Abstract:
Automation of Structural Health Monitoring (SHM) through unsupervised Machine Learning
(ML) is a growing field of research. Most of these algorithms have been evaluated on scaled
and finite element (FE) models which do not accurately represent real-world conditions.
Autoencoder is an unsupervised machine learning algorithm that does not need a scaled or FE
model for SHM. It is applied to accelerometer data from the structure directly without the need
for any feature engineering. Recently, Autoencoders have been proposed for damage detection
using vibration data from real structures without live loads. While autoencoders successfully
detected damage in real structures, their use on in-service structures under ambient conditions,
i.e., under the effect of live load, remains unexplored. This study explores the practical
application of autoencoders for SHM of a through-arch bridge under the effect of live load
from motorcycles, pedestrians, and wind. Different damage cases were simulated by
producing knocking at different sensor locations. The autoencoders were successful in
fulfilling the primary objective of detecting damage and its location under simulated damage
conditions. In addition, they also successfully identified and located an unknown defect
already present in the bridge.
