Special Column on Key Scientific Problems and Technological Breakthroughs in Mega Tunnel Engineering Construction
HE Chuan, CHEN Zi-quan, ZHOU Zi-han, MA Wei-bin, WANG Bo, ZHANG Jin-long
With the rapid development of artificial intelligence, deep learning algorithms for nonlinear propose a new approach for solving the persistent dilemma of tunnels and underground engineering relying on empirical designs. In this study, by fusing multiple indices (mechanical and deformation control indices) with the correlation coefficient of the support system synergy degree, an evaluation standard for support systems, characterized by the degree of fit, was proposed. Using this evaluation standard, the data of 718 highway and 486 railway tunnel sections were collected to build a database for algorithm training. Eight attributes about the background information of tunnel engineering, including rock hardness degree, integrity degree, rock thickness, underground water volume, buried depth level, geological structure, construction method, and internal contour type, were considered input indicators. Eight attributes of the support system, including shotcrete+steel mesh, rock bolt, steel arch, secondary lining, and auxiliary measures, were considered output indicators. The input and output indicators were then quantified. After comparing the characteristics of the PSO-SVM, SA-PSO-SVM, and CLS-PSO-SVM in the application of the intelligent feedback model of the support system, the generated intelligent feedback model was tested. The results show that the evaluation method first eliminates the weak design scheme. The degrees of fit of the strong support and general support schemes are 4.28 and 4.68, respectively, which verifies that the method can evaluate the material utilization rate while ensuring structural safety. Among the three intelligent algorithms, the CLS-PSO-SVM algorithm, with the broadest search range, had the highest feedback accuracy but the longest time consumption, whereas the PSO-SVM algorithm had the shortest time consumption but the lowest accuracy. Finally, the accuracies of the five output labels designed using the CLS-PSO-SVM algorithm are 93.4%, 92.6%, 89.3%, 91.8%, and 94.3%. The collective accuracy of the five output indices is 81.1%.