PSI - Issue 84

Andy Duarte-Taño et al. / Procedia Structural Integrity 84 (2026) 280–287

281

Keywords: Compressed Sensing; Convolutional Residual Autoencoder; Deep Learning; Structural Health Monitoring (SHM); Wireless Sensor Networks (WSNs)

1. Introduction Structural health monitoring (SHM) systems play a key role in ensuring the safety and long-term performance of civil infrastructure, including bridges, buildings, and heritage structures [Delgadillo and Casas (2022), Ni et al. (2021), Zhu et al. (2018)]. Among the available approaches, vibration-based SHM systems are widely adopted because they enable non-destructive assessment of structural conditions without interfering with normal operation. These systems extract dynamic features, such as natural frequencies and mode shapes, from measured vibration responses, which serve as damage-sensitive indicators for early detection and maintenance planning. However, the increasing use of dense sensor networks and long-term monitoring leads to large volumes of data, making storage, transmission, and processing major challenges, particularly for wireless sensing systems with limited power and communication resources [Jaafar et al. (2012), Alsalaet and Ali (2015)]. Data compression has therefore emerged as an enabling technology for scalable SHM. While effective data reduction can significantly lower operational costs, excessive compression may compromise the preservation of modal information required for reliable operational modal analysis (OMA). This trade-off between data reduction and modal fidelity has been reported in recent studies [Talebi-Kalaleh and Mei (2024), Zhou et al. (2025)] and remains an open challenge, especially for real structures affected by environmental variability. Existing approaches include classical compressive sensing (CS) methods [Donoho et al. (2006)] and machine-learning-based strategies. Classical CS frameworks rely on sparsity assumptions and mainly target time-domain reconstruction accuracy [Bao et al. (2020)], whereas optimization-based schemes can improve fidelity at moderate compression levels but are often sensitive to noise [Zhang et al. (2024)]. More recent deep learning approaches have shown promising modal preservation under laboratory conditions [An et al. (2024)], while model-assisted CS techniques exploit prior modal information at the cost of reduced applicability in blind monitoring scenarios [Zonzini et al. (2021)]. Based on the above considerations, there remains a need for compression approaches capable of achieving high time-domain reconstruction accuracy while preserving dynamic identifiability without relying on prior modal knowledge. To address this gap, this paper presents a lightweight residual convolutional autoencoder for vibration signal compression and reconstruction. The proposed framework employs a joint time–frequency reconstruction loss without predefined sparsity assumptions and is validated using a synthetic vibration dataset and a real monitoring deployment on the San Jerónimo bell tower in Granada (Spain), demonstrating effective modal preservation at moderate compression ratios under both controlled and field conditions. 2. Proposed compression–reconstruction framework 2.1. Model architecture and time-frequency reconstruction loss The proposed model follows a progressive residual autoencoder architecture tailored for vibration signal compression and reconstruction. The encoder progressively reduces temporal resolution of the input signal through a series of 1D convolutional residual blocks [He et al. (2016)], while increasing the number of filters to capture higher level representations. The decoder mirrors this structure by gradually upsampling the compressed representation back to the original resolution. The overall architecture is illustrated in Fig. 1. Residual connections are used throughout the encoder and decoder to stabilize training and preserve low-level features. The degree of compression is governed by the number of downsampling stages. The compression ratio is defined as = / , where N is the original signal length and M, the latent representation length after the encoder.