A Hybrid Quantum–Classical Framework for Synthetic Data-Driven Railway Defect Prediction and Safety Enhancement

Abstract

Track geometry defect detection is often limited by the scarcity of defective observations in inspection data sets, which reduces the effectiveness of machine learning models. This study evaluates whether synthetic data can improve defect prediction under such conditions. Classical neural network (CNN) and hybrid classical-quantum neural network (HCQNN) models are trained on data sets where the limited defective samples are supplemented with synthetic data generated using a conditional tabular generative adversarial network (CTGAN) and tabular variational autoencoder (TVAE). This is compared to models trained on limited real defective samples alone. The results show that TVAE produces synthetic data that more closely match the real data distribution, as measured by Kolmogorov–Smirnov, Jensen–Shannon, and Wasserstein distances. This contributes to an improved predictive performance over the limited real-data baseline (HCQNN F1-scores of 0.932 and 0.923, respectively), while CTGAN does not achieve a comparable improvement (0.916). Across all cases, the HCQNN achieves higher performance than the CNN and shows greater gains when trained on TVAE-generated data, with the CNN F1-score ranging from 0.890 to 0.906. These results highlight the effectiveness of combining TVAE-based synthetic data generation with hybrid classical–quantum modeling for defect prediction under limited-data conditions.