Parallel active learning XGBoost for structural reliability analysis with application to an onshore wind turbine tower
提出一种名为PALX的并行主动学习XGBoost方法，通过交叉验证量化预测不确定性，并引入新的学习函数和并行采样策略，在保证精度的同时大幅降低计算成本，适用于复杂工程结构可靠性分析。
Active learning methods have emerged as a powerful tool in structural reliability analysis. However, conventional approaches may still fall short in terms of efficiency, accuracy, and applicability when addressing complex real-world problems. To this end, this study develops a novel active learning method called ‘parallel active learning XGBoost’ (PALX). In this method, the XGBoost model is employed as a surrogate for the true performance function instead of the commonly used Kriging model, with prediction uncertainty quantified through cross-validation. By assuming that the resulting predictions follow a Gaussian process, a convenient failure probability estimator and a robust stopping criterion are introduced, which are adapted from a well-established Bayesian active learning method. The failure probability estimator and stopping criterion are numerically solved using the sequential variance-amplified importance sampling. Furthermore, a new learning function, termed ‘prediction variance-weighted epistemic uncertainty contribution’, is proposed for identifying the best next evaluation point. To enable parallel computing, a multi-point selection method called ‘lower confidence bound believer’ is developed. The effectiveness of PALX is demonstrated through three numerical examples and a practical engineering problem involving an onshore wind turbine tower. It is shown that PALX can significantly reduce computational costs without compromising accuracy, demonstrating its potential for real-world engineering challenges.