Scholarlink

<p>Time-variant reliability analysis (TRA) plays an important role for the lifecycle cost optimization, structural safety design, and preventive maintenance management of many complex engineering structures. For overcoming the huge computational burden of high-dimensional time-variant reliability analysis (HDTRA), this paper proposes a novel surrogate modeling method based on deep Bayesian active learning (DBAL). Firstly, a batch representative points selection-based active learning strategy is developed, selecting the most important points from predictive extreme points and predictive failure points in all time trajectories at each iteration. Then, Bayesian deep neural network (BDNN) is utilized to construct the surrogate model for high-dimensional time-variant performance function under the framework of active learning. Furthermore, a confidence-based convergence criterion is proposed to terminate the active training process timely. Finally, the extreme value-based Monte Carlo simulation (MCS) can be implemented to calculate the time-variant reliability by counting the frequency of failure time trajectories, and several numerical examples are analyzed for validating the effectiveness of proposed method.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Bayesian active learning for high-dimensional time-variant reliability analysis

  • 期刊论文

  • 发表日期：

    2026年5月21日

查看全文
  • Jingfei Liu,
  • Weiwei Chen,
  • Bingyu Ni,
  • Detao Wan,
  • Wei Feng,
  • Xinxing Ban

摘要

Time-variant reliability analysis (TRA) plays an important role for the lifecycle cost optimization, structural safety design, and preventive maintenance management of many complex engineering structures. For overcoming the huge computational burden of high-dimensional time-variant reliability analysis (HDTRA), this paper proposes a novel surrogate modeling method based on deep Bayesian active learning (DBAL). Firstly, a batch representative points selection-based active learning strategy is developed, selecting the most important points from predictive extreme points and predictive failure points in all time trajectories at each iteration. Then, Bayesian deep neural network (BDNN) is utilized to construct the surrogate model for high-dimensional time-variant performance function under the framework of active learning. Furthermore, a confidence-based convergence criterion is proposed to terminate the active training process timely. Finally, the extreme value-based Monte Carlo simulation (MCS) can be implemented to calculate the time-variant reliability by counting the frequency of failure time trajectories, and several numerical examples are analyzed for validating the effectiveness of proposed method.