Semantic Freshness: A System-Level Metric for Real-Time Semantic Digital Twins

Abstract

Real-time semantic digital twins (DTs) depend on mobile networks to keep their knowledge graph (KG) state continuously aligned with the physical world. Existing architectures route this transformation through a central cloud server, forcing full-resolution sensor data across the network before most of it is discarded as semantically irrelevant. This paper proposes a distributed architecture in which semantic processing is moved to the user device so that compact semantic tokens, rather than raw pixels, traverse the mobile network. To evaluate timeliness in such systems, we define Semantic Freshness (SF) as the latency from physical observation to the moment the corresponding semantic state becomes queryable in a KG. We evaluate this metric on a real 5G testbed. Compared with centralized video streaming, token-based transmission reduces uplink bandwidth by 83-fold and mean SF by 3.4-fold, yielding sub-second median semantic-state availability. Stage-wise SF decomposition further reveals that, once raw video transport is eliminated, end-to-end latency is dominated not by the radio link but by the KG commit stage. These results demonstrate that relocating semantic filtering to the user equipment (UE) is an effective and scalable approach to real-time DTs under mobile constraints.