Machine learning-based adaptive degradation model for RC beams. (15th February 2022)

• Record Type: Journal Article
• Authors: Wu, Zi-Nan; Han, Xiao-Lei; He, An; Cai, Yan-Fei; Ji, Jing
• Is Part Of: Engineering structures. Volume 253(2022)
• Publication Date: 2022-02-15
• Subjects: Degradation model -- Adaptive update -- Reinforced concrete beams -- Artificial neural network -- Experimental database -- OpenSees; Structural engineering -- Periodicals; Structural analysis (Engineering) -- Periodicals; Construction, Technique de la -- Périodiques; Génie parasismique -- Périodiques; Pression du vent -- Périodiques; Earthquake engineering; Structural engineering; Wind-pressure; Periodicals; 624.105
• DOI: 10.1016/j.engstruct.2021.113817
• Languages: English
• ISSNs: 0141-0296

Abstract

Highlights: Machine learning-based adaptive-updated degradation model (AUDM) is proposed. The constitutive parameters in AUDM can be adaptive updated at each analysis step. AUDM enhances the deterioration simulation and captures the variable shear-flexural interaction through adaptive update strategy. AUDM is implemented in OpenSees. The accuracy of AUDM over the existing degradation model is revealed based on experimental results. Abstract: Reliable seismic damage assessment of structural systems requires analytical models that are able to capture the cyclic deterioration of components. In this paper, an adaptive-updated degradation model (AUDM) for reinforced concrete (RC) beams is proposed based on the machine learning approach. The proposed model is capable of updating its constitutive parameters from the embedded artificial neural networks, based on the analysis results derived from the previous step. It offers the possibility for generating deterioration without constructing empirical equations, updating the hysteretic shape as the accumulation of damage, and accounting for the effect of variable shear-flexural interaction in nonlinear analysis. The details of the proposed model, including the backbone curve, the hysteretic behavior and the cyclic deterioration rule, were first described. An experimental database consisting of 100 cantilever rectangular RC beams under cyclic loading was collected from the existing literature to identify the constitutive parameters in the proposed model. Artificial neural networks for predicting the constitutive parameters in the proposed model were developed and trained based on the identified data, which was then embedded into the open-source computational platform OpenSees to implement AUDM. The accuracy of the proposed model over the existing commonly-used degradation model is revealed in OpenSees based on the test results of RC beams under cyclic loading. It was found that the proposed AUDM was shown to be superior to the existing degradation models in terms of the lateral bearing capacity, the dissipated energy, the cyclic deterioration and the hysteretic shapes under different failure modes, with the relative error of energy dissipation capacity in the range between −0.19 and 0.19 and the maximum root square error for strength deterioration less than 0.259.