Small-scale damage detection of bridges using machine learning techniques and drive-by inspection methods

The drive-by inspection approach for bridge health monitoring has received a lot of interest recently due to its advantages in mobility, economy and efficiency. The feasibility of it has been demonstrated by many studies via numerical simulations, laboratory experiments, and even field tests, when there is a noticeable damage. In terms of minor damages, however, the dynamic features (e.g., frequencies and mode shapes) of the damaged bridge are highly similar to those of the healthy one, for which traditional drive-by methods are likely to perform poorly. Machine learning techniques, which utilize the entire time-domain responses and are sensitive to tiny signal changes, have the potential to identify small-scale damages and achieve higher detection accuracy. This paper compares the performance of different machine learning methods on the indirect framework and proposes a strong classification algorithm for damage identification of bridges. Laboratory experiments were conducted to build the dataset by employing a steel beam and a scale truck model. It presents an early attempt to experimentally validate the feasibility of the drive-by inspection method to identify small structural changes in the bridge.