In the constantly growing need for sustainable mobility and transportation, on-site inspections of existing reinforced concrete (RC) bridges are critical in ensuring the safety of such infrastructures. However, surveying RC bridges presents several challenges, such as the high costs and effort required by the surveyors, the subjectivity in assessing identified defects, and the possible lapses of attention when inspections are systematically repeated on different bridges. Hence, traditional methods of on-site inspection can be enhanced by leveraging digital innovations and by developing new instruments that support road management companies in ensuring the safety of the existing infrastructure. Among the new technologies, deep learning-based object detection systems provide promising and effective solutions. As such, this research proposes a new, simple, intuitive and efficient tool to support engineers and surveyors in assessing the health state of existing RC bridges. To this end, domain experts gathered and labelled a dataset of real images containing typical defects found in existing RC bridges. Consequently, an improved version of YOLO11, embedding attention mechanisms to allow the network to focus on the most relevant details in each image, was trained, tested, and validated on the provided dataset, showing an overall improvement of quantitative metrics such as precision and recall, while retaining enough computational efficiency to allow real-time implementation on constrained devices. Visual explanations achieved via the Eigen-CAM algorithm were also exploited to evaluate the reliability of the predictions. The model was finally embedded in an end-to-end tool offering a graphical user interface (GUI) to allow an effective interaction between the domain expert and the machine. Overall, the proposal revealed its potential to improve the effectiveness of the survey, lowering the burden on surveyors and engineers and providing a reliable method to improve the overall security in large RC bridges portfolios.

Anomaly detection is a topic of interest in several areas, ranging from Industry 4.0 to Energy Management, Smart Agriculture, Cybersecurity, and Bioinformatics. In a wide sense, detecting anomalies implies finding samples generated within a process that differs from its standard data generation mechanisms. Identifying these samples is extremely important for a variety of reasons, depending on the specific application and scenario, ranging from the minimization of production costs to maintaining the required safety standards. As such, the increasing availability of wide networks of sensors that yield large amounts of data characterizing the processes under observation allowed the large adoption of deep learning techniques, which proved worthy of attention due to their capability of identifying anomalies with large precision, accuracy and reproducibility. Consequently, there is an extensive need to consolidate research results to provide a common framework to understand the topic and ensure a common foundation to establish future research trends. To respond to this need, this work systematically reviews the state of the art of anomaly detection in Industry 4.0, evaluating gaps in the current knowledge and proposing future directions of interest. To pursue this objective, three main dimensions have been considered: the scenario where the anomaly detection methodologies were applied, the sensing equipment used to gather data characterizing the underlying process, and the algorithm employed to properly interpret the phenomena. The study was conducted following the PRISMA protocol, which allowed the identification of a relevant selection of papers by extracting a meaningful dataset of 78 papers of interest. The analysis highlighted the diffusion of autoencoders in several configurations and application scenarios, highlighting their effectiveness and flexibility for anomaly detection.