| With the rapid development of urban rail transit in China,it is often necessary to carry out foundation pit excavation during the construction of rail transit.For rail transit the retaining structure of foundation pit deformation monitoring,timely warning by monitoring data of research reports and prevention is the important method and the link in the process of rail transit construction.Among the existing deformation prediction models,SVM is the representative of the mature algorithm.In view of the SVM model parameter value selection in difficult problems and disadvantages based on a single point of data modeling,Particle Swarm Optimization(PSO)is proposed to optimize the SVM model in this paper and apply it to the settlement prediction modeling of subway deep foundation pit.Paper specific research content is as follows:(1)Machine learning and SVM theory are described,with the return of the regression model as an example this paper introduces the machine learning and classification problems.By introducing machine learning and statistics theory,SVM theory is introduced,including support vector classifier and support vector regression machine.In order to verify that SVM model has good generalization ability in dealing with settlement monitoring of foundation pit with high data and nonlinear,the SVM model and BP neural network performance comparison verification experiment,the results show that the SVM model has the characteristics of high fitting precision and good performance.However,the parameters of SVM model are fixed values,and the selection of model parameters is subject to subjectivity,so specific algorithm improvements are needed.(2)In view of the traditional inertial parameters of the SVM model for fixed value this problem,PSO optimization algorithm of SVM model parameter optimization.Because of the demand of different prediction model on the number of samples is different,build the PSOSVM model steps to solve the optimal training sample size;In addition,under the condition of the optimal training sample,the prediction accuracy of PSO-SVM model was compared with that of the traditional BP neural network and SVM model.The results showed that the prediction effect of PSO-SVM model was the best and the difficulty in parameter selection of SVM model was overcome.The PSO-SVM model is based on single point only monitoring data modeling problems,therefore need to further optimize the PSO-SVM model.(3)In view of the shortcoming of PSO-SVM without considering the interaction between other subsidence deformation factors,the PSO-SVM model is established by introducing three factors.Through correlation analysis,it is found that the three factors of "time","historical data" and "settlement deformation value at adjacent points" have strong correlation with foundation pit settlement deformation,so these three factors are introduced into PSO-SVM model for modeling analysis.The analysis results show that the PSO-SVM model based on multiple factors under short-term sample fitting prediction effect is best,under the condition of the medium and long term sample data based on multi-factor model fitting accuracy is poorer.Based on the multi-factor PSO-SVM model,the 150-phase monitoring data of dbc-02-1,dbc-02-2 and dbc-02-3,a group of monitoring points for deep foundation pit settlement of Liuxiandong station on line 13 of Shenzhen urban rail transit,were taken as sample data to improve the algorithm by combining multi-scale one-dimensional wavelet decomposition function group and cauchy-distribution function.The results show that the Improved PSO-SVM model based on multiple factors can adapt to deformation prediction under different sample sizes,the improved model has obvious parameter optimization,high fitting accuracy and stable prediction performance.Based on "time","historical data" and "near point settlement value" Improved PSO-SVM model fitting performance is much better than the traditional model,the prediction precision is higher.Compared with the PSO-SVM model,the PSO-SVM model based on multiple factors integrates the factors affecting the subsidence deformation,and the model establishment is more in line with the reality,Improved model overcame the PSO-SVM model based on multiple factors in the medium to long term sample under the disadvantage of poor precision,more suitable for the deformation of metro construction of the high risk project early warning,for deep foundation pit deformation monitoring provides a more scientific and effective early warning method,has higher application value. |
