Sensitivity Analysis of Factors that Induce Road Collapses Due to Drainage Pipe Leakage and Traffic Load

Abstract

At persent, quantitatively assessing the sensitivity of disaster-causing factors in road collapses remains challenging using conventional methodologies. This study establishes a physical model test system for road collapse to investigate the evolution of cavity buried depth, equivalent diameter, road settlement, soil stress/settlement, and drainage density/flow under various traffic loads. A deep-learning backbone is created that integrates a convolutional neural network (CNN), bidirectional short-term memory (BiLSTM), attention mechanisms, and the Sobol method (CBAS). This model is applied to three functional units, including data generation, settlement prediction, and sensitivity analysis. The findings show that the buried depth of a cavity decreases exponentially over time, whereas the equivalent diameter of the cavity follows a power-law increase. The drainage density and flow demonstrate fluctuating characteristics, whereas the abrupt increase in soil settlement exhibits a “time lag” effect relative to the sudden surge in soil stress. Notably, the buried depth and equivalent diameter of the cavity are highly sensitive to road settlement, whereas the sensitivity of soil settlement to road settlement varies. By contrast, the sensitivities of soil stress, drainage density, and drainage flow to road settlement are relatively weak. The validity of the CNN–BiLSTM–ATTENTION backbone was verified using root mean square error, mean absolute percentage error, and mean absolute percentage error performance metrics. The CBAS’ sensitivity analysis results were validated using the extended Fourier amplitude sensitivity test. The study findings provide valuable insights for monitoring and early warning of road collapses.