Blind source separation for interference cancellation in CDMA systems

Abstract

Communication is the science of "reliable" transfer of information between two parties, in the sense that the information reaches the intended party with as few errors as possible. Modern wireless systems have many interfering sources that hinder reliable communication. The performance of receivers severely deteriorates in the presence of unknown or unaccounted interference. The goal of a receiver is then to combat these sources of interference in a robust manner while trying to optimize the trade-off between gain and computational complexity.

Conventional methods mitigate these sources of interference by taking into account all available information and at times seeking additional information e.g., channel characteristics, direction of arrival, etc. This usually costs bandwidth. This thesis examines the issue of developing mitigating algorithms that utilize as little as possible or no prior information about the nature of the interference. These methods are either semi-blind, in the former case, or blind in the latter case.

Blind source separation (BSS) involves solving a source separation problem with very little prior information. A popular framework for solving the BSS problem is independent component analysis (ICA). This thesis combines techniques of ICA with conventional signal detection to cancel out unaccounted sources of interference. Combining an ICA element to standard techniques enables a robust and computationally efficient structure. This thesis proposes switching techniques based on BSS/ICA effectively to combat interference. Additionally, a structure based on a generalized framework termed as denoising source separation (DSS) is presented. In cases where more information is known about the nature of interference, it is natural to incorporate this knowledge in the separation process, so finally this thesis looks at the issue of using some prior knowledge in these techniques. In the simple case, the advantage of using priors should at least lead to faster algorithms.