Fast and Accurate Cooperative Radio Map Estimation Enabled by GAN

TL;DR

This work tackles real-time radio map estimation in 6G by eliminating the need for transmitter information and instead exploiting distributed RSS samples and a geographical map through cooperative radio map estimation (CRME). It introduces GAN-CRME, a conditional GAN framework with a UNet generator that ingests a 2-channel image of RSS and map data to produce high-fidelity radio maps ${\mathbf P}$, augmented by adversarial training to improve accuracy and enable error correction when the map is imperfect. The method achieves high accuracy with relatively low data acquisition and computation, and demonstrates robustness to map perturbations by recovering missing environmental features, validated on the RadioMapSeer dataset with comparisons to RadioUNet. The approach promises enhanced scalability and applicability for real-time radio resource management in 6G, including potential extensions to 3D RME, cognitive radio, and digital twin scenarios, while leveraging ground-truth RSS as a supervisory signal.

Abstract

In the 6G era, real-time radio resource monitoring and management are urged to support diverse wireless-empowered applications. This calls for fast and accurate estimation on the distribution of the radio resources, which is usually represented by the spatial signal power strength over the geographical environment, known as a radio map. In this paper, we present a cooperative radio map estimation (CRME) approach enabled by the generative adversarial network (GAN), called as GAN-CRME, which features fast and accurate radio map estimation without the transmitters' information. The radio map is inferred by exploiting the interaction between distributed received signal strength (RSS) measurements at mobile users and the geographical map using a deep neural network estimator, resulting in low data-acquisition cost and computational complexity. Moreover, a GAN-based learning algorithm is proposed to boost the inference capability of the deep neural network estimator by exploiting the power of generative AI. Simulation results showcase that the proposed GAN-CRME is even capable of coarse error-correction when the geographical map information is inaccurate.

Figures (5)

• Figure 1: Diagram of the proposed GAN-CRME training process.
• Figure 2: Models used in the GAN-CRME architecture. Resolution is the number of pixels of the image in each feature channel along the $x$, $y$ axis. Filter is the number of pixels of each filter kernel along the $x$, $y$ axis. The input layer is concatenated in the last two layers.
• Figure 3: A radio map label (left) and estimation results using $300$ (middle) and $1000$ (right) RSS samples over the Standard Dataset.
• Figure 4: Illustration of the error-correction phenomenon by using $300$ RSS samples (middle) and $1000$ RSS samples (right) in GAN-CRME. The ground-truth label is presented in the left hand side.
• Figure 5: Effect of number of users on the estimation accuracy.