Damage detection in railway bridges using Machine Learning: Application to a historic structure

This paper presents a method that uses machine learning to detect and localize damage in railway bridges. Results of the method application to a historical bridge are presented and used to validate the proposed algorithm. For the application of this technique, both air temperature and deck accelerations data, measured under railway traffic at several locations on the bridge, are needed. The method consists of four stages: (1) collection of such data in both reference condition (i.e. when the state of preservation of the structure is known) and current one; (2) pre-processing of acceleration time histories aimed at extracting characteristics of the crossing train (i.e. running direction, speed and number of axles); (3) training of Artificial Neural Networks and Gaussian Processes using data collected in reference condition and (4) health classification of the bridge in current condition through the comparison between predicted and measured responses. During stage 3, a set of neural networks is trained to predict deck accelerations under every environmental and operational condition (i.e. air temperature and crossing vehicle characteristics, respectively) assuming the reference state of preservation. Then, in stage 4, the current response is compared with accelerations predicted under current environmental and operational conditions. Changes in the behavior of the structure due to damage are thus detected as a discrepancy between predicted and measured responses. The application of the proposed technique to data collected on San Michele Bridge (1889), in Northern Italy, has shown good agreement with results from previous studies based on mode shape variation. This shows the potential and confirms the possibility of applying the proposed technique to real bridges. This method can thus be used to detect anomalous responses that can be flagged as possible damage as well as give an indication of the location of the decayed structural region.