Memory-Efficient Low-Rank Fine-Tuning for AoA Estimation in AI-Enabled RANs

The integration of AI into radio access networks (RANs) marks a pivotal advancement toward realizing the vision of 6 G. In this work, we propose a novel AI-RAN framework that leverages low-rank fine-tuning to enhance Angle-of-Arrival (AoA) estimation in massive MIMO systems. Built upon the Open RAN (O-RAN) architecture, our system disaggregates traditional base stations and embeds intelligence at the network edge through a Vision Transformer (ViT) deployed in the near-real-time RAN Intelligent Controller (near-RT RIC). To enable lifelong adaptability under non-stationary environments, a non-real-time RIC (non-RT RIC) monitors data distributions via a FIFO buffer and triggers model updates using a variational Bayesian controller. We propose a low-rank fine-tuning algorithm that reduces training overhead while maintaining estimation accuracy. Extensive simulations demonstrate that our method achieves competitive performance under dynamic scenarios and significantly outperforms classical subspace methods and recent learning-based baselines, confirming its effectiveness and deployability in practical edge intelligence scenarios.