Novel Neural Network Model of Power Amplifier Plus Iq Imbalances

Abstract

Abstract-Traditionally, the transmitter (TX) IQ imbalances distortion and power amplifier (PA) distortion are separately modeled. In this paper, the behavior of the two distortions are unified, and characterized by a single model. Rectangular structured Focused Time-Delay Neural Network (RSFTDNN) is proposed to uniformly model IQ imbalances and PA distortions. As a result, the physical distortions in the analog circuits are further abstracted. It also saves computation resources in simulation. Unlike the polynomial based model, which suffers from condition number effects and inaccuracy for deeply nonlinear system, the proposed RSFTDNN shows high accuracy. Two cases of real experiments are carried out, where RSFTDNN model shows much better performance than the polynomial based model in the sense of model accuracy.

Low-Pass Equivalent Behavioral Modeling of Rf Power Amplifiers Using Two Independent Real-Valued Feed-Forward Neural Networks

“…Volterra series [7][8][9] and artificial neural networks (ANNs) [10][11][12][13][14][15] are the most widely reported techniques that can provide an adequate mathematical description for the PA behavior. The selection of a particular technique targets the improvement of the trade-off between increasing modeling accuracy and reducing computational cost.…”

“…This work first proposes a novel ANN-based PA behavioral model that, contrarily to the previous approaches in [11][12][13][14][15], requires two independent real-valued ANNs, each network having just one output. Second, based on input-output data measured on a GaN HEMT class AB PA, it reports a better tradeoff between modeling accuracy and computational complexity of the proposed ANN-based model than previous approaches.…”

“…ANNs having only real-valued inputs, outputs, weights and biases, are widely reported in literature for the low-pass equivalent behavioral modeling of PAs [11][12][13][14][15]. Since low-pass equivalent behavioral models relate complex-valued envelope signals at the PA input and output, complex-valued signals must be converted into real-valued signals.…”

“…Indeed, different approaches have been proposed to perform these conversions. Specifically, in [11][12][13] the complex-valued input and output envelopes are decomposed into rectangular components, in [14] into polar coordinates and in [15] into amplitude and phase difference components. However, the ANN-based behavioral models in [11][12][13][14][15] use a single ANN having two outputs that estimate real-valued components of the PA output signal.…”

“…In literature, there are various techniques that can provide an adequate mathematical description for the PA behavior. Volterra series [7][8][9] and artificial neural networks (ANNs) [10][11][12][13][14][15] are the most widely reported techniques that can simultaneously describe nonlinear and dynamic behaviors. ANNs have the advantage of requiring a lower number of parameters than the Volterra series and have more general validity than polynomial approximations.…”