A double-loop relevant vector machine-based system reliability analysis method with Meta-IS idea and active learning strategy

Abstract

This paper proposes a double-loop relevant vector machine (RVM) model for system reliability analysis. To reduce the computational load, an adaptive RVM is constructed, which is built by minority initial samples and K-folds clustering. The candidate sample pool constructed by this rough adaptive RVM model improves the computational efficiency. Based on the idea of active learning, another adaptive RVM is established. By combining two adaptive RVMs, the proposed model has the advantages of both active learning and importance sampling, which is called DLRVM. In this model, the failure probability is expressed as a product of the augmented failure probability and the correction factor. From the characteristics of RVM, this model under the Bayesian framework has significant generalization ability which avoids the limitations of many machine learning models. The accuracy and high efficiency are verified via four academic examples and an implicit engineering problem. The results also indicate that RVM is appropriate for system reliability analysis.