Machine-learning-based bridge damage detection using train-borne measurements

This study presents a novel machine-learning-based approach for damage detection using train-borne measurements, under operational conditions (speed > 50 km/h, rail irregularities and noise). To this end, an optimised two-dimensional convolutional neural network with network-in-network architecture is built, trained and tested to detect damage of various severity levels and locations in a bridge, using train-borne measurements only. As an input, cross-correlation of signals from two train bogies is used as a damage-sensitive feature for the first time. The proposed method in this study is applied to a cohort of simulated acceleration measurements on a nominal RC4 power car passing over a 25 m simply supported reinforced concrete bridge. The presented method shows great accuracy in detecting damage under operational conditions. The sensitivity and robustness of the approach are tested and validated for 18 damage severity and location scenarios and 100 random vehicle speeds, ranging between 70 and 130 km/h. This is of particular value, as speed defines the length of the train-borne signal while passing over the bridge and hence the amount of information manifested in a single passing. The results demonstrate the feasibility of the approach for data-driven damage detection using measurement on an instrumented passing train.