PSI - Issue 84

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

282

Fig. 1. Overall architecture of the proposed residual autoencoder. A key element of the proposed framework is a reconstruction loss that simultaneously evaluates fidelity in the time and frequency domains. The time-domain signal captures instantaneous variations and overall dynamic evolution, whereas the frequency-domain representation highlights dominant resonances and periodic components relevant for modal interpretation. The total loss is defined as: ℒ= ⋅ ℒ ( , ̂) + ⋅ ℒ ( , ̂) , , ≥1, (1) ℒ ( , ̂) = ( , ̂) , (2) ℒ ( , ̂) = (Φ( ), Φ( ̂)), (3) where Φ(⋅) denotes a selected frequency-domain representation. The frequency-domain term ℒ was adapted to the dataset characteristics, as detailed in the corresponding case studies. Training setup 2.2. Training setup and evaluation metrics The autoencoder was trained end-to-end using the Adam optimizer with a fixed learning rate of 10 −4 . Early stopping [Prechelt et al. (2002)] with patience of 50 epochs was employed to prevent overfitting, retaining the model weights associated with the lowest validation loss. Unless otherwise stated, Min–Max normalization to [0,1] was applied on a per-window basis, to ensure comparability across sensors and operating conditions. For the San Jerónimo dataset, standard scaling was adopted to account for outliers and environmental variability. Data were divided into training, validation, and testing subsets (70/15/15), applied consistently across all experimental datasets. Reconstruction quality was assessed using complementary signal-level and modal-level metrics, depending on the case study. For the synthetic dataset, waveform fidelity and linear similarity were quantified using the relative reconstruction error (RRE) and the Pearson correlation coefficient , together with the Modal Assurance Criterion (MAC) computed from the identified vibration modes. For San Jerónimo case, where the primary objective is to evaluate the preservation of dynamics features under field conditions, performance was evaluated through operational modal analysis, using the MAC to quantify mode-shape similarity between original and reconstructed responses. All metrics were computed on a window-wise basis and averaged across datasets where applicable.