Stochastic nonlocal damage analysis by a machine learning approach

Abstract

A machine learning aided stochastic nonlocal damage analysis framework is proposed for quasi-brittle materials. The uncertain system parameters, including the material properties and loading actions, have been incorporated and analysed within a unified safety assessment framework against various working conditions. A three-dimensional integral-type nonlocal damage model through finite element method (FEM) has been adopted. For the purpose of investigating the probabilistic damage analysis problems, a freshly established machine learning approach, namely the capped-extended-support vector regression method (C-X-SVR), is proposed to eliminate the influences of random outliers in the first step, then establish the relationship between the uncertain systemic inputs and structural responses. Such that the training robustness and computational adaptability of the proposed regression model can be reinforced. Moreover, the proposed approach is competent of efficiently predicting the statistical information (i.e., means, standard deviations, probability density functions and cumulative density functions) of structural behaviours under continuous information update of the uncertain working condition from mercurial environment. One real-life experimental validation and two numerical investigations are implemented to further verify the effectiveness and efficiency of the uncertainty quantification framework against probabilistic damage analysis.