Data Augmentation for Predictive Digital Twin Channel: Learning Multi-Domain Correlations by Convolutional TimeGAN

Abstract

In order to realize advanced system design for the sophisticated mobile networks, predictive digital twin (DT) channel is constructed via data-driven approaches to provide high-accuracy channel prediction. However, lacking sufficient time-series datasets leads to overfitting, which degrades the prediction accuracy of the DT channel. In this article, data augmentation is investigated for constructing the predictive DT channel, while enhancing its capability of tackling channel aging problem. The feature space needs to be learned by guaranteeing that the synthetic datasets have the same channel coefficient distribution and time-frequency-space domain correlations as the original ones. Therefore, convolutional time-series generative adversarial network (TimeGAN) is proposed to capture the intrinsic features of the original datasets and then generate synthetic samples. Specifically, the embedding network and recovery network provide a latent space by reducing the dimensions of the original channel datasets, while adversarial learning operates in this space via sequence generator and sequence discriminator. Simulation results demonstrate that the synthetic dataset has the same channel coefficient distribution and multi-domain correlations as the original one. Moreover, the proposed data augmentation scheme effectively improves the prediction accuracy of the DT channel in a dynamic wireless environment, thereby increasing the achievable spectral efficiency in an aging channel.