Transformer-Enhanced Successive Interference Cancellation for Channel-Resilient NOMA Systems

Non-orthogonal multiple access (NOMA) with successive interference cancellation (SIC) has emerged as a promising technology for next-generation wireless networks. However, conventional SIC performance degrades significantly under imperfect channel state information (CSI), a common scenario in practical implementations. This paper presents a Transformerenhanced SIC architecture that leverages the self-attention mechanism to maintain robust performance despite channel estimation errors. Our proposed method features a cascade of Transformers that process received signals and likelihood outputs sequentially, enabling effective interference cancellation without relying on precise channel knowledge. The architecture demonstrates remarkable adaptability to various NOMA configurations, including changes in power allocation and user numbers, while maintaining performance even without retraining. Numerical results show that our Transformer-enhanced SIC consistently outperforms conventional SIC and existing deep learning approaches under imperfect CSI conditions, exhibiting robust performance across different system configurations. This work offers a practical solution for NOMA systems operating under realistic channel conditions.