Active learning surrogates in system reliability analysis: a review

Purpose This review aims to comprehensively summarize recent progress in surrogate modeling for reliability analysis of complex mechanical systems, particularly aeroengine mechanisms operating under harsh, multi-physics conditions. It focuses on how active learning strategies enhance modeling efficiency and accuracy, addressing challenges such as small failure probabilities, multidisciplinary coupling, high-dimensional inputs and time-varying uncertainties. Design/methodology/approach The study reviews regression-based surrogates (e.g. Gaussian process regression, artificial neural networks, random forest regression) and interpolation-based surrogates (e.g. radial basis function, Kriging), outlining their strengths and limitations. It examines active learning methods – such as expected improvement (EI), upper confidence bound (UCB) and reliability-based expected improvement function (REIF) – and their integration with variance reduction techniques like subset simulation and importance sampling, supported by practical aeroengine case studies. Findings Active learning surrogate modeling significantly improves computational efficiency and predictive accuracy for high-dimensional and multidisciplinary reliability problems. Coupled with advanced variance reduction strategies, it enables accurate estimation of small failure probabilities with reduced simulation costs. Aeroengine applications demonstrate substantial acceleration of design cycles and more effective resource allocation. Originality/value This review offers the first structured synthesis of surrogate modeling and active learning in aeroengine reliability analysis, linking methodological advances with real-world engineering needs. It identifies key research gaps and future directions, particularly the integration of real-time monitoring, multi-fidelity modeling and digital twins for adaptive, online reliability assessment, providing valuable guidance for both researchers and engineers.