Komprimerade maskininlärningsmodeller för digital frontend-behandling i trådlösa kommunikationssystem

Digital Front-End (DFE) processing plays a crucial role in modern wireless communication systems by shaping baseband signals and mitigating hardware imperfections before radio frequency transmission. An important challenge in this domain is the non-linear behavior of Power Amplifiers (PAs), which introduces signal distortion and degrades system performance. Traditional compensation techniques, such as Volterra Series and Generalized Memory Polynomial (GMP) models, lack flexibility and struggle to generalize across varying PA architectures. This thesis investigates the use of Machine Learning (ML) models as a solution to replace Digital Pre-distortion (DPD) blocks, offering improved adaptability and linearization capabilities. However, the computational complexity of dense Neural Networks (NNs) brings significant limitations to resource-constrained environments. To address this, two widely used model compression techniques, network pruning and quantization, are explored. The goal is to reduce memory footprint and computational demands without significantly compromising key performance metrics such as Normalized Mean Squared Error (NMSE), Adjacent Channel Leakage Ratio (ACLR), and Peak-to-Average Power Ratio (PAPR). Experimental results demonstrate that pruning up to 80% of model parameters can be achieved while achieving adequate linearity metrics. Similarly, quantizing the model to 8- and 16-bit precision retains satisfactory accuracy, proving the robustness of ML-based solutions under compression. These findings highlight the potential of compressed ML models for efficient and high-performance deployment in DFE systems.