A diffusion model-driven data augmentation framework for improving machine learning prediction of rockburst levels

Rockbursts are a major hazard in deep underground engineering, posing serious threats to construction safety and human life. Accurate rockburst prediction is therefore critical for engineering practice. Although machine learning has shown promise for rockburst prediction, its performance is often constrained by limited datasets, costly data collection, and class imbalance. This study proposes RTB-DDPM, a diffusion-based data augmentation model for rockburst tabular data, to generate high-fidelity, engineering-feasible synthetic samples. RTB-DDPM integrates class-guided sampling, a residual multilayer perceptron noise predictor, a time-embedding module, and boundary-level physical constraints to improve generation fidelity, class consistency, physical plausibility, and controllability. Experiments show that RTB-DDPM achieves the lowest distributional discrepancies, with a maximum mean discrepancy (MMD) of 0.007 and an earth mover's distance (EMD) of 0.117 and 0.122 in two representative cases. It also reduces training time to 27.3 s and 30.8 s and increases generation throughput to 35.9 and 67.1 samples per second compared with the baseline diffusion model. Results from both cases further indicate that RTB-DDPM outperforms several state-of-the-art generative methods in distributional agreement and statistical feature reconstruction. When applied to rockburst prediction, the augmented data improve average accuracy by 20.3% in Case 1 and 5.7% in Case 2 across six mainstream models, indicating better generalization and robustness. In addition, an intelligent early-warning software platform has been developed to support practical rockburst prediction. Overall, the proposed framework offers a practical pathway to alleviate data scarcity in rockburst engineering.