A Combined Logistic Regression and Learning Kriging Method for Reliability-based Design Optimization

摘要

Reliability-based design optimization (RBDO) is an effective methodology to eliminate potential failures in the design of products. However, the heavy computations of expensive models are great challenges for existing RBDO methods. To address this issue, a combined Logistic Regression and Learning Kriging (LR-LK) method is proposed in this work. LR-LK constructs the reliability Kriging model (RKM) between design variables and their corresponding reliabilities in the highly probable region of the global optima. Logistic regression is integrated into the RKM to highly improve the approximation accuracy. To approximate the objective function, the objective Kriging model (OKM) is built at the feasible region recognized by the RKM. This way, both the RKM and the OKM are sufficiently accurate only at the active constraints around the global optima, thereby ensuring high efficiency. Further optimization can be conducted on the RKM and OKM without any call to the original expensive models.