An active learning method combining Reliability Analysis Transformer and attention-based sampling for high-dimensional reliability analysis

Reliability analysis faces the challenge of accurately estimating failure probabilities while minimizing the number of time-consuming model evaluations. Surrogate models can significantly alleviate the computational burden. However, the increasing dimensionality of modern engineering systems poses substantial challenges to the performance of traditional surrogate models. To address these issues, this paper introduces a novel cross-domain paradigm for high-dimensional reliability analysis, where the dimensionality of random variables is conceptually mapped to the sequence length in natural language processing (NLP). Building on this mapping, a Reliability Analysis Transformer (RAT) is designed to capture complex dependencies between random variables and system responses. Furthermore, based on the principle that similar samples contribute less new information, an attention-based sampling strategy is proposed to measure sample similarity and iteratively select the least similar samples to existing ones from the candidate dataset, thereby improving the accuracy of the RAT. The effectiveness of the proposed method is validated through four high-dimensional numerical examples and a liquid rocket engine case study. Results show that the proposed method achieves high accuracy and efficiency in handling high-dimensional and large-scale engineering problems.