Study On Nonlinear Deformation Behavior Prediction Of Rock And Soil With Support Vector Machines

Prediction is one of heart problems of scientific study. And it is also a significant prerequisite for scientific decision-making and planning. The prediction about the behavior of rock and soil, which is a rub and problem crying out for solutions in the field of geotechnical engineering, gradually becoming an active domain of academics and engineering all over the world.The dissertation integrates support vector machines that an excellent machine learning algorithm with the prediction of nonlinear deformation behavior. This paper studies the representation of deformation time series (DTS) data, the construction of support vector kernels, and the hyper-parameters selection, and the deformation time series extrapolation prediction. The main work includes the following:(1) For the data representation, data reproduction and reconstruction methods are proposed for deformation time series. Some simple pretreatment works were done for past forecasting, but it is not clearly enough to accurate prediction. Aim at the lower efficient modeling due to limited data number, the piecewise interpolation and Bootstrap technique are introduced to supplement or expend the sample capacity, and enrich evolution details of rock and earth mass system. Still, low dimension coordinate transform and high dimension transform of phase space are posed. By reconstructing representation form of deformation time series to bring out rewrite data regularity, and then discover the complicated essential characteristic.(2) By analyzing the curve characters of classical displacement time series, and discussing some general properties of Mercer kernels, an interrelated selection principle of kernel for SVR prediction and modeling is presented. And we use closure properties theoretically to create optimal or complicated kernels to match the domain problem that SVM prediction of deformation measurement data. The complicated kernel enlarged functions capacity and strengthened applied flexibility, and thus the better trade-off performance among single Mercer kernels was obtained. This is favorable to improvement of regression precision and generalization. In terms of kernel mapping, the prior knowledge, which input data point more relative with adjacent data points about deformation time series, were utilized to adjust kernel matrix or kerne 1 itself for increasing performance matching problem on hand. In addition, considering good kernels are of similar form, an all-data- information translation invariant kernel and a uniform kernel function are introduced after seriously analyzed good kernel form. The former retain local and global input data information, thus input sample data and kernel matrix information on feather space or model selection are extended. The latter, in fact, a limited uniform kernel function of distance and dot product kernel, leaves out the complicated selection of kernel in the course of SVR modeling.(3) For hyper-parameters selection of SVR algorithm, an orthogonal experimental design procedure for hyper-parameter selection (ODPS) and boosting committee techniques are proposed. Analyzed the range value of hyper-parameters for SVM, determined experimental levels for different parameters, and taken orthogonal and interaction effect of hyper-parameters into account, the ODPS method which is better than the others on the time effect could choose hierarchically the SVM model of optimal hyper-parameters combination. On the other hand, in process of the parameters optimal of boosting algorithm, each sub-SVR is trained independently and the training set is created based on weighted random sampling from the original dataset. And according to regression error to update the sample's probability distribution and re-sample to form train set of the next sub-SVR. The final SVR solution is produced by aggregating the sub-SVR results through certain methods such as least-squares estimation based weighting etc.(4) To improve extrapolation precision of rock and earth mass deformation time series, a united SVR prediction method base on advanced judgment for deformation tendency is presented. There are two way to realize tendency advanced judgment: expert recognition, firstly, depends on qualitative and quantitative style to identify extension model which defined after the feature analysis of deformation and curve of time series. The second is case based reasoning (CBR). To find the similar case from cases database, the method mainly uses Euclidean distance to assess attribute similarity between target case and base case, and build up matrix structure of time series to carry through the similarity research on time series. The ultimate aim of method is judge tendency of target case time series refer to similar base case in future. Support vector machines learning and modeling based on advanced judgment could achieve efficient models of good generalization in which the number of sample data is very limited or time series without key point information.