PSI - Issue 80

Alessandro De Luca et al. / Procedia Structural Integrity 80 (2026) 403–410 Alessandro De Luca / Structural Integrity Procedia 00 (2019) 000 – 000

408

6

Figure 4. Density distributions of (a) silhouette score, (b) purity, and (c) balance for all pair (red lines) and triplet (blue lines) feature combinations. The plots show that although average values are modest, certain combinations achieved indicator values approaching unity. Figure 5 presents the three best-performing combinations of features in clustering healthy and damaged states for the tested panels. The criterion used to define “best” combinations prioritized purity as the main objective, followed by balance, and finally silhouette score. This hierarchical approach reflects the need of identifying feature combinations that not only form compact clusters but, more importantly, produce cluster partitions that closely correspond to the actual structural state (healthy or damaged) while ensuring balanced cluster sizes. The three identified best combinations consistently correspond to data acquired at 300 kHz for the I path category. Two of these are the triplets 1-10-23 and 1-10-26, while the third is the pair 10-11. Remarkably, all three top performing combinations share feature 10, highlighting its critical role in distinguishing the presence of impact damage under these specific loading condition. The purity values for these combinations are particularly high (0.92), confirming the strong alignment between clustering assignments and the true physical condition of the structure. The balance (0.74) is also notable, indicating that the clusters formed are well-proportioned in terms of size. Although the silhouette scores are modest compared to the purity and balance (0.56 for the pair and 0.64 for the triplets), they still reflect a reasonable separation between clusters. These findings suggest that feature 10, when used in appropriate combinations, is highly effective at capturing the differences in UGW response between pristine and damaged conditions in this experimental scenario.

Figure 5. Visualization of the three best clustering combinations. Colours indicate GMM-predicted clusters (blue, red); marker shapes show actual condition (circles for pristine and squares for damaged).