Adaptive Predistortion With Direct Learning Based on Piecewise Linear Approximation of Amplifier Nonlinearity

We propose an efficient Wiener model for a power amplifier (PA) and develop a direct learning predistorter (PD) based on the model. The Wiener model is formed by a linear filter and a memoryless nonlinearity in which AM/AM and AM/PM characteristics are approximated as piecewise linear and piecewise constant functions, respectively. A two-step identification scheme, wherein the linear portion is estimated first and the nonlinear portion is then identified, is developed. The PD is modeled by a polynomial and its coefficients are directly updated using a recursive least squares (RLS) algorithm. To avoid implementing the inverse of the PA's linear portion, the cost function for the RLS algorithm is defined as the sum of differences between the output of the PA's linear portion and the inverse of the PA's nonlinear portion. The proposed direct learning scheme, which is referred to as the piecewise RLS (PWRLS) algorithm, is simpler to implement, yet exhibits comparable performance, as compared with existing direct learning schemes.

I. Introduction

Digital baseband predistortion has been recognized as a cost effective technique for linearizing PAs. In this scheme, the PA input signal is distorted by a predistorter (PD) whose characteristics are the inverse of those of the amplifier. Both PAs and PDs are often modeled by using a polynomial, and PD coefficients are adaptively adjusted [1]–[10]. There are two types of learning techniques for PD parameter adaptation, direct learning [1], [2] and indirect learning methods [3]–[10]. The former adaptively identifies PA model parameters and directly updates PD coefficients based on the PD input and the PA output, while the latter employs an additional device called an adaptive post-inverse filter in the feedback path to identify inverse characteristics of the PA, and coefficients of the post-inverse filter are copied by the PD. Although the direct learning approach can perform better than indirect learning [2], the former is less popular because the PA model parameters need to be identified.