期刊
RELIABILITY ENGINEERING & SYSTEM SAFETY
卷 219, 期 -, 页码 -出版社
ELSEVIER SCI LTD
DOI: 10.1016/j.ress.2021.108258
关键词
Structural system; Deep reinforcement learning; Deep neural network; Dominant failure mode
资金
- National Key R&D Program of China [2017YFC1500603]
- National Natural Science Foundation of China [51978216, 52192664, U1711265]
A deep reinforcement learning-based method is proposed for searching the dominant failure modes of a structural system. The method transforms the probability-based failure component selection process into a sequential decision process, and utilizes a deep neural network and a new reward function for learning the failure component selection policy.
The dominant failure modes (DFMs) of a structural system are significant for structural analysis and failure probability estimation. However, existing failure modes (FMs) searching methods often face problems of combinatorial explosion. To address this issue, a deep reinforcement learning (DRL)-based method is proposed for DFMs searching, which transforms the probability-based failure component selection process into a sequential decision process. First, the failure stages and the selected failure components of a structural system are transformed to be the states and actions in the DRL. Second, a deep neural network (DNN) is established to observe the failure stages and select failure components. Finally, a new reward function is designed to guide the network to learn the failure component selection policy. The proposed method was tested through a roof truss structure and a truss bridge structure. It was demonstrated that the trained DNN could learn to observe the failure stages and select the most critical components in a completely unknown testing set. High accuracy of the identified DFMs can be achieved. In comparison with the calculation results of Monte Carlo Simulation (MCS) and p-unzipping method, this proposed method shows significant computational efficiency advantages with high accuracy in dealing with combinatorial explosion.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据