Development of digital predistorters for broadband power amplifiers in OFDM systems using the simplicial canonical piecewise linear function

Abstract

Power amplifiers (PAs) are inherently nonlinear devices. Linearity of a PA can be achieved by backing off the PA to its linear region at the expense of power efficiency loss. For signals with high envelope fluctuation such OFDM system, large backoff is required, causing significant loss in power efficiency. Thus, backoff is not a favourable solution. Digital predistorters (PDs) are widely employed for linearizing PAs that are driven to the nonlinear regions. In broadband systems where PAs exhibit memory effects, the PDs are also required to compensate the memory effects. This thesis deals with the development of digital PDs for broadband PAs in OFDM systems using the Simplicial Canonical Piecewise Linear (SCPWL) function. The SCPWL function offers a few advantages over polynomial models. It imposes a saturation after the last breakpoint, making it suitable for modelling nonlinearities of PA and PD. The breakpoints of the function can be freely placed to allow optimum fitting of a given nonlinearity. It is suitable for modeling strong nonlinearities. Analysis of the SCPWL spectra property shows that the function models infinite order of intermodulation distortion, even with small number of breakpoints. The accuracy of the model can be improved by increasing the number of breakpoints. The original real-valued SCPWL function is extended to include memory structure and complex-valued coefficients, resulting in the proposed baseband SCPWL model with memory. The model is adopted in the development of the Hammerstein-SCPWL PD and memory-SCPWL PD. Vector projection methods are developed for static SCPWL PDs identification. Adaptive algorithms employing the indirect and direct learning architectures are developed for identifying the Hammerstein-SCPWL PD and memory-SCPWL PD. By exploiting the properties of the SCPWL function, the algorithms are simplified. A modified Wiener model estimator is employed to circumvent the non-convex cost function problem of block models. This further reduces the complexity of the Hammerstein PD algorithms. The thesis also analyses the effects of measurement noise on indirect learning SCPWL filter. Due to its linear basis function, the SCPWL filter coefficients do not suffer the coefficient bias effects which are observed in polynomial models. The performance of the proposed SCPWL PDs are compared with state-of-the-art polynomial-based PDs by simulations and measurements.