An implementation of the redirected learning architecture for digital pre-distortion

Abstract

<p>Digital pre-distortion is a digital signal processing technique that's used to linearize the output of various systems. A common application of digital pre-distortion is to linearize microwave power amplifier circuits, because non-linear distortion can lead to inefficient performance and out of band emissions. Since its conception, several different architectures have been developed for digital pre-distortion. There are online architectures, such as the indirect learning architecture, where signal processing is done while the amplifier is running. There are also offline architectures, such as the direct learning architecture and the relatively new redirected learning architecture, where the signal processing is done using previous input and output data from the amplifier to create a pre-distorted signal. The choice of which architecture to use often comes down to a trade off between performance and complexity. However, a common problem exists between these architectures; the complexity of the pre-distortion technique is bound to the complexity of the system's architecture. Most digital pre-distortion systems in use today use Volterra series filters and their derivatives for behavioral modeling or simple look-up tables. The complexity of applying a given behavioral model to an input signal varies little between architectures, so for a given model the question becomes which architecture will yield the greatest performance. Online methods have excellent performance, though the system required to train the models is computationally complex, as the algorithms to implement them require many calculations in a short period of time; whereas offline methods do not require an expensive training system but may not perform as well. For this reason, it is often desirable to use offline methods to save on system costs and engineering time. Most offline digital pre-distortion systems use the direct learning architecture, however newer architectures may be able to outperform the direct learning architecture with a given behavioral model.In this thesis it is shown how the redirected learning architecture was used to mitigate harmonic distortion by about 30~dB more than the indirect learning architecture. The direct, indirect, and redirected learning architectures are presented, as well as various behavioral models. This is followed by an analysis of the redirected learning architecture. Finally an implementation of the redirected learning model is presented using ADS-Matlab co-simulation. The results are then discussed to show the potential of the redirected learning method.</p>