Drive-by damage detection in railway bridges subject to operational vari- abilities using deep autoencoder

Resumo

One of the major challenges in the design and management of railway bridges is ensuring their safety and
structural integrity throughout their lifespan. This is due to the fact that the loss of structural function and eventual
failure of these structures have catastrophic consequences. Additionally, climate change projection which permeate
the present show a tendency towards an increase in the frequency and intensity of extreme events, accelerating the
deterioration process of railway infrastructure. In this context, there is a demand to define strategies in structural
health monitoring (SHM) in order to minimize disruptions in railway operations and maximize its profitability
through the safe operation of the system. There are two approaches to acquiring structural data for evaluating
integrity: the direct monitoring approach and the indirect monitoring approach, or drive-by. In direct monitoring
of railway bridges, sensors are installed directly on the bridge to capture responses caused by train excitations on
the structure. On the other hand, in the drive-by monitoring approach, sensors are installed on the train to capture
vibrational responses from the dynamic interaction of the train-bridge system during its passage. The advantages
of the indirect approach over the direct approach involve the ability to obtain spatial information along the entire
length of the bridge without the need to interrupt train operation, in addition to a substantial decrease in the cost
associated with monitoring an entire railway line. However, one of the main challenges of indirect monitoring is
dealing with the variability of operational and environmental conditions which affect monitoring data and result
in false negatives or false positives in damage identification process. This study focuses on the indirect structural
health monitoring of railway bridges using deep autoencoder model. For this purpose, numerical simulations of
the dynamic train-via-structure interaction (TTBI) are performed. These simulations aimed to gather acceleration
responses as trains crossed the target bridge under different levels of bridge foundation scour damage. Operational
variability involving train speed and track irregularity, and measurement data noise are considered to simulate
conditions closer to reality. The results show that the applied methodology is highly effective for detecting the
scour damage of railway bridge foundations.