Abstract
With the ever-increasing number of well-aged bridges carrying traffic loads beyond their intended design capacity, there is an urgency to find reliable and efficient means of monitoring structural safety and integrity. Among different attempts, vibration-based indirect damage identification systems have shown great promise in providing real-time information on the state of bridge damage. The fundamental principle in an indirect vibration-based damage identification system is to extract bridge damage signatures from on-board measurements, which also embody vibration signatures from the vehicle and road/rail profile and can be contaminated due to varying environmental and operational conditions. This study presents a numerical feasibility study of a novel data-driven damage detection system using train-borne signals while passing over a bridge with the speed of traffic. For this purpose, a deep Convolutional Neural Network is optimised, trained and tested to detect damage using a simulated acceleration response on a nominal RC4 power car passing over a 15 m simply supported reinforced concrete railway bridge. A 2D train–track interaction model is used to simulate train-borne acceleration signals. Bayesian Optimisation is used to optimise the architecture of the deep learning algorithm. The damage detection algorithm was tested on 18 damage scenarios (different severity levels and locations) and has shown great accuracy in detecting damage under varying speeds, rail irregularities and noise, hence provides promise in transforming the future of railway bridge damage identification systems.