Structural Damage Prediction of a Concrete and Steel Bridge Using Acceleration Signal Processing and Artificial Intelligence Algorithms

Nicole Xiomara Diaz; Aldahir William Ortega; Rick M. Delgadillo · Applied Mechanics and Materials (2026) · DOI: 10.4028/p-mvkm6i

Resumen

This study presents a methodology for predicting structural damage in a concrete and steel bridge using acceleration signal processing and artificial intelligence techniques. Structural vibrations were recorded continuously for 24 hours using five LARA (Low-cost Adaptable Reliable Anglemete) triaxial sensors located on the metal beams under the bridge, capturing data in three axes. The signals were normalized in order to be able to have a fairness of accelerations in the 3 axes and processed through a Convolutional Autoencoder (CAE), which achieved a signal reconstruction fidelity of 97.22%, enabling the generation of realistic synthetic data. To evaluate the separability between real and synthetic signals, a Domain-Adversarial Neural Network (DANN) was applied, successfully classifying both domains. Subsequently, K-Means clustering was performed on the compressed latent space, identifying three distinct structural states: healthy, transitional, and anomalous, with a silhouette coefficient of 0.8947. Notably, Sensors 2 and 4 were grouped into the anomalous cluster, indicating potential localized structural degradation. Finally, a Temporal Convolutional Network (TCN) was implemented to predict the future structural condition based on sequences of latent features. The model achieved an overall accuracy of 85.29% and an F1-score of 0.9954 for transitional states, demonstrating its effectiveness in anticipating early structural changes and reinforcing the potential of the proposed methodology for real-time, predictive bridge monitoring applications.

Instituciones: Universidad Peruana de Ciencias Aplicadas