Ensemble wavelet-learning approach for predicting the effective mechanical properties of concrete composite materials

摘要

This paper proposes a high-accuracy and efficient ensemble wavelet-neural network method to predict the equivalent mechanical parameters of concrete composites. The doubly random uncertainties in structural heterogeneities and mechanical properties of concrete composites result in a challenging task to handle high-dimensional data properties, highly-complex mappings and huge computational cost for the repeated prediction of their mechanical parameters. The significant characteristics of this study are: (i) The random uncertainties both of structural heterogeneities and mechanical properties of concrete composites are modeled based on authors' previous work and Weibull probabilistic model, respectively. (ii) Asymptotic homogenization method (AHM) and the proposed background mesh technique are introduced to thoroughly extract the doubly random geometric and material characteristics of concrete composites for establishing concrete material databases. (iii) The wavelet transform is used to preprocess the high-dimensional data features of the material database, and the wavelet coefficients are used as the new input neurons of the artificial neural network (ANN) to establish the ensemble wavelet-neural network model. It should be noted that the wavelet-based learning strategy can not only extract important data features and resist noise from material database, but also achieve a great reduction in input data of neural networks from the entire material database and ensuring the successful training the neural networks. Finally, numerical experiments indicate that the proposed ensemble approach is a robust method for the high-accuracy and efficient prediction of equivalent mechanical properties of concrete composites.