Research On Structural Reliability Analysis Based On Active Learning Surrogate Model

In engineering areas,various parameters and external environmental loads of structures have uncertainties.These uncertainties will affect the results during structure design process.Thus it cannot meet the actual requirements when we only use deterministic design criteria.Therefore,it is important to carry out basic theoretical research on structural reliability and its engineering applications.Traditional methods of reliability analysis,such as first order reliability method(FORM)and the second order reliability method(SORM),often have large prediction errors for a strongly nonlinear structural system.While Monte Carlo simulation method needs a large number of random sample points,which will cause much computation time in practical engineering applications.Therefore,these reasons result in great limitations to the engineering of the above methods.Many researchers proposed a sequential sampling method based on the surrogate model,which can greatly improve the efficiency of establishing a approximate model.However,the existing researches of the active learning method for reliability analysis most focus on the Kriging model,but there are few investigations for other types of approximate models,such as radial basis function(RBF),artificial neural network(ANN),etc.Therefore,in this paper,we study the active learning method of reliability and sensitivity analysis based on the RBF model,and apply it to engineering structures.The main research contents and innovations of this thesis are as follows.(1)Based on the RBF surrogate model,this paper studies the failure probability of structures using the active learning method with single kernel function and multiple kernel functions.For the single kernel function method,the cross-validation method is used to obtain the local uncertainty of the predicted sample points,and the multiple kernel functions method constructs the local uncertainty according to the difference between the prediction results of different kernel functions.Moreover,the proposed active learning functions consider both the distances between the samples and the limit state function and the distances between the training sample points.And a new convergence criterion for active learning process is presented in this paper,which is based on the relationship between the mean and the standard deviation of the final several predicted failure probabilities.(2)For the proposed active learning algorithm,the influences of the initial sampling methods(including Latin hypercube sampling and probability density sampling)and the number of sample points on the reliability analysis efficiency and accuracy are studied based on some classical examples.In addition,the effects of the parameters of the active learning function on the efficiency and accuracy are also investigated,including the uncertainty parameter ? of the surrogate model,the distance parameter ? between the training sample points,the number and type of kernel functions in the multiple-kernel-function method,and the threshold value of the stopping criterion ?.(3)An active learning method based on two-stage surrogate model is proposed to solve the small failure probability and multiple failure areas problem.The establishment of the approximate model is divided into two stages.Firstly,based on the extended joint probability density function,the first stage of active learning process is carried out,and the corresponding RBF model is established.Then combined with the hierarchical clustering method,the surrogate most probable points of each failure domain can be obtained,which can avoid the difficulties to obtain the multiple most probable points by using the traditional method.Finally,based on the important sampling method and the approximate most probable points,the second stage of the active learning process is conducted to establish a more accurate RBF model.Therefore,it is called the two-stage approximate model active learning method.(4)An active learning algorithm is also proposed to obtain the variance-based global sensitivity index,and the coefficient of each sensitivity indicators is used to construct the stopping criterion.Moreover,considering the uncertainty of the distribution parameters,we provide a two-layer surrogate model combined with active learning strategy to carry out the parametric global sensitivity analysis,which can highly improve the efficiency.(5)The reliability and sensitivity analysis method proposed in this paper is applied in the engineering areas based on the vibration characteristics of CNC turret punch and orbital sander.We use ADAMS and other software to realize the parameterized processing of CNC turret punch,sanding machine dynamics model and joint surface stiffness.Moreover,combined with MATLAB and Isight,it is easy to construct an active learning RBF surrogate model that can accurately predict the vibration response of CNC turret punch.Thus,the failure probability and global sensitivity of the CNC turret punching vibration under different working conditions are obtained.The parametric global sensitivity analysis of the CNC turret punch and sanding machine are also conducted when considering the uncertainty of the distribution parameters,which are useful for the structure optimization and the vibration reduction.