Knowledge-Driven 3D Semantic Spectrum Map: KE-VQ-Transformer Based UAV Semantic Communication and Map Completion

TL;DR

This paper introduces a novel AI-driven framework for 3D spectrum map completion using UAVs, integrating physical knowledge and a specialized transformer model to improve accuracy and physical consistency in spectrum monitoring.

Contribution

It proposes a knowledge-enhanced semantic spectrum map completion framework with a new transformer model and novel metrics, addressing interpretability and physical realism in spectrum map reconstruction.

Findings

Outperforms state-of-the-art benchmarks in RKMSE metrics.

Effectively captures physical signal propagation in spectrum map completion.

Enhances spectrum monitoring accuracy in complex environments.

Abstract

Artificial intelligence (AI)-native three-dimensional (3D) spectrum maps are crucial in spectrum monitoring for intelligent communication networks. However, it is challenging to obtain and transmit 3D spectrum maps in a spectrum-efficient, computation-efficient, and AI-driven manner, especially under complex communication environments and sparse sampling data. In this paper, we consider practical air-to-ground semantic communications for spectrum map completion, where the unmanned aerial vehicle (UAV) measures the spectrum at spatial points and extracts the spectrum semantics, which are then utilized to complete spectrum maps at the ground device. Since statistical machine learning can easily be misled by superficial data correlations with the lack of interpretability, we propose a novel knowledge-enhanced semantic spectrum map completion framework with two expert knowledge-driven constraints from physical signal propagation models. This framework can capture the real-world physics and avoid getting stuck in the mindset of superficial data distributions. Furthermore, a knowledge-enhanced vector-quantized Transformer (KE-VQ-Transformer) based multi-scale low-complex intelligent completion approach is proposed, where the sparse window is applied to avoid ultra-large 3D attention computation, and the multi-scale design improves the completion performance. The knowledge-enhanced mean square error (KMSE) and root KMSE (RKMSE) are introduced as novel metrics for semantic spectrum map completion that jointly consider the numerical precision and physical consistency with the signal propagation model, based on which a joint offline and online training method is developed with supervised and unsupervised knowledge loss. The simulation demonstrates that our proposed scheme outperforms the state-of-the-art benchmark schemes in terms of RKMSE.