Data-Driven Performance Evaluation of A Concrete Slab Bridge Using Machine Learning

Abstract

Field load testing of bridges is often used as a reliable method for evaluating bridge performance. One of the downsides of field testing is that it usually requires a heavy instrumentation setup. This paper investigates the efficacy of using an artificial neural network (ANN) to predict a concrete slab bridge response and potentially reduce the number of instruments needed for field testing. The diagnostic test results from a single-span bridge are incorporated as the input dataset. Test truck location from the edge of the bridge, loading on the truck axles, and distance covered along the bridge by each axle are set as the input parameters, while the measured strains from 13 strain gauges are set as the target output. The neural network is then trained, tested, and validated, showing a good correlation with an acceptable average error percentage. Parametric studies are conducted next using the developed neural network to examine the influence of the number of strain gauges on the results. The network involving only three strain gauges with peak response shows a nearly similar correlation as the network with all 13 strain gauges. The developed neural networks are then used to predict the bridge response compared with the same bridge's proof load test results. The networks are found to predict the bridge response with high accuracy within a range of − 13.7 to + 18.6%, even with the reduced number of sensors. The results from this study demonstrate the potential of using ANNs to predict the bridge response and to optimize the sensor plans for on-site bridge load testing.