channels are investigated and highlight the problems of channel estimation using delayed and often incorrect symbol estimates. The relative sensitivity of Bayesian equalisation techniques to errors in the channel estimate is studied and demonstrates that the structure's equalisation capability is also susceptible to such errors. Applications of multiple channel estimator methods are developed, leading to reduced complexity structures which trade performance for a smaller computational load. These novel structures are shown to provide an improvement over the conventional techniques, especially for rapidly time-varying channels, by reducing the time delay in the channel estimation process. Finally, the use of confidence measures of the equaliser's symbol estimates in order to improve channel estimation is studied and isolates the critical areas in the development of the technique — the production of reliable confidence measures by the equalisers and the statistics of symbol estimation error bursts.