Deep learning-based indirect bridge damage identification system

Abstract

With the growing number of well-aged bridges and the urgency in developing reliable, (pseudo-) real-time monitoring of structural safety and integrity, there is a worldwide and widespread campaign toward transforming structural health monitoring practice. Among these attempts, the application of data-driven approaches in developing damage identification techniques has received particular attention in recent years. Given the growing volume of structural health monitoring data, the power of data-driven approaches has been further exploited. These efforts have been predominantly focused on building and training algorithms using direct measurements from bridges. Although recent years have seen transformative technologies in producing cheap and wireless sensors, network-wide bridge instrumentation is logistically difficult and expensive. This has led to a new group of structural health monitoring systems entitled indirect or drive-by approaches. In drive-by systems, measurements from an instrumented vehicle are used to extract structural damage signatures. In other words, in these systems, the instrumented vehicle acts as both actuator and receiver while passing over a bridge. The main challenge in deploying drive-by approaches for damage identification purposes is that the signals collected on drive-by vehicles also embody signatures from the vehicle, road/rail profile and are easily contaminated by environmental and operational conditions. Furthermore, the majority of current drive-by damage identification systems rely on prior knowledge of vehicle or bridge dynamic characteristics which has led to limited application of the concept in practice so far. To address these challenges, this study employs a powerful class of deep learning algorithm to develop a damage identification system using measurements on an instrumented travelling train. The proposed algorithm is capable of automatically extracting damage signatures from train-borne measurements only. To demonstrate the algorithm’s capability, the method is applied to measurements collected on a model instrumented train travelling on a simply supported model steel bridge. For this purpose, a deep convolutional neural network is built, optimised, trained and tested to detect damage using acceleration signals collected on the instrumented train only. The hyperparameters of the algorithm are optimised using the Bayesian optimisation technique. The accuracy of the algorithm is experimentally tested for four positive damage scenarios (combination of two different locations and intensity) and three different travelling speeds. This is the first demonstration of the data-driven drive-by damage identification system under scaled operational environment conditions. The performance of the proposed method is discussed under different travelling speeds and different damage states. The result shows that the proposed method can accurately and automatically detect and classify damage under varying speed, rail irregularities and operational noise using train-borne measurements only and offers a great promise in transforming the future of bridge damage identification system.