LIU Boyang, LIU Yisheng, BAO Xueying, YANG Yi
Journal of Railway Engineering Society. 2026, 43(9): 90-97.
Research purposes：To address multi-source heterogeneity, quality fluctuation and error propagation of low-reliability modalities in alpine tunnel monitoring data, a quality-aware cross-modal fusion method is proposed. Structural response, environmental action, apparent defect and engineering inspection data are organized as window-level samples. A five-dimensional quality evaluation scheme is established, encompassing completeness, accuracy, consistency, timeliness, and traceability. A quality-aware gating correction is performed based on the modal-specific quality scores and reliability to suppress the contribution of low-quality data. LSTM is used to extract temporal degradation feature, and Transformer is employed to explore global cross-modal correlations, thereby enabling health state assessment and remaining useful life prediction.
Research conclusions：(1) A window-level representation and quality evaluation method for four types of monitoring data in alpine tunnels is established, enabling differentiated model input according to data quality. (2) The quality-aware gating mechanism can reduce the interference of low-credibility modalities on fusion results. (3) Engineering case validation shows that the proposed method achieves a health-state assessment accuracy of 89.7%, a RUL prediction MAPE of 6.3%, and an accuracy of 84.1% under 15% noise, outperforming traditional fusion methods. (4) Frost heave pressure, cumulative freeze-thaw fatigue damage, damage degradation rate, crack width, and temperature zero-crossing frequency are key features affecting prediction performance. (5) The proposed fusion method can provide a methodological foundation for operational monitoring analysis, defect identification and early warning, and predictive maintenance decision-making of tunnels in cold regions and complex environments.
Just Accepted