Multi-point Bayesian active learning reliability analysis

Abstract

This manuscript presents a novel Bayesian active learning reliability method integrating both Bayesian failure probability estimation and Bayesian decision-theoretic multi-point enrichment process. First, an epistemic uncertainty measure called integrated margin probability (IMP) is proposed as an upper bound for the mean absolute deviation of failure probability estimated by Kriging. Then, adhering to the Bayesian decision theory, a look-ahead learning function called multi-point stepwise margin reduction (MSMR) is defined to quantify the possible reduction of IMP brought by adding a batch of new samples in expectation. The cost-effective implementation of MSMR-based multi-point enrichment process is conducted by three key workarounds: (a) Thanks to analytical tractability of the inner integral, the MSMR reduces to a single integral. (b) The remaining single integral in the MSMR is numerically computed with the rational truncation of the quadrature set. (c) A heuristic treatment of maximizing the MSMR is devised to fastly select a batch of best next points per iteration, where the prescribed scheme or adaptive scheme is used to specify the batch size. The proposed method is tested on two benchmark examples and two dynamic reliability problems. The results indicate that the adaptive scheme in the MSMR gains a good balance between the computing resource consumption and the overall computational time. Then, the MSMR fairly outperforms those existing leaning functions and parallelization strategies in terms of the accuracy of failure probability estimate, the number of iterations, as well as the number of performance function evaluations, especially in complex dynamic reliability problems.