PSI - Issue 64

Francesco Pentassuglia et al. / Procedia Structural Integrity 64 (2024) 254–261 F. Pentassuglia et al./ Structural Integrity Procedia 00 (2019) 000 – 000

257

4

However, its efficiency and applicability to other concrete bridge typologies, such as beam and slab bridges, cast-in situ slabs, continuous box-girders, that are commonly used worldwide still remains unknown. To make the methodology more widely applicable there is a need to test and validate it across a broader spectrum of bridge types. Moreover, this paper suggests further validation of the machine learning k-NN (k-Nearest Neighbor) methodology through comparing its predictions for the damage sustained by the bridge asset with those obtained via utilising Finite Element (FE) model updating. FE model updating involves adjusting the parameters of a numerical model to better match the behaviour of the actual structure, thereby improving the accuracy of the model predictions. By incorporating FE model updating as a validation step, the study aims to enhance the reliability and robustness of the k-NN methodology for bridge damage identification, ensuring its effectiveness across different bridge types and scenarios. 2. Enhancing bridge damage characterization through advanced methodologies The intended progress compared to the past study by Kazantzi et al. (2024a, 2024b) includes two key objectives: • Application to various bridge typologies: the aim is to extend the methodology to different types of bridges, in particular continuous box girders, cast-in situ slabs and precast I-beams, to assess its effectiveness in producing accurate results. By repeating and testing the methodology on different bridge typologies, the study seeks to determine its applicability and reliability in identifying damage states in concrete bridges beyond the specific focus on balanced cantilever bridges of the previous study. • Validation through comparison: the methodology will be validated by comparing its predictions with those generated through model updating. This comparison serves as a critical step in ensuring the accuracy and robustness of the methodology. If the comparison results are not satisfactory, the study commits to refining the existing method by exploring more advanced machine learning techniques or expanding the training set to enhance the methodology's predictive capabilities. The initial step in the process (Figure 1), that involves the asset modelling, is fundamental in the methodology for identifying the damage state of a bridge. This step focuses on examining each bridge typology (continuous box-girders, cast-in situ slabs, precast I-beams) separately. In fact, each of these bridge types presents unique structural characteristics and behaviours that need to be accurately captured in the modelling process. To create precise representations of these bridge structures, Finite Element (FE) models are developed. These models are constructed using a combination of available design drawings and digital data obtained through advanced Structural Health Monitoring (SHM) techniques. SHM methods provide valuable insights into the structural health of bridges by monitoring key parameters such as deflections, vibrations, and material properties. The asset modelling process relies on sophisticated surveying methods to gather essential geometry and deflection data. Point clouds, generated through high-resolution 3D laser scanning, offer detailed spatial information about the bridge's geometry and condition. Additionally, digital twins of the bridge, created using photographic methods and other imaging technologies, provide a virtual representation of the structure for analysis and assessment purposes. By leveraging these advanced surveying techniques, the methodology ensures precise data collection, enabling researchers to create accurate Finite Element (FE) models that reflect the real-world behaviour of the bridges under investigation. This detailed asset modelling process is crucial for laying the foundation for subsequent analyses, such as damage identification through ML algorithms, and plays a key role in enhancing the overall reliability and effectiveness of bridge condition assessment and management practices. The subsequent step in the process focuses on utilising ML to identify the damage states of bridges by interpreting the observed bridge deflections. This step begins with a comprehensive parametric analysis of the Finite Element (FE) models. Different scenarios are created to establish a diverse training set for the k-Nearest Neighbor (k-NN) Machine Learning algorithm. These scenarios involve generating a spectrum of deflected shapes for various types of bridges and spans. This is achieved by simulating the reduction in the cross-sectional area of prestressed tendons to replicate the effects of corrosion on the bridge structure. On the account of the percentage of the prestressing tendon losses, damage states are defined. The efficiency of the kNN methodology in identifying damage state in bridges is then tested by inspecting its ability to predict damage states of test samples, for each bridge type, that represent bridges with similar characteristics. The third step in the process involves model updating and optimisation procedures. The primary objective of FE model updating in this proposal is to validate the machine-learning-driven k-NN methodology by comparing the results obtained from the updating process with the damage state predictions derived from the ML algorithm. Finite