A COMBINED FNM METHOD FOR TIME-DEPENDENT RELIABILITY EVALUATION OF EXISTING BRIDGES

Based on three work states, in recognizing that reliability assessment of existing bridges by using finite element analysis and a complete set of random samples is very time-consuming, a combined method of finite element analysis, neural networks and Monte Carlo simulation (FNM) is developed in this study to analyze the time-dependent reliability of bridges. At first, analytic hierarchy process is used to calculate the weight of reliability factors. A fuzzy-clustering-based method is firstly used for removal of abnormal samples obtained from full-scale bridge tests, and statistical analysis is performed to derive the probability density functions of the normal samples. A total of n groups of samples are numerically generated following these prescribed probability density functions. Considering that computational efficiency of finite element method is relative low, among them, N of n groups of the samples are selected equably and their response values are obtained by repetitively inputted to a finite element code to compute structural responses. A neural network is then trained by using the N groups of samples and their corresponding structural responses. The structural responses values of the other remaining n-N groups of samples are obtained by feeding the n-N groups of samples of random variables to the trained neural network. According to the results from direct finite element (FE) Monte Carlo simulation as well as those from the trained neural network, the injury damage probability, the failure probability and the corresponding reliability indices are obtained. Working states current condition of existing bridges can be determined directly by the two indices as well. The developed FNM method is finally illustrated with a full-scale bridge. The example studies demonstrate a simple and effective quantitative analysis of the reliability of existing bridges. And reference basis for maintenance decision of the bridge is further provided.