Detecting Structural Changes on a Long-Span Steel Bridge

The dependence of modal parameters on stiffness and mass properties makes the Operational Modal Analysis (OMA) an interesting strategy of Structural Health Monitoring (SHM). However, crucial aspects of OMA-based SHM are related to the low sensitivity of modal parameters to structural changes and the need for a correct minimization of environmental and operational variability (EOV) effects on modal parameters, as those effects may mask changes instead due to structural damages. In the last few years, the field of Artificial Intelligence has made major progress and Deep Learning (DL) tools have become more common, especially in the SHM of bridges. Among the advantages of DL procedures, one should mention: (a) the use of short datasets; (b) the avoidance of any system identification and (c) the possibility of implicitly handling the EOV within the input data. Within this context, the paper presents a comparative study between OMA-and DL-based approach to detect structural changes. In particular, the DL procedure is based on the use of a sparse autoencoder (SAE) network to reconstruct the measurement data collected during continuous dynamic monitoring. The SAE network turns out to provide reconstruction errors that increase as soon as the monitored system departs from its healthy condition. The application of the detection procedures is exemplified using data collected on the benchmark KW51 bridge (a steel bowstring railway bridge in Leuven, Belgium): both the investigated strategies allow detecting the structural changes due to a 4-months retrofitting performed on the bridge, with the SAE being capable of implicitly accounting for the EOV.