Probabilistic machine learning approach for structural reliability analysis

In recent years, the surrogate modeling technique has been increasingly employed for the failure probability estimation of civil structures. The surrogate modeling technique acts as a black box that predicts the performance of expensive-to-evaluate problems. However, the uncertainty of failure probability estimation due to the discrepancy between the surrogate model and physical model has not been well studied. This study introduces a new surrogate modeling technique for failure probability estimation of civil structures that utilizes a Bayesian neural network. Specifically, a fully connected neural network is constructed to approximate the limit state function. The hyperparameters in the neural network are obtained within the Bayesian paradigm, and the probability distributions are estimated through Laplace approximation. When the Bayesian neural network method is used for reliability analysis, the failure curve and the corresponding uncertainty can be estimated. Numerical studies on 2d truss, a simple support beam under moving load, and dynamic analysis of subway station considering the soil–structure interaction are conducted to validate the efficiency of the approach. Results with Monte Carlo simulation and subset simulation are also presented. The results demonstrate the proposed method's potential for improving prediction accuracy by factoring in the uncertainty embedded in the surrogate model. Additionally, it exhibits higher levels of efficiency than traditional methods.