Efficient framework for structural reliability analysis based on adaptive ensemble learning paired with subset simulation

This study develops a novel, accurate and efficient framework for the reliability analysis of multi-component civil structures, which involves many random variables and repeated time-consuming structural analysis processes. The simple yet effective idea is to construct a binary classification surrogate model to detect the structure’s condition (failure/safe) given structural parameters, external loads, and predefined safety thresholds. However, building a surrogate model that can accurately detect the structure condition is challenging because of the heavy imbalance nature of the considered problem, i.e., only a very small portion of the data corresponds to the failure condition. To overcome this problem, an ensemble learning model which stacks six different classification machine/deep learning models is engineered. The ensemble model can improve the classification performance by leveraging the model diversity rather than manually tuning a set of hyperparameters. Besides, a subset simulation scheme is leveraged to address the scarcity of relevant samples, i.e., providing more training samples from the potential failure region. In addition, the sampling weighting technique is adopted to assign higher weights for samples corresponding to the class with smaller probability, allowing the training process to achieve a faster convergence rate and higher final performance. The applicability and efficiency of the proposed approach are successfully demonstrated through three case studies with different complexity and dimensionality, showing that the proposed approach can provide accurate reliability results with up to two orders of magnitudes less computational costs.