PSI - Issue 77

Douaa Benhaddouche et al. / Procedia Structural Integrity 77 (2026) 152–160 Douaa BENHADDOUCHE/ Structural Integrity Procedia 00 (2026) 000 – 000

153

2

on the full-scale Tianjin bridge in order to demonstrate its efficiency in identifying the presence of damage and assessing the global structural condition of the bridge. © 2026 The Authors. Published by ELSEVIER B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) Peer-review under responsibility of ICSI organizers Keywords: Structural Health Monioring; Unsupervised learning; Graph Neural Network 1. Introduction Large-scale bridges play a vital role in ensuring transport efficiency and regional economic growth, making their safety and functionality essential. To address the challenges posed by external factors such as temperature variations, traffic loads, wind, and seismic events, data-driven SHM systems have become increasingly important. These systems complement traditional visual inspections by enabling continuous, objective assessments and early detection of damage. Considerable attention has been devoted to vibration-based monitoring, as structural damage modifies key physical properties such as stiffness, and consequently alters the dynamic response. Recent studies have explored various deep learning strategies in vibration-based SHM systems, often differing in how features are extracted from vibration signals. Early works combined wavelet- and Fourier-based transforms to generate frequency-domain features subsequently analyzed by models such as DBMs (Rafiei and Adeli, 2018) or DCNNs (Wang, 2023; Yu et al., 2019). These approaches typically relied on preprocessing steps like wavelet packet decomposition, synchronized wavelet transformation, or FFT to denoise signals, reduce dimensionality, and enhance damage-sensitive characteristics before model training. In contrast, other methods emphasize statistical modeling of time-series residuals. For instance, an Auto-Encoder (AE) have been applied to compress ARMA residuals, with Mahalanobis distance used in latent space for structural conditions classification (Entezami et al., 2020). More recent efforts have focused on directly leveraging neural networks to automatically extract damage-sensitive features from raw data, avoiding handcrafted transformations. This includes Denoising AE (DAE) with probabilistic modeling based on reconstruction errors and distribution comparisons (e.g., Kolmogorov – Smirnov test (Jiang et al., 2021)) or Variaional AEs (VAE) for robust probabilistic feature representation (Ma et al., 2020). Convolutional Neural Networks (CNNs) are especially prominent due to their efficiency and ability to capture hidden patterns in vibration data. Both 1D-CNNs (Avci et al., 2021; Sony, n.d.; Wang, 2023) and 2D-CNNs (Rastin et al., 2021; Teng et al., 2020) have been applied for different detection levels : from damage identification and severity classification to real-time localization, often with feature comparison against reference states using distance-based metrics. Hybrid designs combining CNNs with Long Short Time Memory (LSTMs) (Fu et al., 2021; Hung et al., 2020) further extend this capacity by jointly modeling spatial and temporal dependencies. More advanced approaches incorporate semi supervised strategies and emerging architectures. For exemple, (Dang et al., 2023) introduces an encoder that integrates CNNs with Transformer – Graph modules, enabling attention-driven feature refinement and improved generalization with limited labeled data. Most developped SHM methodologies rely on supervised learning, thus they require labeled data from both undamaged and damaged scenarios. In practice, collecting and labeling such data is costly and often infeasible for structures in service. Moreover, dimensionality reduction methods may cause information loss and compromise detection accuracy (Soleimani-Babakamali et al., 2023), while deep learning models such as CNNs, autoencoders, and hybrid networks (Fu et al., 2021; Hung et al., 2020; Ma et al., 2020; Wang and Cha, 2021) often demand high computational resources, limiting their applicability in real-time monitoring. Another drawback is the lack of physical interpretability of features extracted from vibration signals (Rafiei and Adeli, 2018). Finally, as highlighted in (Meruane, 2016), subtle changes in structural integrity often produce minimal variations in time- or frequency-domain responses, making small-scale damage particularly difficult to detect with conventional techniques. In this paper, we propose a new unsupervised method for damage identification and global structural state assessment using sensor data and a hybrid neural network. This neural network model combines a graph convolutional layer and LSTM layer, outperforming traditional deep learning models by better representation and extraction of the spatio-temporal features from the monitoring data. The hybrid model is trained on acceleration data and sensor graph