Monitoring of a bridge experiencing scour using frequency domain decomposition mixed with DBSCAN algorithm: unsupervised modal analysis of output-only system

• Published in: Archives of Civil and Mechanical Engineering Vol. 25; no. 4; p. 193
• Main Authors: Belmokhtar, Mohamed, Schmidt, Franziska, Chevalier, Christophe, Ture Savadkoohi, Alireza
• Format: Journal Article
• Language: English
• Published: London Springer London 05.06.2025; Springer Nature B.V; Springer
• Subjects: Algorithms; Automation; Civil Engineering; Decomposition; Eigenvalues; Eigenvectors; Engineering; Engineering Sciences; Expected values; Fourier transforms; Frequency domain analysis; Frequency ranges; Hydraulics; Hypotheses; Machine learning; Mechanical Engineering; Modal analysis; Modal assurance criterion; Monitoring; Original Article; Parameter identification; Piers; Power spectral density; Resonant frequencies; Sensors; Signal processing; Structural Materials; Subspaces; Unsupervised learning; Water levels; Unsupervised machine learning; Signal filtering; Structural health monitoring; Frequency domain decomposition; Scour; Operational modal analysis; Structural Health Monitoring
• ISSN: 2083-3318; 1644-9665; 2083-3318
• DOI: 10.1007/s43452-025-01235-1

This study proposes a new method for monitoring bridges under scour considering an output-only system. From classical operational modal analysis (OMA), the revisited frequency domain decomposition (FDD) technique is applied to identify the modal parameters of the system. An eigenvalue decomposition (EVD) of the power spectral density matrix using the modal assurance criterion (MAC) value leads to defining a continuous base of eigenvectors through the studied frequency range. The method called EVD-MAC is then applied to a bridge subjected to varying environmental and traffic loads. In this real case study, the subspaces of EVD-MAC are used to represent the frequency behavior of the system with uncorrelated components. This makes it possible to identify clusters using unsupervised machine learning algorithms where each cluster corresponds to one presumed mode. The analysis is carried out with respect to specific environmental factors, including temperature and water level, which influence the system’s behavior. Scour assessments are conducted based on variations in natural frequencies observed in more than a year of monitoring data, demonstrating the utility and automation potential of this approach.