Study And Application Of SVM Optimization Model On Foundation Pit Deformation Prediction

With the rapid development of the city, the deep foundation pits are increasing emerged, so that the deformation monitoring and analysis of foundation pit is essential to ensure safety of deep foundation pit construction. Due to the deep foundation pit is mainly located in densely populated districts and commercial compact districts of the cities, which resulted in a tremendous pressure load of it. Therefore, to effective real-time monitoring and analysis the security of foundation pit as well as surrounding buildings to make timely forecasts to avoid accidents is crucial. However, the monitoring data are very sensitive to many factors when the foundation pit is monitored, including underground water level, deep horizontal displacement, soil pressure and temperature. In addition, the randomness each factor has led to the complex relationship between each other, which resulted in the traditional mathematical model that unable to quantitatively describe the relationship between these factors and foundation settlement. Sometimes, because of the limitation of conditions made it difficult for us to obtain complete and accurate data of these factors that led to the collected information of the data deficiency and the lager error noise effect. Hence, in order to increase the predicting accuracy of deformation monitoring of deep foundation pit, the establishment of combination deformation monitoring and alerting model of deep foundation pit deformation has become an important issue. In this study, an improved grey prediction model(GM(1,1)), neural networks model, least squares support vector machine model(LSSVM), particle swarm optimization(PSO) and its improved algorithm were proposed in combining an engineering example of metro railway foundation pit. And the results of different combination model were compared in this study.The main contents researched in this paper are as follows:(1) This paper addressed grey system discusses and neural network basic theories. In this paper, we analysis the improved method GM(1,1) model, the mechanism of neural network modeling, RBF neural network modeling and the derivations of formulas of initial value correction improvement method. By combining with the advantages and disadvantages of two models, we discussed the combination type of neural network and gray system(serial, parallel) and the various combinations of modeling process were described in detail in this paper. Thus, a modified residuals model of improved GM(1,1) and radial basis function(RBF) neural network fusion(GM-RBF) were established. This paper analyzes the feasibility of this combination model base on an engineering example, results showed that GM-RBF model feasible, and this one has higher precision than single GM(1,1) model and RBF neural network model.(2) The basic theory of LSSVM, PSO and its improved algorithm were study in this paper. The results indicated that LSSVM modeling larger affected by penalty factor c and kernel width coefficient ?, and the PSO has self-defect. Therefore, an improved strategy of PSO and improved scheme of the adjusting inertia weight and contraction factor were proposed in this paper. Accordingly, we founded a based on chaotic particle swarm optimization algorithm of LSSVM of optimal combination model(CPSO- LSSVM). Through instances, the improved CPSO-LSSVM model has higher precision than LSSVM model.(3) A CPSO-LS SVM combination optimization model and GM-RBF residual error correction model was established in this paper, and we analysis and discussion of the effect of a combination of application model by case study. By comparison of CPSO-LSSVM and GM-RBF combination model, LSSVM model and CPSO-LSSVM model in applied in deep foundation pit monitoring, it is showed that the forward model has higher precision than the latter two.