PSI - Issue 84

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

286

Fig. 7. San Jerónimo bell-tower case: Raw versus reconstructed acceleration response corresponding to the selected reference sensor channel. (a) Full time-domain signal, (b) zoomed-in time window highlighting local waveform fidelity, and (c) corresponding power spectral density estimated using Welch’s method.

Fig. 8. San Jerónimo bell-tower case: a) Comparison of reference modes of vibration obtained from the original and reconstructed acceleration signals. The modal displacements extracted from the original signals are depicted with solid lines, while those obtained from the reconstructed signals are shown with dashed lines. b) Top : MAC matrix between the selected modes identified from the original and reconstructed acceleration signals. Bottom : Auto-MAC matrix of the modes identified from the original signal. 4. Conclusions This paper has presented a residual convolutional autoencoder for vibration signal compression and reconstruction in SHM applications, designed to achieve high time-domain reconstruction accuracy while preserving modal information for operational modal analysis. A joint time–frequency reconstruction loss was employed to balance waveform fidelity and spectral consistency, enabling the retention of physically meaningful dynamic features after compression. The proposed network was validated using both a synthetic vibration dataset and real-world ambient vibration data recorded at the San Jerónimo bell tower in Granada, Spain. For the synthetic case, accurate signal reconstruction and preservation of vibration mode shapes were achieved at compression ratios up to = 8 . The real world deployment required a more conservative compression ratio ( = 2 ) to ensure reliable modal identification,