Rail-GAN: Domain-Aware QoS Data Augmentation Framework for Link-State Prediction in Railway Cooperative Transmission

In heterogeneous railway cooperative transmission systems, accurate link-state prediction (e.g., QoS metrics like throughput and latency) is crucial for multipath scheduling. However, collecting large-scale real communication data in operational railway systems is costly, while conventional tool-based simulation pipelines struggle to reproduce the complex coupling between network evolution and link-state dynamics. This results in a scarcity of effective data for intelligent scheduling and performance prediction. To address this issue, this study proposes Rail-GAN, a domain-aware adaptation of the TimeGAN sequence-generation framework for heterogeneous railway cooperative transmission. Rail-GAN introduces QoS-specific preprocessing, full-dimensional moment matching, engineering consistency post-filtering, and source-to-target fine-tuning to generate multivariate QoS sequences under severe target-domain data scarcity. The generated data are evaluated from both statistical fidelity and downstream practical utility perspectives. Experimental results show that Rail-GAN achieves substantially better discriminative and predictive scores than the original TimeGAN baseline. In downstream LSTM-based prediction tasks, augmenting a scarce target-domain dataset with Rail-GAN synthetic samples reduces the MAE for TCP rate and packet loss rate by 31.19% and 26.61%, respectively, while RTT exhibits a mild 4.20% degradation. This result suggests that the current smoothing and moment-matching design attenuates rare RTT bursts, which remains an important limitation and motivates future burst-aware generative modeling. Overall, Rail-GAN provides a practical synthetic-data foundation for intelligent operation, congestion control, and resource allocation in highly dynamic railway communication networks.