A Novel Reliability Analysis Approach With Collaborative Active Learning Strategy-Based Augmented RBF Metamodel

Yanxu Wei,Guangchen Bai,Lu-Kai Song

doi:10.1109/access.2020.3035670

Yanxu Wei, Guangchen Bai + Show 1 more

Open Access

https://doi.org/10.1109/access.2020.3035670

Copy DOI

Abstract

Metamodels in lieu of time-demanding performance functions can accelerate the reliability analysis effectively. In this paper, we propose an efficient collaborative active learning strategy-based augmented radial basis function metamodel (CAL-ARBF), for reliability analysis with implicit and nonlinear performance functions. For generating the suitable samples, a CAL function is first designed to constrain the new samples being generated in sensitivity region, near limit state surface and keep certain distances mutually. Then by adjusting the adjustment coefficient of CAL function, the CAL-ARBF is mathematically modeled and the corresponding reliability analysis theory is developed. The effectiveness of the proposed approach is validated by four numerical samples, including global nonlinear problem, local nonlinear problem, nonlinear oscillator and truss structure. Through comparison of several state-of-the-art methods, the proposed CAL-ARBF is demonstrated to possess the computational advantages in efficiency and accuracy for reliability analysis.

Highlights

Various uncertainties widely exist in real structural engineering such as aerospace equipment, mechanical component and civil structure [1]-[4]
The coefficient of variation Cv indicates the uncertainty on failure probability and CV = (1 − pf ) pf nMCS ; Ns represents the number of samples; δ1 is evaluated by |pf-pcf|/pcf, where pcf the failure probability obtained by direct Monte Carlo Simulation [11] (MCS) method
Some conclusions are derived as follows: (1) The generated samples distribution reveals that the presented collaborative active learning (CAL) strategy can constrain new samples be generated in sensitivity region, near limit state surface and keep certain distances mutually

Summary

INTRODUCTION

Various uncertainties widely exist in real structural engineering such as aerospace equipment, mechanical component and civil structure [1]-[4]. As one of accurate interpolation methods, RBF can efficiently deal with high dimensional problems with exponentially converge rate [25][27] and the Gaussian function-based augmented RBF (ARBF) is one of the most widely used RBF, which can make full use of all the samples and possesses potentials to accomplish an accurate approximation [28] It had been widely applied in reliability evaluation and design fields [22], [25]-[28]. The generated samples concurrently conforming to the three constraints can acquire appropriate samples and accomplish high-accuracy and highefficiency metamodeling To meet these constraints and generate appropriate samples, it is urgently desired to develop a novel active learning function to achieve efficient and accurate reliability analysis.

CAL-ARBF method

A G a g

Method

CONCLUSIONS