Deep learning-based condition assessment for bridge elastomeric bearings

Mida Cui

doi:10.1007/s13349-021-00540-6

Deep learning-based condition assessment for bridge elastomeric bearings

2022-01-06

18

PRE

AI

DOI：10.1007/s13349-021-00540-6

Original

Original request for help

Save

Abstract

En 中文

Although computer vision-based methods are emerging due to the advances in mobile imaging technologies, the current approach to assessing elastomeric-bearing conditions heavily relies on human visual inspection. Furthermore, few computer-vision efforts are found for automatic condition assessment for the bearings, partially due to the challenge in acquiring a specific and large-scale database in practice. Through developing a unique imagery database with engineering-meaningful condition labels, this paper first benchmarks the performance by utilizing the traditional computer vision framework using scale-invariant feature transform (SIFT) features and support vector machine (SVM) classifier. By adopting three different kinds of convolutional neural networks (CNN) architectures (AlexNet, VGG-11, and ResNet-18), this paper contributes by evaluating different CNN architectures on small size elastomeric bearing dataset. Also, transfer learning (TL) techniques are applied to improve the performance of CNN models. Furthermore, different training strategies including using pretrained weights as fixed feature extractors and fully finetune the network architecture are evaluated in this paper. The authors conclude that the CNN models equipped with fully fine-tuned TL techniques possess much satisfactory performance and hold promising to be used for real-world applications for automating elastomeric-bearing condition assessment.

Keywords:

Elastomeric bearings

Damage detection

Deep learning

Convolutional neural network

Small data

Deep learning-based condition assessment for bridge elastomeric bearings

Abstract

Journal

Organization