Mechanism-Based Prediction and Multi-Field Modelling of Shallow-Buried Tunnel Subsidence: A Dual-Channel Algorithmic Approach

Against the backdrop of frequent extreme short-duration heavy rainfall events, the precursors to sudden ground subsidence in the overlying strata of shallow-buried tunnels are often subtle, and the relationships between multiple field responses remain unclear; consequently, traditional single-field threshold-based early warning systems struggle to balance early detection with reliability. To address the challenges of multi-source, heterogeneous monitoring and graded early warning for ground subsidence in shallow-buried tunnels induced by heavy rainfall, a risk-state judgement framework integrating statistical learning and mechanistic constraints has been developed. Based on networked parallel electrical resistivity tomography (ERT), fibre Bragg gratings (FBGs) and digital image correlation (DIC) physical experiments, the temporal rates of change and spatial evolution characteristics of the three-field responses were organised into six-dimensional inputs. Following baseline normalisation, these were fed into a random forest algorithm, with ensemble voting used to characterise the relative support levels for four state categories; an asymmetric misclassification cost was introduced, and the initial level was determined by minimising conditional risk, with the final output constrained by mechanism-triggered risk thresholds and hysteresis rules. The results indicate that, in this experiment, representative precursors related to electrical properties, internal strain and macroscopic deformation appeared in sequence, preceding the collapse by 2157, 1203 and 166 s respectively. Based on the existing sample, the internal validation accuracy was 96.5 per cent, the macro-average F1 score was 94.7 per cent, and the recall rate for the “red” level was 91.3 per cent; the global feature contributions of FBG, ERT and DIC were 0.42, 0.37 and 0.21, respectively. Equivalent hydromechanical calculations provide a supplementary explanation for the relationship between seepage, stress adjustment and deformation development. This framework maps multi-field characteristics onto discrete early-warning levels that balance the cost of misclassification with physical conditions. The proposed method transforms the objective limitation of macroscopic precursor lag into the necessity for early, coordinated multi-field early warning, thereby providing a mechanistically sound and transferable graded early-warning framework for shallow-buried underground structures under heavy rainfall conditions.