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

256

3

1.2 Machine Learning (ML) for bridge damage detection Machine learning is a subset of artificial intelligence that specifically deals with the study, design and development of algorithms that allow computers to learn from data (Alpaydin, 2004). ML and data are inherently interconnected, since data plays a central role in the training, evaluation and deployment of ML models. ML methods can be categorised into three main types: supervised learning, unsupervised learning, and reinforcement learning (Alpaydin, 2016). In supervised learning, algorithms learn from labelled training data to predict outputs for new data points. This type includes tasks like regression and classification. Unsupervised learning involves learning from unlabelled data to identify patterns or relationships without explicit guidance. Common tasks include clustering and dimensionality reduction. Reinforcement learning is a learning approach where an agent learns to make decisions by interacting with an environment and receiving feedback in the form of rewards or penalties. The agent aims to maximise cumulative rewards through trial and error. ML has the potential, through advanced mathematical frameworks and algorithms, to analyse large volumes of data collected from structures to identify patterns, anomalies, and trends that indicate structural health conditions and damage levels (Fan et al., 2021). Therefore, Structural Health Monitoring (SHM) is of vital importance for gathering sufficient and quality data to provide the necessary support for the AI-based bridge damage identification technologies (Bao and Li, 2021). The use of the different ML techniques depends on the type and quality of data available as well as on the aim of the assessment. Deep convolutional neural networks are used to classify image blocks as containing damage or not (Zhang and Yuen, 2022). Deep learning is a subset of machine learning that utilises neural networks with multiple layers to learn complex patterns and features from images. Kim et al. (2018) employed UAVs and R-CNN for the detection and characterisation of cracks on bridge surfaces. Their research involved the collection of 384 crack images with a resolution of 256 x 256 pixels to create a dataset. They then analysed two essential crack parameters, area and width, based on the extracted crack image blocks. Ma et al. (2020) introduced a UAV-assisted ML tool for detecting post-earthquake bridge damage. They utilised Faster R-CNN to rapidly identify and locate cracks and spalling on bridge surfaces. Point cloud data from terrestrial laser scanning (TLS) enables the creation of detailed 3D visualisations of bridge components, providing a realistic representation of the structure within a Building Information Modelling (BIM) framework. Point cloud data can be semantically segmented at the pixel level using deep learning models, allowing for the extraction of specific geometric parameters and component information from the scanned data (Lee et al., 2021). PointNet is a neural network that has the ability to handle unstructured point cloud data efficiently, enabling the classification of bridge components based on the spatial relationships and shapes within the point cloud (Kim et al., 2020). Assessing internal damage in bridge structures is particularly challenging. Advanced techniques like Infrared Thermography, Ground-Penetrating Radar (GPR), Vibration Response Analysis, Acoustic Emission Testing, and X ray Imaging have been developed to address this challenge (Zhang and Yuen, 2022). These methods detect internal damage by analysing temperature variations, imaging the subsurface, studying the dynamic behaviour, monitoring stress waves, and providing detailed internal images which can be exploited by ML algorithms for bridge damage detection. Zhang et al. (2022) presented a method for the automated detection of bridge deck corrosion utilising the Single Shot MultiBox Detector (SSD) and scanned images of bridge decks. The training phase involved 10316 scanned images and testing was conducted on the basis of 2578 images of 300 x 300 pixels. The SSD model demonstrated a high level of accuracy by correctly identifying and locating 677 corroded bars within the scanned images with a precision of 98%. However, deck scanning techniques are often expensive therefore are rarely adopted by transportation stakeholders and operators. Intensive research has been conducted in developing damage identification methods that exploit modal properties of bridges such as natural frequencies, mode shapes and modal damping obtained through the interaction of passing vehicles and bridge structures (Huth et al., 2005). Ghiasi et al. (2022) adopted a supervised algorithm, the k-Nearest Neighbour (kNN) classifier, to classify the level of damage of steel railway bridges by exploiting modal properties. However, it has been shown that modal properties may not be sensitive enough to detect local damage (Ono et al., 2019). Furthermore, physical interaction with the asset may not always be possible due to its fragile nature or safety concerns (Liu et al., 2024). This paper aims to extend an existing methodology that was developed before by Kazantzi et al. (2024a, 2024b). The methodology exploited bridge deck deflections for identifying the damage state of a bridge through ML. The original methodology was built upon and demonstrated on account of a case study balanced cantilever concrete bridge.