Bridge health monitoring using deep learning based on sinusoidal gratings

This article proposes a novel vibration-based bridge health monitoring approach using convolutional neural networks (CNNs). Unlike other works that rely on precise feature extraction and suffer from environmental variability, the proposed methodology utilizes acceleration signals transformed into images of sinusoidal gratings via fast Fourier transform as an alternative solution. These images capture the natural frequencies of the bridge, which are sensitive indicators of damage. Three structures are analyzed to verify this novel approach: the Old ADA Bridge, the Z24 Bridge, and the Hell Bridge Test Arena. We employed ResNet50, DenseNet, and InceptionV3 CNN architectures to classify the structural condition. The models of the Old ADA Bridge, the Z24 Bridge, and the Hell Bridge Test Arena achieved F1 scores of 94.15, 98.39, and 99.97%, respectively, surpassing the performance of baseline techniques using acceleration signals as input by up to 13.46%. Furthermore, it was observed that the variability of the modal damage-sensitive features and the type of vibration of the system were not determining factors in the models’ performance. These results indicate the effectiveness of the proposed methodology in the early identification of potential structural problems on real-world bridges.