PSI - Issue 84

Valentina Giglioni et al. / Procedia Structural Integrity 84 (2026) 481–488

487

4. Results: selection of damage scenario-related features The considered feature space consists of 54 input variables, including a combination of modal frequency-related features and nodal rotation measurements, while the classification problem involves 17 distinct damage scenarios. Firstly, a normalization process between -1 and 1 is carried out to avoid scale-induced bias in mean-based metrics. Since not all features are equally informative for discriminating against every damage scenario, the OvR sensitivity analysis is performed to identify the most relevant features for each class independently. For each scenario, the corresponding samples are contrasted against the aggregation of all remaining scenarios, and feature sensitivity is quantified as the absolute difference between class-conditional and rest-conditional mean values. Based on this analysis, the top 10 most sensitive features are selected for each scenario. These reduced, damage-specific feature subsets are subsequently intended to be employed as inputs to 17 parallel damage-tailored DANNs, each specialized in recognizing a single damage scenario under domain shift conditions. Fig. 3 presents the resulting OvR sensitivity heatmap, where rows correspond to damage scenarios and columns to the input features. The colour intensity reflects the relative importance of each feature for a given scenario, enabling a direct visual comparison of feature relevance across all damage classes and confirming that different damage mechanisms activate distinct subsets of features. In particular, frequency-based features exhibit minimal sensitivity for scenarios 9–14, which are associated with pseudo-damage conditions affecting the cable system, indicating that modal frequencies are weak indicators for these damage types. Note that the influence of nodal rotation features is not uniform. The three rotational components exhibit markedly different sensitivities, reflecting the fact that the structural deformation pattern varies across simulated damage scenarios, thereby altering the contribution of each rotational degree of freedom. This behaviour confirms that both the location and the directional component of rotational measurements are damage-dependent, further motivating the adoption of scenario-specific feature selection rather than a global feature set.

Fig. 3. OvR sensitivity heatmap

5. Conclusions This paper has presented a preliminary investigation into sensitivity-based feature selection as a supporting tool for the development of damage-tailored DANNs in bridge SHM, whose final goal would be to address two critical challenges in practical applications: the lack of labeled damage data in real structures and the heterogeneity between