A decision support system for damage identification in RC bridges using deep neural networks

The constant monitoring of critical infrastructures, such as bridges and viaducts, is mandatory to ensure transportation sustainability. The first and immediate approach to monitor to assess the health state of such infrastructures is represented by visual inspections, carried out within onsite surveys performed by domain experts. However, some human factors inherently bias these inspections, leading to costly, time-consuming procedures, which may also be affected by the subjectivity of the surveyors, potentially leading to biased outcomes. To improve traditional onsite inspections, recent technological advancements can be exploited to develop novel tools and instruments to support road management companies. Specifically, the deep learning paradigm can provide intuitive and reliable solutions for these tools. With this goal in mind, this work proposes an end-to-end framework that, starting from a dataset of real images gathered during onsite surveys, yields a graphical prediction to its users, effectively supporting decisions and automating the most time-consuming operations during surveys. Specifically, domain experts first labelled the dataset to identify seven types of damage commonly characterising reinforced concrete (RC) bridges. Afterwards, deep-learning-based object detectors were trained and embedded in an end-to-end tool, providing a simple yet effective interaction using a graphical user interface. The tool was tested on real use case scenarios, showing its effectiveness in defect detection and prediction. Overall, the tool had the potential to improve the reliability of the inspections, reducing the burden on surveyors while providing a reliable tool to aid in the assessment of large RC bridge portfolios.