Efficient Machine Learning Algorithm And Its Application In Structural Reliability Analysis

With the leap-forward development of design and manufacturing technology in the engineering field and with the requirements of high reliability and safety for Major equipment/engineering,the application of structural reliability analysis theory becomes very important.Uncertainty in design,manufacture and service affects the stability of structural performance,increases the workload of product maintenance and repair,and leads to the reduction of products life.Although the current development of structural reliability analysis methods is relatively mature,the accuracy and efficiency of existing simulation methods and analytical methods are relatively low for low failure probability or high computational cost structural reliability analysis problems.How to use efficient machine learning algorithms coupled with structural reliability analysis methods to solve more complex problems is the key to the development of current structural reliability theory and method.Therefore,to deal with the insufficiency of the current structural reliability analysis methods in solving specific structural reliability analysis problems,the following work is carried out in this thesis:(1)Considering the development of current structural reliability analysis methods,this thesis summarizes the development of simulation methods and analytical methods,and provides the principles and case studies of two commonly used Monte Carlo simulation and first-order reliability methods.The results revealed that the simulation methods are robust and accuracy,but requiring high computational cost.In contrast,the analytical methods are more efficient,but the accuracy of the result is affected by the performance function and random variables.(2)Considering the problem of low failure probability,the simulation method has high computational cost and low efficiency.This thesis proposed an enhanced Monte Carlo method based on support vector regression.Compared with the traditional enhanced Monte Carlo method,the proposed method provides a selection strategy for sample size and training interval and a more robust scaling formula.Case studies illustrate the effectiveness of the proposed method.The results show that the proposed method outperforms Monte Carlo simulation in efficiency and accuracy.(3)In view of the accurate results of finite element analysis in structural reliability analysis with the high computational cost,this thesis proposed an active kriging algorithm coupled with conjugate first-order reliability method which considering a resampling based active learning strategy to build the kriging surrogate model and using conjugate first-order reliability method based on the equivalent numerical differentiation strategy for structural reliability analysis.The effectiveness and efficiency are illustrated by actual engineering cases.Through the comparison of the results,it is found that the number of system responses and surrogate model responses called by the proposed method is lower.Meanwhile,the proposed method is more efficiency with premising of ensuring the accuracy and robustness.