Adaptive multilayer perceptron networks for detection of cracks in anisotropic laminated plates

Abstract

Listen

Translate article

In this study, an adaptive multiplayer perceptron (MLP) technique is proposed for the detection of cracks in anisotropic laminated plates. The displacement response on the surface of plate, excited by a time-harmonic line load, is used as the input of the MLP. The crack parameters that specify the location and size of the cracks in the anisotropic laminated plates are taken as the output of the MLP. The MLP model is first trained to establish the nonlinear relationship between the scattered surface displacement response and the corresponding location and size of the cracks. The scattered displacement responses required in training samples are calculated from the strip element method (SEM). To facilitate this training process, the correlation analysis for the outputs of neurons in the hidden layers of the MLP model is carried out to optimize the MLP architecture. A modified back-propagation learning algorithm with a dynamically adjusted learning rate and an additional jump factor is developed to speed up the convergence of the MLP model in the training process. The concept of orthogonal array is adopted to generate the representative combinations of the crack parameters, which significantly reduces the number of samples while maintaining the completeness of sample data. The well-trained model is then used to reconstruct the crack parameters by feeding in the measured displacement response on the plate surface. These reconstructed crack parameters are further examined by comparing their resulting displacement response from the SEM forward calculation with the measured displacement response. If the comparison is satisfactory, the reconstructed crack parameters would be considered to be true and the computation ends. Otherwise, the MLP model would go another round of re-training process until the satisfactory reconstruction is obtained. The proposed technique was verified numerically using an anisotropic laminated plate [C0/G+45/G–45] s with four types of horizontal cracks. The verification includes the detection for both the location and the size of cracks using the simulated response data with and without noise.