Mitigating potential risk via counterfactual explanation generation in blast-based tunnel construction

Machine learning and deep learning have significantly enhanced the ability to mitigate risks in blast-based tunnel construction. However, most studies fall short in model constraints, data quality, and explainability, making nonrobust risk mitigation strategies. Therefore, this study aims to investigate the following questions: how to accurately assess risk for blast-based tunnel construction using limited data, and develop effective risk mitigation strategies? This research leverages counterfactual explanation generation, a key technique of explainable artificial intelligence, along with data augmentation to develop a framework for guiding risk mitigation, which includes: (1) a two-stage data augmentation technique to address data shortage and imbalance; (2) a novel counterfactual explanation generation algorithm to optimize blasting parameters and reduce risk; and (3) a post-hoc explainable approach to provide insights on feature importance. A railway tunnel in Hubei is conducted as a case study to test the validity of the proposed method. The results show that the proposed method accurately predict overbreak, achieving the highest R2 (0.883) and the lowest RMSE (1.335) compared to baseline models. Additionally, it effectively optimizes the blasting parameters to mitigate risk, reducing the average overbreak in six scenarios. The explainable analytic identifies key factors (e.g., periphery hole spacing) influencing construction risk, thereby enhancing personnel's understanding of complex construction systems.