Vibration-based SHM of railway bridges using machine learning: The influence of temperature on the health prediction

Abstract

Civil engineering structures continuously undergo environmental conditions changes that can lead to temporary variations of their dynamic characteristics. Therefore, damage detection techniques have to be able to distinguish abnormal changes in the response due to damage from those normally related to environmental conditions variability. This paper addresses this issue by presenting a damage detection method that uses machine learning to detect and localize damage in railway bridges under varying environmental conditions (i.e. temperature). Results of the application to simulated data are shown with validation purposes. The first stage of the proposed algorithm consists in training a set of Artificial Neural Networks (ANNs) to predict deck accelerations during train passages assuming the bridge to be undamaged (or in a known state of preservation). In the second stage, the currently measured response is compared with that predicted by the trained ANNs. Since possible changes in the bridge state of preservation (damage) decrease the predictive accuracy of the ANNs, this comparison allows for the damage detection. During both stages, air temperature is given as input to the networks together with the train characteristics (i.e. speed and load per axle). The application results in the paper prove the ability of the algorithm to detect and localize damage. Furthermore, when the same procedure was applied neglecting the environmental factor, a noticeable decrease of the prediction power was met. This proves that changes in structural properties due to temperature variation can mask the damage occurrence and prevent its detection. The importance of accounting for environmental variations in damage detection is thus highlighted.