Efficient adaptive Kriging-based reliability analysis combining new learning function and error-based stopping criterion
| Autor: | Qi Zhou, Jiaxiang Yi, Yuansheng Cheng, Jun Liu |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
Mathematical optimization
Control and Optimization Computer science 0211 other engineering and technologies Probability density function 02 engineering and technology Function (mathematics) Computational resource Computer Graphics and Computer-Aided Design Computer Science Applications 020303 mechanical engineering & transports 0203 mechanical engineering Control and Systems Engineering Bounding overwatch Approximation error Kriging Engineering design process Software Reliability (statistics) 021106 design practice & management |
| Zdroj: | Structural and Multidisciplinary Optimization. 62:2517-2536 |
| ISSN: | 1615-1488 1615-147X |
| Popis: | The Kriging-based reliability analysis is extensively adopted in engineering structural reliability analysis for its capacity to achieve accurate failure probability estimation with high efficiency. Generally, the Kriging-based reliability analysis is an active-learning process that mainly includes three aspects: (1) the determination of the design space; (2) the rule of choosing new samples, i.e., the learning function; and (3) the stopping criterion of the active-learning process. In this work, a new learning function and an error-based stopping criterion are proposed to enhance the efficiency of the active-learning Kriging-based reliability analysis. First, the reliability-based lower confidence bounding (RLCB) function is proposed to select the update points, which can balance the exploration and exploitation through the probability density-based weight. Second, an improved stopping criterion based on the relative error estimation of the failure probability is developed to avoid the pre-mature and late-mature of the active-learning Kriging-based reliability analysis method. Specifically, the samples that have large probabilities to change their safety statuses are identified. The estimated relative error caused by these samples is derived as the stopping criterion. To verify the performance of the proposed RLCB function and the error-based stopping criterion, four examples with different complexities are tested. Results show that the RLCB function is competitive compared with state-of-the-art learning functions, especially for highly non-linear problems. Meanwhile, the new stopping criterion reduces the computational resource of the active-learning process compared with the state-of-the-art stopping criteria. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|
Full text from SpringerLink