PSI - Issue 78

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

897

SV dynamics and temperature variations under anomalies, suggesting that temperature residuals can indicate structural issues.

(a)

(b)

(c)

(d)

(e)

(f)

Fig. 5: Diagnostics metrics distributions for Train, Test-healthy and Test-damaged samples. Here, ‘wk’ stands for ‘week’. (a) Reconstruction error; (b) OGSR; (c) Cosine similarity; (d) Wasserstein Distance; (e) KL Divergence; and (f) JS Divergence

5.1. Reconstruction-Based Anomaly Detection (RBAD) Metrics

Six distinct spectral reconstruction metrics were employed to assess anomaly detection performance, includ ing Reconstruction Error, Overall Gradient Similarity Ratio (OGSR), Cosine Similarity, Wasserstein Distance, Kullback-Leibler (KL) Divergence, and Jensen-Shannon (JS) Divergence. For healthy training and test samples, these metrics generally remained below a 95th percentile threshold es tablished from the training data, as shown in Figure 5. Damaged samples consistently exhibited clear elevations in metrics like Reconstruction Error, Wasserstein Distance, KL Divergence, and JS Divergence, followed by a decline over time, suggesting a gradual structural resettlement post-seismic disturbance. Cosine Similarity for damaged samples sharply dropped to approximately 0.2 or lower, indicating substantial directional dissimilarity. Quantitatively, metrics such as Reconstruction Error demonstrated high performance, achieving an accuracy of 0.9864, sensitivity of 1.00, specificity of 0.9767, and an F1-score of 0.9839. Notably, JS Divergence showed the highest accuracy (0.9932) and F1-score (0.9919). These results collectively a ffi rm the model’s ability to robustly identify subtle yet persistent anomalies while e ff ectively compensating for operational and environmental e ff ects. Table 1 summarises the performance parameters of the adopted methodology.

6. Conclusion

This study introduces a data-driven framework for unsupervised anomaly detection in the structural health monitoring (SHM) of historical masonry structures. Utilising a multivariate convolutional autoencoder (CAE), the methodology integrates dynamic spectral responses—focused on dominant-SV features—with key environmental variables, such as temperature metrics, enhancing anomaly detection robustness and interpretability. The CAE model is trained exclusively on data from healthy structures, identifying anomalies by detecting significant devia tions from a healthy manifold. This dual-domain approach e ff ectively distinguishes real structural anomalies from