Vibration-based SHM of railway steel arch bridge with orbit-shaped image and wavelet-integrated CNN classification

Abstract

The study presents the data-driven applications of continuous wavelet transforms (CWT), orbit-shaped analysis and convolutional neural networks (CNN) using GoogLeNet for Dębica railway bridge health monitoring in Poland. Training and validation data sets are the dynamic behavior of the bridge deck recorded through an IEPE vibration sensor with a sampling frequency of 128 Hz from vibration-based structural health monitoring (SHM) system over a nine-month period from December 2019 to September 2020. Utilizing Morse, Morlet, and Bump wavelet, the vibration signal scalogram images are produced in the frequency–time domain as the input for CNN classification models, while the output is to predict health states based on the experimental tension force of eight hangers using label thresholds developed by calibrated finite element model. Moreover, the vibration-based orbit-shaped image patterns, acquired through a bidirectional sensor on each hanger are processed with CNN classification models for automated hanger health diagnostic. The accuracies and weighted F1-scores of the wavelet-assisted and orbit-shaped CNN models are more than 83% on imbalanced validation images. The results show that the proposed CNN approaches can classify the potential structural problems.