Investigations on Plate Crack Damage Detection Using Convolutional Neural Networks

摘要

This paper presents a prediction model to detect the location and level of cracks in plates using a method based on the convolutional neural network (CNN), which can automatically learn damage characteristics without prior knowledge. A simply supported plate with cracks is established in Abaqus/CAE. Node accelerations at representative locations are adopted as the CNN input parameters. Effects of the initial learning rate, kernel size, and pool size on CNN performance are discussed so as to select the appropriate combination of hyperparameters to train the CNN. The results show that, when predicting the location and level of single- and multi-crack damage on noise-free and noisy data sets, the proposed method is more accurate than the multi-layer perceptron method and multi-layer perceptron by wavelet packet transformation. In addition, the proposed method produces a reasonable prediction of crack lengths.