Mechanical Parameters Inversion Of Rock Mass With Support Vector Machine And Its Engineering Application

The displacement back analysis based on support vector machine (SVM) is a newly appeared parameters identification method of geotechnical engineering, which realizes the complex mapping relationship between mechanical parameters and deformations of rock mass using SVM model determined by a few learning examples instead of numerical model and improves inversion efficiency. Compared with artificial neural network (ANN), which is earlier used for back analysis, the SVM bears more excellent characteristics both in basic theory and solution method, so it has been paid increasingly more attention by researchers in geotechnical engineering.In allusion to the disadvantages of present back analysis method, the numerical modeling method of underground engineering is discussed first in detailed, and then the method of rock mechanical parameter identification based on SVM is studied systematically from three aspects, parameters optimization of SVM, type of SVM, and the kernel function form, which are crucial factors affecting the generation ability of SVM, finally the research results was applied to real project. The major contents are as follows:(1) The numerical modeling method of underground engineering based on natural element method (NEM) was discussed. Aimed at the defect that the errors was inevitably caused by selecting a certain finite region and setting the boundary conditions subjectively when treating infinite or half infinite domain problem in geotechnical engineering, the infinite element method(IEM) was introduced to simulate the boundary conditions at infinitely distant place, thus the coupling analysis method of NEM and IEM was provided. Through an example, the correctness of the computation method was proved. The results show that the coupling method improves the calculation precision and reduces the requirement for calculation range; meanwhile, visco-elastic analysis based on the coupling method is successfully implemented, which extends the application scope of NEM in geotechnical engineering.(2) In order to overcome the shortcomings of conventional parameter selection methods, the parameter selection method based on particle swarm optimization (PSO) is addressed. By modifying the comparison mode of fitness value and the motion mode of individual particle of conventional PSO, and designing an auto-adaptive dynamic inertia weight which considers both the time and clustering degree of particles, an improved particle swarm optimization (IPSO) is proposed. Numerical results showed that, compared with traditional parameter selection methods, the new method obviously increases the parameter searching efficiency and the predictive precision of corresponding SVM.(3) The surrounding rock parameters identification method based on least squares support vector machine (LS-SVM) is studied. By substituting LS-SVM for standard SVM, and combined with the improved particle swarm optimization (IPSO), a novel displacement back analysis method is provided. To ensure the sparsity of support vectors of LS-SVM, an iterative regression algorithm based on quadratic Renyi entropy and incremental learning algorithm is chosen to solve the LS-SVM model. Through examples of back analysis, the feasibility and effectiveness of applying LS-SVM to parameters inversion of rock mass is demonstrated, and the results show that, compared with standard SVM, inversion efficiency is greatly improved using LS-SVM.(4) The influence of kernel function form on inversion results is discussed. Due to the lack of orthogonality by translation, the RBF kernel function, which is generally used for back analysis at present, restricts the generalization performance of corresponding SVM and impairs inversion precision indirectly. In view of this, the wavelet function and scaling function were used as kernel function for parameter inversion. In addition, aiming at the deficiency of locality of conventional scaling function, a new compact support scaling kernel function with good smoothness was constructed according to the related theory. The results of an inversion example showed that the approximation ability of kernel function has serious influence on parameter identification precision, and when the admissible accuracy is satisfied, less training examples is needed if using kernel function with better approximation ability, which means less computation cost.(5) Based on above work, application study on back analysis was performed in Wangkeng expressway tunnel project. Using the coupling method of NEM and IEM, the 3D numerical model simulating tunnel construction was build, and the influence of excavation damage zone was considered. Through influence degree analysis, the parameters that would be identified were determined. Then the compact support scaling kernel LS-SVM was used for mechanical parameters inversion of surrounding rock. The space displacement sequence grey correction analysis indicated that the inversion results were credible. Finally, based on the identified parameters, the deformations of surrounding rock caused by subsequent excavation were successfully predicted.