Evaluation of structural integrity of railway bridge using acceleration data and semi-supervised learning approach

Abstract

In order to investigate the structural integrity of a railway bridge, supervised and semi-supervised deep learning techniques incorporating wavelet transform (WT) were used where the measured acceleration data were transformed into images. Specifically, the structural integrity was evaluated both by acceleration datasets of a newly built bridge and by damaged datasets calculated from numerical analyses, and assumed damage in the structure was detected by supervised and semi-supervised learning methods. For the supervised learning, the well-known AlexNet and VGG16 convolutional neural network (CNN) models were employed, and during the one class (OC) classification of the semi-supervised learning approach, wherein only the label of the normal data was known a priori, a transfer learning method was utilized. It was found that the minimum value of damage with which novelty was detected was found to be at least 15% reduction of the stiffness in our case. It was also found that the cosine rather than Euclidean distance metric was more accurate during damage prediction. Although both methods showed very reliable prediction results, the semi-supervised learning method was shown to be more practical under the given field conditions.