Nonlinear Behavior Prediction Of Rock And Soil Based On Particle Swarm Optimization And Relevance Vector Machine

The problems existing in geotechnical engineering can be divided into prediction,optimization and modeling,etc.Using traditional engineering methods to solve these problems has disadvantages such as long cycle,high cost and difficult survey.Therefore,no matter considering the development of disciplines or the practical application of geotechnical engineering,it is crucial to apply the machine learning methods to the field of geotechnical engineering and conduct cross-research on geotechnical engineering.The commonly used machine learning methods include BP neural network model,gaussian process,and support vector machine model.The disadvantages of the above models themselves are difficult to meet the complex geotechnical engineering problems,and the prediction results are often unsatisfactory,too.Therefore,the prediction model with higher accuracy needs to be proposed and studied.Aimed at these problems this paper put forward the relevance vector machine model.Considering the complexity of rock and soil and the optimization of related parameters,Particle swarm optimization is carried out on the basis of relevance vector machine,and a relevance vector machine model based on particle swarm optimization is provided for the prediction problem of geotechnical discipline,trying to provide a new method for the acquisition of relevant data.In general,the research work of this paper is as follows:1.For nonlinear,high-dimensional and uncertain prediction problems in rock and soil,a relevance vector machine model based on particle swarm optimization is proposed.The applicability of PSO-RVM model is proved by using the compressibility coefficient of foundation soil and peripheral displacement during construction of wushi tunnel.At the same time,the PSO-RVM model is compared with the neural network model and the generalized regression neural network model,and the results show that the PSO-RVM model has the highest accuracy.The mean relative error,root mean square error,Thiel inequality coefficient and confidence interval were used to verify the compression coefficient prediction results.The sensitivity factor analysis of the peripheral displacement was carried out to explore the influence status of the influence factors,and the relationship degree of the nine influencing factors was ranked.2.Kernel function plays a key role in the prediction model,whose function is to map the relation of low-dimensional space to high-dimensional space,so as to facilitate the search of mapping relation.Gaussian kernel function,spline kernel function,cauchy kernel function and bubble kernel function were used to predict the height of water conducting fracture zone.The results show that the PSO-RVM model based on bubble kernel function is more accurate than the support vector machine model.Gaussian kernel function,cauchy kernel function and bubble kernel function were used to predict the frost heave rate of the microscopic characteristic structure of seasonal frozen soil,among which gaussian kernel function is the most accurate.The above proves that different kernel functions have different application scopes and different kernel functions have different accuracy of prediction models.3.For geotechnical problems with more information of influencing factors and difficult to find out the mapping relationship,principal component analysis can be used to optimize the PSO-RVM model.Principal component analysis reduces the dimensions of many influencing factors to a few linearly independent principal component variables by analyzing the influence status and contribution rates among the influencing factors,and then uses the principal component variables to make predictions,which further improves the accuracy and operating efficiency of the PSO-RVM prediction model.The PCA-PSO-RVM model is applied to the soft down hard on the formation of the surface subsidence.Firstly,PCA was used to reduce the dimensions of the 7 influencing factors to 4 principal components,and then PSO-RVM model was used for prediction analysis.Compared with its own PSO-RVM model and the six kinds of support vector machine models in the literature,the accuracy of PCA-PSO-RVM model is proved.The correlation index verifies that the predicted value and the measured value have high fitting degree.