Silveira, D. D. (2007). Black-Box modeling of microwave amplifiers for linearization [Dissertation, Technische Universität Wien]. reposiTUm. https://resolver.obvsg.at/urn:nbn:at:at-ubtuw:1-14892
This thesis is a research about power amplifier (PA) behavioral models (BMs) structures and their estimation strategies for digital pre-distortion purposes. The work starts with a brief overview of BMs excitation signals and partitioning of data used in the modeling process. Figures of merit, tools to measure BMs quality, are analyzed.<br />An investigation of linear estimation techniques and parametrization of linear systems is also performed, showing advances in the finite impulse response filter estimation. Following, techniques for nonlinear systems estimation are described, focused on PAs. Static nonlinear models and dynamic ones are outlined, together with their linear estimation methods. A study about PA BMs estimated under different noise levels shows that models estimated with noise corrupted data achieve better general results than models estimated with noise free data. The next application was a PA pre-distorter implementation. The pre-distorter presented only reasonable results, so the need for more elaborated models was evident. Further on, advances in BMs are shown, as a Volterra series model approximation that uses linear estimation techniques, having a well organized structure. A method for improving the condition number of the least-squares Hessian is depicted, based on a resampling factor used in the regression matrix. Later on, a figure of merit was developed to analyze the modeling of nonlinear distortions of the output signal. Different models presented better performance for systems with specific memory effects (memoryless, linear memory, nonlinear memory, linear and nonlinear memory). A pre-processor for PA BMs estimation is presented. New models are introduced, composed by branches with look-up tables and pruned Volterra series approximation. Using a resampling factor in the fitting process, these models have high accuracy and operate in parallel, what is more computationally efficient than models with one branch having operating with a high number of parameters. The results have shown that this model was more accurate and capable to a better representation of a PA than previous analyzed and developed models. Simulated and/or measured data are used for modeling and validation in all cases.