Incremental learning of parameter spaces in machine-learning based reliability analysis

Significant advances in machine-learning based reliability analysis occurred recently. These allowed it to be performed with high effectiveness. However, in most applications the problem of reliability is treated within a closed setup, and if a change in the problem is needed a posteriori, the reliability needs to be re-assessed, which often results in an inefficient usage of resources.In this context, the present work argues that established reliability knowledge can inform the assessment of a similar problem under parameter changes. It uses incremental learning in an augmented space to solve a reliability analysis with dependence on parameter variations. It is shown that only the points that swap their classification are of interest to reassess reliability, which has large synergy with machine learning and classification.A learning approach that uses this synergy is proposed to search for the points that are under a class change. It is tested in four examples and uses adaptive kriging. The results show that with only few additional evaluations of the true function it is possible to accurately (at <1% loss in accuracy) assess the reliability for a relatively complex problem experiencing changes in its parameters.This approach opens a new frontier on the topic and is aligned with the demand for dynamic assessment techniques in engineering decision-making processes.