RecuGAN: A Novel Generative AI Approach for Synthesizing RF Coverage Maps

This is the authors' abstract. We don't add key points for this paper.

Radio-frequency coverage maps (RF maps) are essential in wireless communication, but obtaining them through site surveys can be labor-intensive and sometimes impractical. To address this challenge, we propose RecuGAN, a generative adversarial network (GAN)-based approach for generating RF maps. RecuGAN leverages the principles of information maximizing GAN (InfoGAN) to capture latent properties of RF maps, enabling unsupervised categorization and generation of new and diverse RF maps. Unlike traditional methods, RecuGAN does not require labeled data or conditional input, reducing complexity, time, and cost. We enhance the RecuGAN objective function with a customized gradient penalty-based Wasserstein GAN (WGAN) function and a gradient-based loss function for stable training and accurate map generation. We also provide the option to incorporate multiple generators in RecuGAN, enabling high-resolution RF map generation. As demonstrated through extensive training with both experimental and simulation data, RecuGAN can synthesize diverse high-quality RF maps and categorize them based on the RSS distribution. Compared to a UNet-based conditional GAN (cGAN), RecuGAN achieves a mean average percentage error (MAPE) of 1.18%, outperforming the cGAN model, which achieves a MAPE of 2.5%.