Current Structural Health Monitoring (SHM) systems remain constrained by their reliance on handcrafted feature extraction and centralized cloud processing, limiting their real-time performance, scalability, and deployment on resource-constrained infrastructure. This study seeks to overcome these limitations by introducing a hybrid edge/cloud SHM framework that facilitates low-power, real-time anomaly detection directly on embedded devices. We introduce a temporal transfer learning approach based on the WaveNet and Recurrent Neural Network architectures that operates on raw time-series data and eliminates the need for statistical or frequency-domain feature engineering. The models were trained and validated using real-world datasets from the Z24 (Switzerland) and S101 (Austria) bridges and deployed on edge hardware (Raspberry Pi 5 and LattePanda Mu, Intel® N100) for on-device inferences.

The framework implements a hierarchical inference strategy in which edge devices perform continuous, low-latency anomaly detection, whereas cloud resources are selectively engaged for the high-precision analysis of critical events. This design reduces bandwidth requirements and enhances system autonomy without compromising accuracy. The results demonstrate that the proposed approach achieves a high anomaly detection performance in time-domain transfer learning while maintaining an efficient edge deployment. This study provides a validated pathway toward scalable, real-time, and feature-free SHM systems for deployment in operational bridge networks, supporting continuous monitoring, early damage detection, and maintenance decision-making in the future.

Dam displacement monitoring is imperative to assess the operational status and structural safety of dams under various environmental conditions and operational loads. Although most of the dam structures are instrumented with robust in-situ sensing systems, long-term field monitoring is often constrained by practical challenges such as instrumentation costs, sensor drift or malfunction, data gaps, and the inherent complexity of operating and maintaining large-scale monitoring networks. Machine learning offers an alternative solution to deal with these challenges by enabling data-driven prediction and interpretation of structural responses. Despite numerous regression-based predictive models for predicting dam structural responses, a demanding issue arises from incomplete training data stemming from the unavailability of some influential environmental and operational factors. This study aims to address the aforementioned engineering and technical limitations by proposing an intelligent hybrid regressor. This model integrates ridge regression with a convolutional neural network (CNN), leveraging the strengths of both statistical learning and deep learning paradigms, centralized on a residual correction mechanism. First, the ridge regression model performs initial dam displacement predictions using the available environmental and operational factors. Second, prediction errors (residuals) of the ridge regression model are fed into the CNN to capture hidden nonlinear relationships embedded in the residuals and subsequently enhance overall prediction accuracy. Given this sequential dual-stage prediction architecture, the proposed hybrid regressor can address the limitation of unavailable environmental and operational factors that significantly influence dam deformation behaviour. A real-world dam structure, along with limited data including reservoir levels and temperature, is employed to validate the proposed predictive method. Results show that the proposed hybrid regressor achieves a prediction accuracy of approximately 89% using incomplete data while effectively capturing strong nonlinear characteristics in dam displacement responses.

Strain and displacement measurement is a fundamental task in structural health monitoring. Traditionally, Digital image correlation (DIC) is used for image-based deformation measurement by comparing corresponding local regions in two images to determine the surface deformation field. However, DIC requires the surface to be coated with speckles to create unique texture features to achieve matching. This requirement significantly limits the application of DIC technology, as it is often impractical to apply speckle patterns on a large scale. Recognizing that natural material surfaces often possess natural texture features, we propose a speckle-free approach to deformation field estimation by leveraging deep learning techniques to extract these natural textures. We frame this challenge as a generalized dense correspondence matching problem, encompassing two specific tasks: optical flow estimation and disparity estimation. To address this, we introduce a unified cost volume transformer designed to extract 4D cost volume information, facilitating the aggregation of long-range dependencies. Additionally, we employ a compact latent space strategy to compress the cost volume dimensions, thereby minimizing memory costs. Furthermore, we implement a masked fine-tuning strategy to adapt the pre-trained model to various texture characteristics in the absence of ground truth reference outputs. Finally, we construct a binocular measurement system using two GoPro cameras. The effectiveness of our proposed method is validated through experiments conducted with different materials.

Modal-based structural health monitoring requires both reliable long-term extraction of modal features and statistical models capable of separating normal environmental and operational variability from structurally induced change. This paper presents a fully automated framework that combines frequency-domain-decomposition-based mode identification and tracking with a latent-factor state-space model with exogenous inputs (SSM-X) for innovation-based novelty detection. Tracked natural frequencies are treated as the observation vector, while measured environmental and operational variables are incorporated as exogenous inputs. The healthy evolution of the modal features is represented through a low-rank linear-Gaussian dynamical system, estimated from baseline data by expectation-maximization using Kalman filtering and Rauch–Tung–Striebel smoothing. Novelty is quantified through the one-step-ahead normalized innovation squared, yielding a probabilistically normalized multivariate damage indicator that naturally accommodates missing observations. The framework is demonstrated on two months of continuous monitoring data from an aluminium arch pedestrian bridge. Results show stable predictive performance on unseen healthy data and clear sensitivity to simulated novelty scenarios. In particular, the low-rank multivariate formulation exploits cross-modal structure to detect departures that may remain weak at the single-mode level, while preserving robustness to incomplete and irregular modal time series.

Concrete structure crack identification and measurement is traditionally conducted by human inspectors who visually identify and quantify defects such as cracks, assisted by tools such as a ruler or tape measure. Between inspection intervals, inspectors monitor changes in crack width as well as its propagation. However, this approach can be subjective, time-consuming, and inconsistent, particularly in poor or varying lighting conditions. To address these challenges, extensive research has explored the use of image processing and artificial intelligence techniques to automatically classify pixels in images of concrete surfaces as cracked or uncracked, and monitor their growth over repeat inspection intervals. Further studies have crack segmentation techniques can be enhanced through controlled scene illumination. One approach to this is to capture multiple images, each individually illuminated with lighting from a different angle and direction. This approach has been shown to highlight smaller cracks which may not be visible under regular diffused lighting approaches, but has not been demonstrated for crack width monitoring.

Despite these advances, most existing studies and datasets have focused on clean, unchanging surfaces, neglecting the effects of environmental weathering such as dirt accumulation, surface discolouration, and biological growth. As a result, the performance of these methods under realistic, evolving surface conditions remains largely untested.

This paper extends existing work, demonstrating how directional lighting-enhanced pixel-level concrete crack segmentation methods can monitor crack width growth over time. This allows crack width monitoring at repeat inspection intervals. In this study, a laboratory dataset of concrete slabs, each containing cracks that originated at 0.1 mm in width, was created, with the cracks iteratively widened and propagated until approximately 2 mm in width. At each width iteration, the cracks were imaged under various lighting conditions, with controlled lighting angle and direction. Captured directional lighting data was given to a crack segmentation model to classify pixels as cracked or uncracked. The segmentation output was measured using a skeleton-based crack width measurement algorithm, resulting in automated measurement of crack widths for each dataset interval.

Between inspection intervals of crack width growth, the samples were deliberately artificially “weathered” to simulate realistic surface changes that occur on real-world concrete structures. This process aimed to test the robustness of the proposed crack segmentation algorithm under conditions that more closely represent real-world inspections, where both cracks and their surrounding surfaces evolve over long periods between repeat inspections.

The proposed approach demonstrated superior crack segmentation performance to diffused lighting methods in terms of F1 score, recall and precision. Measured widths were also compared to a crack measurement ruler (i.e. the current state of practice), with minimal deviation. These findings highlight the potential for more reliable and automated concrete crack monitoring, particularly in low-light or variable environmental conditions.

Preserving the structural integrity of heritage masonry arch bridges presents unique challenges, particularly within historically dense environments like Venice where non-invasive methods are paramount. Ambient vibration monitoring (AVM) offers a well-established starting point, allowing us to capture the dynamic behavior of these structures under their normal operating conditions. Our ongoing work involves a large-scale AVM campaign across Venice, using synchronized velocimeters to systematically gather vibration data. From this, we extract key dynamic parameters like dominant frequencies, mode shapes and damping ratios, alongside essential geometric features. Simultaneously, careful visual inspections provide the critical qualitative context regarding each bridge's state of conservation. By integrating these distinct quantitative and qualitative data streams, we have been developing a unique baseline archive crucial for moving beyond characterization towards meaningful interpretation.

The central aim of this paper is to explore how this integrated dataset can be leveraged to address the significant challenge of automated condition assessment. We employ a supervised machine learning framework to bridge the gap between measured data and assessed structural condition. The curated archive, which links the dynamic and geometric measurements (our inputs) to discrete condition categories derived from the visual inspections (our outputs), provides the training ground for established classification algorithms – specifically Support Vector Machines (SVM) and Random Forest (RF). Standard pre-processing techniques, including feature scaling, are incorporated to ensure the classifiers can effectively learn from the data without scale-induced biases.

Initial explorations with this framework have already yielded valuable insights. For instance, feature importance analysis of the Random Forest model, strongly suggests that the dynamic parameters captured via AVM—particularly damping ratio and dominant frequency—offer considerably more diagnostic information about condition than the simpler geometric descriptors alone. This finding reinforces the value embedded in vibration monitoring. Moreover, our work confirms the feasibility of training these models to distinguish between different condition states using only the combined instrumental and geometric data. Importantly, from an engineering perspective, we recognize that raw classification accuracy is insufficient; ensuring high sensitivity (Recall) in identifying potentially compromised structures is critical for responsible heritage management and risk mitigation.

This combined AVM-ML methodology provides a scalable foundation for developing more objective, data-supported tools to assist in condition assessment and guide proactive maintenance efforts for vital cultural infrastructure. Further research will necessarily involve expanding the dataset and undertaking rigorous validation of the models.