TY - JOUR

T1 - A GRU-based ensemble learning method for time-variant uncertain structural response analysis

AU - Zhang, Kun

AU - Chen, Ning

AU - Liu, Jian

AU - Beer, Michael

N1 - Funding Information: The paper is supported by the National Natural Science Foundation of China (Grant No. 51905162 ), the Foundation for Innovative Research Groups of the National Natural Science Foundation of China (Grant No. 51621004 ) and the Fundamental Research Funds for the Central Universities ( 531107051148 ). The author would also like to thank reviewers for their valuable suggestions.

PY - 2022/3/1

Y1 - 2022/3/1

N2 - Owing to the influence of manufacturing and assembly errors, material performance degradation, external loads and unpredictability of the environment during service, structural response analysis should consider the time-invariant uncertainties and time-variant uncertainties simultaneously. In this paper, a mixed uncertainty model with random variable and stochastic process is adopted to handle this issue. A time-variant uncertain structural response analysis method is proposed based on recurrent neural network using gated recurrent units (GRU) combined with ensemble learning. In the proposed method, by performing Latin hypercube sampling (LHS) of random variables, multiple GRU networks can be trained to estimate the time-variant system response under fixed random variables. During the process of training GRU models, an active learning strategy is developed and applied to improve model accuracy and reduce training samples. On this basis, a set of augmented data is generated using the trained GRU models. Then the mapping relationship between random variables and structural responses through the Gaussian process (GP) regression is built accordingly. Eventually, the global surrogate model of time-variant uncertain structural response can be obtained by integrating the GRU networks and the GP models. Two numerical examples are used to demonstrate the effectiveness and accuracy of the proposed method. The results indicate that the proposed method can effectively calculate the expectation and standard deviation of the system response under the mixed uncertainty model with random variables and stochastic processes. In addition, it has higher computational efficiency under the premise of ensuring the calculation accuracy.

AB - Owing to the influence of manufacturing and assembly errors, material performance degradation, external loads and unpredictability of the environment during service, structural response analysis should consider the time-invariant uncertainties and time-variant uncertainties simultaneously. In this paper, a mixed uncertainty model with random variable and stochastic process is adopted to handle this issue. A time-variant uncertain structural response analysis method is proposed based on recurrent neural network using gated recurrent units (GRU) combined with ensemble learning. In the proposed method, by performing Latin hypercube sampling (LHS) of random variables, multiple GRU networks can be trained to estimate the time-variant system response under fixed random variables. During the process of training GRU models, an active learning strategy is developed and applied to improve model accuracy and reduce training samples. On this basis, a set of augmented data is generated using the trained GRU models. Then the mapping relationship between random variables and structural responses through the Gaussian process (GP) regression is built accordingly. Eventually, the global surrogate model of time-variant uncertain structural response can be obtained by integrating the GRU networks and the GP models. Two numerical examples are used to demonstrate the effectiveness and accuracy of the proposed method. The results indicate that the proposed method can effectively calculate the expectation and standard deviation of the system response under the mixed uncertainty model with random variables and stochastic processes. In addition, it has higher computational efficiency under the premise of ensuring the calculation accuracy.

KW - Active learning

KW - Ensemble learning

KW - Gaussian process

KW - GRU

KW - Time-variant response

UR - http://www.scopus.com/inward/record.url?scp=85123680384&partnerID=8YFLogxK

U2 - 10.1016/j.cma.2021.114516

DO - 10.1016/j.cma.2021.114516

M3 - Article

VL - 391

JO - Computer Methods in Applied Mechanics and Engineering

JF - Computer Methods in Applied Mechanics and Engineering

SN - 0045-7825

IS - 391

M1 - 114516

ER -