PSI - Issue 78

Marco Pirrò et al. / Procedia Structural Integrity 78 (2026) 1641–1648

1642

1. Introduction Continuous monitoring of Civil Engineering structures, combined with Operational Modal Analysis (OMA), offers a promising approach to Structural Health Monitoring (SHM) (Farrar and Worden 2007). This strategy leverages the sensitivity of modal parameters to variations in stiffness and mass. Nevertheless, applying conventional OMA-based techniques in practical SHM scenarios can be challenging due to potential inaccuracies in modal parameter estimation. Moreover, these parameters are influenced not only by actual structural modifications but also by environmental and operational variability (EOV) – such as temperature fluctuations and vehicular loads on bridges (Magalhães et al. 2012, Borlenghi et al. 2023). Since such external influences can significantly affect modal characteristics, it is essential to account for and reduce their impact to prevent misinterpreting normal variability as structural damage. Deep Learning (DL) frameworks are garnering growing attention for SHM, complementing the success of traditional OMA techniques in early detection of anomalies . DL’s appeal lies in its capacity – through multiple layers of interconnected artificial neurons – to capture complex data patterns. In particular, the sparse autoencoder (SAE) (Pathirage et al. 2018) has emerged as a strong candidate for SHM applications: latent data structures are highlighted by compressing input data into a lower-dimensional representation. Moreover, SAEs are thought to manage EOV intrinsically, which could obviate separate correction procedures (Finotti et al. 2022, Wang and Cha 2022). Unlike conventional techniques that require individual networks for each data channel (Finotti et al. 2022), the method herein presented employs a unified network trained on all channels simultaneously (Pirrò and Gentile 2025). This training process assumes the structure is in an undamaged state and exposed to EOV. Once training is complete, the trained SAE can process new input data to reconstruct the original signals. If the structure and EOV conditions remain consistent with those during training, the SAE is expected to reconstruct the signals with high accuracy. To assess structural integrity, the Mean Absolute Error (MAE) between the input and the reconstructed signals is computed. Any significant increase in MAE suggests the presence of damage, with higher error values typically indicating compromised areas, thereby aiding in their detection. The ADA bridge, a steel truss structure introduced by Kim et al. (2021), serves as the benchmark for evaluating the effectiveness of the proposed method. This approach demonstrates strong performance in detecting and pinpointing real structural damage under minimal environmental variation, as data were gathered over two consecutive days with relatively stable temperatures. Previous investigations of the data collected on the ADA bridge were performed to identify damage occurrence – Chang and Kim (2016), for example, explored changes in mode shapes, noting that only slight variations occurred, primarily when the damage was asymmetrical. Also, in Al-Ghalib (2022), power spectral density functions of acceleration time series proved useful in detecting the presence of inflicted damage, though no information is provided on the damage location. On the contrary, the current study enhances those findings by employing a SAE network directly from raw measurement data, eliminating the need for modal analysis. Additionally, the presented method outperforms earlier efforts by accurately localizing all instances of induced damage to the structure. Although for ADA bridge case the influence of EOV is practically negligible, the proposed SAE-based procedure has been successfully applied in continuous monitoring activities under a wide EOV range (Pirrò and Gentile 2025), where reference is made to the benchmark KW51 bridge to detect structural changes due to retrofitting in 14 months of continuous monitoring. 2. Damage detection based on SAE An AE network is aimed at reconstructing input time series through three distinct types of neural layers (see Fig. 1): the encoder, the hidden layer, and the decoder. The encoder is aimed at compressing the input data, so that data dimensionality is reduced and underlying patterns are effectively identified to generate a compact representation. The decoder then leverages the features extracted by the hidden layer to reconstruct the original input sequence at the output layer. Given a time series x n  I , then the encoding procedure can be mathematically expressed as: h n = f ( Wx n + b 1 ) (1) where h n  J (J < I) collects the hidden patterns learned by the encoder, W  J  I is the unknown weight matrix and