Conference Agenda

Reliable deployment of data-driven Machine-Learning (ML) based Structural Health Monitoring (SHM) depends strongly on the quality of the data and on the features extracted from it. Environmental and operational variabilities (EOVs), sensor faults and malfunctions, different type of damages or other confounding influences present in SHM data can make it difficult for pattern recognition models to identify structural changes, a problem widely observed in long-term monitoring applications and highlighted throughout recent research on data-driven SHM.

This contribution investigates how different types of damage-sensitive features (DSFs) support damage identification tasks when data contain combined sources of variability. A numerical benchmark was first developed to reproduce realistic long-term SHM data, retaining full control over ground-truth conditions. The dataset simulates the structural response under Ambient Vibration (AV) monitoring, affected by EOVs, sensor faults and malfunctions, and both fast-varying and slow-varying damage mechanisms. Using this dataset, two categories of features were extracted and compared: (i) physical features derived from established formulations (like statistical descriptors, spectral and energy-based indicators, time-frequency features, correlation-based measures, and model-based quantities) and (ii) latent features learned from an autoencoder (AE) model. The AE was designed and trained following dedicated preprocessing, careful normalisation, and windowing procedures; emphasis was placed on constraints specific to AV data and long-term monitoring. Both reconstruction errors and latent codes were evaluated as DSFs.

The comparison focused on two tasks aligned with common SHM objectives: unsupervised detection of fast-varying damage using One-Class SVM models, and prediction of slow-varying degradation trends using an LSTM-based forecasting approach. Performance was evaluated in terms of damage detectability and predictability, robustness to EOVs and sensor faults, minimum detectable change, and computational cost. Results show that latent features outperform physical DSFs in detecting fast, abrupt stiffness changes and in separating structural variations from sensor-related anomalies. Conversely, spectral and energy-based DSFs provide more stable indicators for slow damage progression compared to latent features.

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.

Reliable structural health monitoring (SHM) of wind turbines requires methods capable of handling high-dimensional, heterogeneous multi-sensor data under both labeled and label-scarce conditions, while remaining computationally efficient for real-world deployment. This paper presents a unified SHM framework that integrates feature-level data fusion, Vision Transformer (ViT) classification, multi-objective AutoML optimization, and unsupervised anomaly detection within a single pipeline.

Multi-channel signals from 26 sensors are preprocessed and encoded into a compact three-channel feature tensor combining statistical descriptors, spectral and wavelet features, and a PCA-denoised inter-sensor correlation matrix. This representation enables efficient storage of time-window information and captures both local signal characteristics and global cross-sensor dependencies. A lightweight ViT is trained for supervised fault classification, while its architecture is optimized using NSGA-II to jointly maximize predictive performance and minimize computational cost, enabling deployment on resource-constrained edge devices.

To address label scarcity, an autoencoder trained solely on normal-condition data is used for anomaly detection. Comparative evaluation of reconstruction- and latent-space-based metrics shows that Mahalanobis distance in the latent space provides superior sensitivity to subtle faults.

Validation on the ETH Aventa AV-7 dataset demonstrates up to 98.4\% macro-F1 in classification and robust anomaly detection performance. The results confirm that the proposed multisensor fusion strategy provides a reliable and scalable pipeline for SHM of real structures, applicable to both supervised and unsupervised scenarios under practical computational constraints.