2024-11-272024-11-272024Jiménez Alonso, J.F., Naranjo Pérez, J., Renedo, C.M.C., García-Palacios, J.H. y Muñoz Díaz, I. (2024). Integrating CWT and CNN for damage detection in high-speed railway bridges. En Bridge Maintenance, Safety, Management, Digitalization and Sustainability - Proceedings of the 12th International Conference on Bridge Maintenance, Safety and Management, IABMAS 2024 (2319-2326), CRC Press.978-103277040-6https://hdl.handle.net/11441/164983© 2024 The Author(s)High-speed railway bridges, reinforced by external post-tensioning systems, are prone to suffer from fatigue problems under corrosion environments. In order to perform the structural assessment of these structures, during their overall life-cycle, structural health monitoring strategies are currently being implemented. Among the different strategies, vibration-based ones are normally employed in civil engineering applications due to its ability to global damage detection. According to these strategies, the measured dynamic response of the structure is compared between two states (the current and the original one) in order to assess the performance level. Nowadays, deep learning, a machine-learning computational tool, has been used to assist in this comparison task. An artificial neural network is usually trained for this purpose. For this training process, two extreme states are normally considered: (i) undamaged state; and (ii) damaged state. As it is not normally possible to obtain data of the damaged state of the structural system, an updated physics-based model is usually considered for this purpose. Therefore, the dynamic response of the damaged structure is simulated based on this updated model. In this manuscript, the performance of a classifier is assessed when it is implemented for the damage detection of a high-speed railway bridge. As an original (undamaged) state, an updated Bayesian finite-element-model of the structure is considered. Different level of uncertainty has been considered (variation range of the updating parameters) for this updated model. As damaged state, different levels of numerical damage have been generated. The dynamic response of the structure (vertical accelerations) has been compared for the damage detection. For the feature extraction, the continuous wavelet transform has been considered. As deep learning technique, a pre-trained convolutional neuronal network has been taken into account. Thus, the performance of the proposed classifier has been assessed, for this particular case-study, via the determination of both the confusion matrix and the accuracy ratio. Finally, the variation of the accuracy ratio in terms of the uncertainty of the updated model and the damage level has been studied in detail.application/pdf8 p.engAttribution-NonCommercial-NoDerivatives 4.0 Internationalhttp://creativecommons.org/licenses/by-nc-nd/4.0/info:eu-repo/semantics/conferenceObjectinfo:eu-repo/semantics/openAccess10.1201/9781003483755-275