Ambient vibration measurement-aided multi-1D CNNs ensemble for damage localization framework

Abstract

Damage localization in civil infrastructure, such as large-scale reinforced concrete (RC) pedestrian bridges, is essential for conducting precise maintenance and avoiding catastrophic failures. In this study, multiple one-dimensional convolutional neural networks (1-D CNNs) are developed for automatically extracting implicit damage-sensitive features from the structural raw dynamic responses to localize damage in the pile foundations of pedestrian bridges considering uncertainties such as environmental and operational variations (EOVs) inherent in dynamic responses. For this purpose, transient dynamics numerical computation models are established to simulate the multi-point dynamic response of the structure under different typical damage scenarios, forming the baseline dataset. Then, on-site vibration tests are conducted on the structural prototype. Ambient vibrations of the real intact bridge are considered EOVs and integrated into the baseline dataset, forming the test dataset. Additionally, the intact structural dynamic response with measured EOVs replaces the simulated intact structural dynamic response in the baseline dataset to form a reference dataset. The network architectures based on one-dimensional convolutional layers proposed in this paper are trained on the baseline dataset and reference datasets to obtain baseline and reference models. Subsequently, model performance evaluation is conducted on the test dataset, and the results indicate a significant decrease in the performance of damage models based on a single deep learning when EOVs are present. However, integrating the baseline and reference models achieves zero false negative/positive predictions which is safety-oriented and an exemplary classification accuracy of up to 97.2 %.