Impairments Compensation in 5G PAs Using Neural Networks

A Master of Science thesis in Electrical Engineering by Reem Al Najjar entitled, “Impairments Compensation in 5G PAs Using Neural Networks”, submitted in May 2024. Thesis advisor is Dr. Oualid Hammi. Soft copy is available (Thesis, Completion Certificate, Approval Signatures, and AUS Archives Consent Form).

Abstract

In the fields of behavioral modeling and PAs predistortion, neural networks have recently demonstrated their superior performance. These networks perform well as predistorters because they efficiently carry out complex calculations and capture the essential traits of nonlinear systems. This study presents a novel hybrid model that combines a neural network, in combination with a look-up table, to create a digital predistorter for PAs linearization. The main motivation being to use the look-up table to eliminate the highly nonlinear static distortions of the PA, and subsequently focusing the neural networks on the compensation of dynamic distortions in a manner that both sub-models complement each other. -Such approach was found to lead to excellent results-. The ZHL-42 driver and the CREE CGH40010 PA were used in the experimental setup. The instrumental equipment was the Anritsu MS2830A, which included a vector signal generator and a vector signal analyzer. The signal in use was a fifth-generation with a 40MHz four carrier bandwidth. The mean square error metric was used to assess the neural network model performance, while the adjacent channel leakage ratio was used to assess the effectiveness of the cascaded neural network and look-up table predistorter and their ability to effectively linearize the PA. In order to reduce the complexity of each block independent and obtain the best performance, this research focuses on merging a look-up table with a neural network model. Since scalability is an advantage of using a neural network, another goal is to achieve scalability and linearize the PA on various signals. Through these encouraging results, this all-encompassing strategy attempts to ultimately advance PAs linearization methods by taking advantage of the operational efficiency and synergy of the neural network and look-up table models when combined together.