Objective Collapse is a major engineering hazard during the foundation pit construction using semi-covered excavation method, and risk assessment is one of the significant means to substantially reduce such hazards. Therefore, in-depth research on risk assessment methods is required to achieve advanced and accurate assessment of metro foundation pit collapse risks.

Method In view of the practical scenario when on-site decision-makers lack sufficient response time, a high-accuracy multi-data fusion method is proposed based on machine learning and improved Dempster-Shafer (D-S) evidence theory. This method employs a multi-step rolling approach to compare three small-sample machine learning models, thus obtaining the optimal prediction results, which are then transformed into Basic Probability Assignment (BPA) values using a cloud model. A novel evidence correction parameter is defined based on conflict degree and divergence degree, while a fine-tuning term is introduced to reduce the impact of subjectivity in global focal element assignment. By fusing multiple prediction data sets (ground settlement, crown displacement, horizontal convergence displacement), the failure rate of high-conflict evidence fusion is reduced, and the accuracy and robustness of the risk assessment results are enhanced.

Result & Conclusion  The proposed method has been successfully applied to the Shiyao Station project of Nanchang Metro Line 3, providing decision-makers with longer response time. Ultimately, only minor deformation occurs in the surrounding ground and a foundation pit collapse is successfully prevented. The proposed high-accuracy advanced risk assessment method demonstrates good applicability and effectiveness in assessing collapse risks for metro foundation pits.