Drive-by scour damage detection in railway bridges using deep autoencoder and different sensor placement strategies

Abstract

Foundation scour is a critical phenomenon that may lead to the collapse of railway bridges. This issue is even more concerning in the current scenario where extreme weather events, such as floods, are becoming more severe and recurrent. Among different methodologies for assessing the structural integrity of railway bridges, vehicle-assisted monitoring has emerged as promising due to its low-cost and straightforward sensor installation compared to direct instrumentation of bridges. This paper provides a proof of concept of employing vehicle acceleration measurements from passing trains to detect the occurrence of bridge scour. To assess the effectiveness of accelerometer placement in data acquisition, vertical acceleration responses are collected from various positions throughout the vehicle and for different vehicles in the train, considering operational variabilities and measurement noise. A deep autoencoder model is used to process raw acceleration measurements collected during multiple train passages over a bridge affected by scour, where the scour damage is simulated as a local reduction in stiffness within a specific pier-foundation system. The difference between model-based and vehicle responses obtained from various observed events is the prediction error evaluated by the mean absolute error. The Kullback–Leibler divergence-based damage index is proposed to assess the number of vehicle-crossing events required to infer the damage. Finally, the approach’s accuracy is evaluated using Receiver Operating Characteristic curves. The results demonstrate that the applied methodology is highly effective in detecting both 5% and 10% levels of scour damage for sensors placed on the front and rear bogies of the first and last vehicles, without any prior data preprocessing.