Pipeline for end-to-end training of DL-augmented LS channel estimation

Abstract

This work presents an enhanced training strategy for LSDNN (Least Squares Deep Neural Net- work) [1], a lightweight deep neural network designed to improve LS channel estimation in pilot-based OFDM systems. The approach offers two key advantages: (1) it enables LSDNN to be trained using practical data sources, such as transmitted codewords and demapper LLRs, and (2) it achieves superior Bit Error Rate (BER) performance through this end-to-end training. By removing the need for perfect CSI during training, the proposed method simplifies implementation in real-world scenarios and aligns with the goal of deploying the system on hardware such as Software Defined Radios (SDRs). Compared to our baseline model, LSDNN strikes a balance between performance and efficiency, with significantly lower computational complexity, fewer parameters, and reduced memory requirements, while delivering competitive results. The model’s ability to generalize to unseen channel models further underscores its potential for real-world applications. This work demonstrates how deep learning can complement traditional signal processing to create efficient, adapt- able, and high-performance wireless communication systems.