RF power amplifiers behavioral modeling based on extended neural network with additional dynamic fuzzy weights

Abstract

This paper presents an extended neural network (NN)-based behavioral model of RF power amplifiers (PAs) with memory effects. PA behavioral models are traditionally constructed based on the relationship between the input and output of PA, without regard to the physical mechanisms of distortions in PA however. To more accurately predict the dynamic nonlinear behavior of PA, the novel proposed NN-based behavioral model takes into account the characteristics of the amplitude modulation to amplitude modulation and amplitude modulation to phase modulation distortions that are incorporated into the behavioral model by means of additional dynamic fuzzy weights. In addition, the NN-based behavioral model employs orthonormal Hermite polynomial functions rather than commonly used sigmoid function, as nonlinear activation functions resulting in significant improvement in accuracy and convergence. Experimental results on a class-AB PA driven with the wideband Mobile Multimedia Broadcasting signals show a superior performance of the proposed extended NN-based model over the traditional NN-based ones. IntroductionPower amplifiers (PAs) are essential devices for almost all communication systems which attract much attention among researchers. Despite recent years' advance in PA design, [1][2][3][4] there are still issues of nonlinearity and efficiency. To analyze and address these problems, various physical and behavioral modeling strategies have been proposed. Physical modeling uses an equivalent circuit description of PA and is appropriate to circuit-level simulation. However, it has a high price in simulation time and the need for a detailed description of the PA internal structure. When such an equivalent circuit is not available, or whenever a complete system-level simulation is desired, behavioral modeling is preferred, as it is only based on the relationship between the input and output of PA while providing a precise prediction of PA output. [5] There have been many kinds of PA behavioral models proposed and discussed in the literature, [6][7][8][9][10][11][12][13][14][15] which can, in general, be categorized into two classes, i.e., memoryless PA models and PA models with memory. Memoryless behavioral models, such as Saleh model [6], polynomial model [7] and so on, have been well established and widely used for many years due to their easy implementation, relative efficiency, and acceptable accuracy in narrowband applications. However, these memoryless models fail to precisely predict PAs with memory widely used in modern communication systems, thus