Exploiting age of information in network digital twins for AI-driven real-time link blockage detection

• An automated large-scale data collection framework is proposed for machine learning models integrating network digital twins , raytracing, and crowd-sourcing. This approach enables scale-out solutions for deploying site-specific machine learning models efficiently and effectively • The study proposes a resolution reduction technique for MIMO-OFDM channel samples, achieving computational speedup factors of 16 and 32 times within the simulated settings, while simultaneously enhancing accuracy performance. • The age of information metric is introduced within the context of network digital twins data, and its integration into the loss function for deep learning is proposed. The integration of the metric into the training process effectively addresses model performance degradation in dynamic wireless environments, achieving robust results with only 1% of the available data. • An extensive simulation study has been conducted to evaluate the proposed data collection framework, age of information integration, and resolution reduction techniques, achieving over 98% accuracy in the blockage detection task. The Line-of-Sight (LoS) identification is crucial to ensure reliable high-frequency communication links, especially those vulnerable to blockages. Network Digital Twins and Artificial Intelligence are key technologies enabling blockage detection (LoS identification) for high-frequency wireless systems, e.g., 6 > GHz. In this work, we enhance Network Digital Twins by incorporating Age of Information (AoI) metrics, a quantification of status update freshness, enabling reliable real-time blockage detection (LoS identification) in dynamic wireless environments. By integrating raytracing techniques, we automate large-scale collection and labeling of channel data, specifically tailored to the evolving conditions of the environment. The introduced AoI is integrated with the loss function to prioritize more recent information to fine-tune deep learning models in case of performance degradation (model drift). The effectiveness of the proposed solution is demonstrated in realistic urban simulations, highlighting the trade-off between input resolution, computational cost, and model performance. A resolution reduction of 4 × 8 from an original channel sample size of ( 32 , 1024 ) along the angle and subcarrier dimension results in a computational speedup of 32 times. The proposed fine-tuning successfully mitigates performance degradation while requiring only 1 % of the available data samples, enabling automated and fast mitigation of model drifts.