A novel version of hierarchical genetic algorithm and its application for hyperparameters optimization in CNN models for structural delamination identification

Abstract

This paper proposes a novel approach for modelling the dynamic characteristics of composite laminated structures. The proposed numerical model employs the first-order shear deformation theory (FSDT) in combination with non-uniform rational B-splines (NURBS) to accurately capture the behavior of the laminated plate. The free vibration response of the composite plate is obtained by applying Hamilton's principle in conjunction with isogeometric analysis (IGA). The study investigates the effects of delamination location and size on the natural frequencies of the composite plate. The efficiency and precision in identifying the location and size of delamination are specified through a comparison of the computed results with those that exist in the published literature. To further enhance the accuracy of delamination prediction, this study employs two convolutional neural network (CNN) architectures. One CNN model is designed to predict the damage degree of delamination, while the other is a multi-output regression model used to predict the X and Y coordinates of the delamination center. Using frequency shifts as the input for the network, which is readily obtainable data. Compared to other machine learning models, this method offers significant advantages in delamination prediction, achieving up to 95% accuracy in evaluating the damage degree quantification. However, parameter settings have an effect on the prediction accuracy of delamination in laminated plates. In response to this issue, a novel delamination prediction model is developed in this paper. The hierarchical genetic algorithm is used to optimize the architecture of CNN. The CNN based on a hierarchical genetic algorithm (HGACNN) achieves a prediction accuracy improvement of 20–30% compared to conventional machine learning networks currently in use.