PSI - Issue 78

Akshay Rai et al. / Procedia Structural Integrity 78 (2026) 891–898

896

sigmoid for the dominant SV channel and tanh for the temperature channels to account for their di ff ering value distributions. All convolutions utilise a kernel size of 3 and a stride of 1, intentionally omitting MaxPooling to maintain spectral granularity. Training is performed using the Adam optimiser (learning rate 1 × 10 − 3 ) and a batch

Fig. 3: Schematic of 1D-CNN based autoencoder

size of 30 over 750 epochs. A dynamic weighting strategy is applied to a multivariate loss function, balancing the reconstruction of dominant SVs (mean-squared error, Huber, and Wasserstein terms), both temperature channels ( T 1 and T 0 ), peak-amplitude preservation, and latent feature consistency via an L2 penalty. These weights are linearly interpolated across epochs to enhance convergence and generalisation. The total training time for the CAE is approximately 20 minutes, highlighting its computational e ffi ciency. The training loss history of the 1D Convolutional Autoencoder (CAE) demonstrated rapid convergence within the initial few hundred epochs. This was followed by a stable descent across all loss components, indicating e ff ective and robust multivariate learning during the training process, as shown in Figure 4.

Fig. 4: Training performance of CAE: Total loss history versus the total epochs

5. Anomaly detection performances

The diagnostic performance of the 1D-Convolutional Autoencoder was thoroughly assessed, emphasising its ability to accurately reconstruct dominant SV spectra and ambient temperatures across both healthy and damaged structural states. The key findings are: 1. Spectral Reconstruction Fidelity: The CAE demonstrated strong generalisation capabilities with healthy test datasets, e ff ectively capturing spectral signatures and showing high similarity between real and generated mean dominant-SV profiles. However, for damaged samples, the model struggled with unfamiliar patterns, leading to notable discrepancies and underestimations of dominant SV magnitude in the 2-4 Hz range. This highlighted its sensitivity to damage-induced changes. 2. Temperature Reconstruction Accuracy: The CAE e ff ectively reproduced non-linear dynamics in temperature profiles ( T 1 and T 0 ) for healthy samples, closely matching real measurements. While there were minor de viations during rapid internal temperature ( T 0 ) increases, the overall quality was consistent. In contrast, for damaged samples, the model had trouble with abrupt temperature changes, resulting in smoother yet accu rate thermal traces. This indicates that the CAE struggled to generalise the non-linear relationship between