PSI - Issue 84

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

486

degraded performance. Second, the use of a large, heterogeneous feature set may obscure class-specific damage signatures, further hindering effective domain alignment. To address these challenges, a damage-tailored domain adaptation strategy is proposed. Instead of training a single multi-class DANN to recognize all damage scenarios simultaneously, the proposed approach employs multiple DANNs operating in parallel, each specifically tailored to a single damage scenario. Therefore, for each damage class: 1. A preliminary sensitivity analysis is performed on the source domain to identify a subset of features that are most responsive to that specific damage scenario. 2. Only these class-sensitive features are used as inputs to a binary domain-adversarial neural network, trained to distinguish between the healthy state and the corresponding damage state in the source domain. 3. The DANN is trained in a semi-supervised domain adaptation setting, where labeled source data guide damage discrimination, while unlabeled target data are used exclusively for domain alignment. Each DANN therefore learns damage-specific, domain-invariant representations, focusing exclusively on the transferability of a single damage pattern rather than competing multi-class objectives. Once monitoring data from the target bridge become available, the same feature subsets are extracted and passed independently through each of the parallel DANNs. For each damage-tailored network, TL performance is evaluated using statistical metrics derived from the network outputs. The most probable damage scenario in the target structure is then identified by comparing the responses of the parallel DANNs. Specifically, the damage class associated with the DANN exhibiting the highest TL confidence or classification accuracy is considered the most likely damage condition affecting the target bridge. This competitive evaluation across damage-specific DANNs is thought to enable robust damage identification even when certain damage classes are weakly transferable or absent in the source domain, thereby mitigating negative transfer effects. 3.1. Preliminary sensitivity analysis within the source domain This paper particularly focuses on the sensitivity analysis useful to select the group of features that are needed to feed the damage-specific DANN. The One-vs-Rest (OvR) analysis is a decomposition strategy for multi-class problems that enables class-specific investigation by reducing the original problem into a set of binary comparisons. Given a multi-class dataset with classes and a feature vector ∈ℝ , the OvR approach constructs binary problems, where each class ∈ {1, … , } is compared against the aggregation of all remaining classes. For a given class , binary labels are defined as: ( ) = { 1, if = 0, if ≠ (4) where denotes the original class label. Feature sensitivity for class is then evaluated by comparing the statistical properties of each feature between the samples belonging to class and those belonging to the rest of the dataset. In this study, sensitivity is quantified through the absolute difference between the class-conditional mean and the mean of the remaining samples: , =∣ , − ¬ , ∣ (5) where , and ¬ , represent the mean value of feature for class and for all other classes, respectively. The resulting sensitivity score , reflects how strongly feature discriminates class from the rest of the population. By repeating this procedure for all classes, a class-feature sensitivity matrix is obtained, enabling the identification of features that are particularly informative for specific damage scenarios. Unlike global multi-class feature ranking methods, the OvR framework prevents dominant classes from masking class-specific patterns and provides a transparent and interpretable basis for damage-tailored feature selection.