Predicting for Ground Surface Settlement Induced by Shield Tunneling in Upper-soft and Lower-hard Ground Based on RS-SVR

To improve the prediction accuracy of ground surface settlement induced by shield tunneling in heterogeneous ground, a model based on rough set-support vector regression (RS-SVR) for predicting ground surface settlement was established and applied to ground settlement in actual subway tunnel engineering. Conditional attributes affecting ground settlement including geometric, ground, and shield construction factors were selected according to specific geological conditions. Pawlak's degree of attribute method of rough set theory was used to delete redundant data to obtain the optimal set of attribute sets for ground settlement.On this basis, support vector regression (SVR) was applied to establish an RS-SVR ground settlement prediction model and was compared with the SVR model without attribute reduction.Moreover, to compare the influence of different kernel functions, a radial basis function (RBF), sigmoid function, and polynomial function were applied as kernel functions for regression prediction for training samples and test samples in RS-SVR and SVR models.Finally, the models were tested with 20 sets of ground settlement monitoring data of upper-soft and lower-hard ground in the Nanhu section of Foshan Metro Line 2.The results show that attribute reduction can condense 12 conditional attributes that affect ground settlement to the optimal conditional attribute set containing 7 items (hard layer ratio α, cohesive force c, internal friction angle φ, pressure of the soil bin, total thrust, torque of the cutter disk, and the driving time). Classification results with attribute reduction are the same as those without attribute reduction.When compared with a similar model,the prediction errors of RBF as a kernel function on RS-SVR and SVR models are 5.54% and 13.10%, respectively, which are lower than the prediction error when the sigmoid and polynomial functions are used as kernel functions.The prediction errors of the RS-SVR model are 5.54%, 11.48%, and 13.26%, respectively, which are lower than the SVR model prediction errors of 13.10%, 15.71%, and 19.68% when the same core function is used for longitudinal contrast. © 2018, Editorial Department of China Journal of Highway and Transport. All right reserved.