Reliability and sensitivity analysis of civil and marine structures using machine-learning-assisted simulation

Abstract

Civil and marine structures are subjected to various deterioration mechanisms due to aggressive environmental effects or mechanical loads. In order to maintain an acceptable performance level of these structures, previous research have focused on developing methodologies to quantify their reliability and provide optimized management plans that can reduce the life-cycle cost and failure risk. However, the successful implementation of these methodologies is contingent upon the ability to consider various uncertainties associated with structural performance. These include uncertainties associated with environmental and human-induced stressors, as well as those affecting material and geometrical characterization as well as performance prediction models. Monte Carlo simulation (MCS) with a sufficient number of samples can provide accurate quantification of the structural performance under uncertainty. However, for complex problems that require detailed finite element (FE) modeling to predict the system performance, the computational cost can be very high. This problem can be addressed by using advanced sampling techniques that can provide an accurate estimation of the reliability with a significantly lower number of samples. Another approach is to use surrogate models to establish an accurate approximation of the complex system behavior. These models can provide statistically equivalent results of a complex simulation model, with no known closed-form solution, through a limited number of original model executions.

The proposed research focuses on developing probabilistic approaches for the performance assessment of civil and marine structures using machine-learning-assisted MCS. In this approach, machine learning is used to generate a surrogate model of the system response and is next integrated into the MCS to quantify the failure probability of the structure. Sensitivity analysis is conducted to identify the key contributing variables that significantly affect the system response. This process helps reduce the number of random variables associated with the problem resulting in a more efficient probabilistic simulation process. The developed approach was applied to solve two major research problems in civil and marine engineering: (a) reliability quantification of eccentrically loaded steel connections employing both welds and bolts for force transfer and (b) characterizing the crack propagation in stiffened panels and quantifying the reliability of ship hulls under realistic loading conditions.