A novel global prediction framework for multi-response models in reliability engineering using adaptive sampling and active subspace methods

Abstract

The computational cost associated with structural reliability analysis increases substantially when dealing with multiple response metrics and high-dimensional input spaces. To address this challenge, an innovative global prediction framework is proposed which leverages multi-output Gaussian process (MOGP) modeling. This framework reduces the computational burden for high-dimensional, multi-response systems by incorporating active subspace and adaptive sampling. The adaptive sampling technique strategically selects the most informative new data points for multi-response prediction by leveraging correlations between responses. Notably, the framework prevents premature termination in low-dimensional scenarios with unknown distributions. Additionally, a multi-response dependent active subspace dimension reduction method is employed to manage high-dimensional data, enhancing the stability of projected structural responses in the reduced-dimensional subspace. The effectiveness of the proposed framework is demonstrated through comprehensive case studies and comparative analyses with traditional approaches. The results demonstrate significant advantages in model dimension reduction, improved accuracy of structural prediction, and enhanced stability, making it well-suited for structural performance prediction.