PSI - Issue 78

Alina Elena Eva et al. / Procedia Structural Integrity 78 (2026) 387–394

389

Fig. 1. Domain Adversarial Neural Network

p(Ys | Xs) and p(Yt | Xt) are di ff erent)[P. Garden et al. (2020)]. A DA-based damage detection approach has two main odjectives: on the one hand, to correctly distinguish data from di ff erent classes within the same domain, and on the other hand, to identify similarities between classes across the two domains. One of the advantages of DANN is that they require labeled data only from the source domain. The objective is to make accurate predictions on the target domain by learning domain-invariant representations through adversarial training. DANN consists of three core components: a feature extractor, a label classifier, and a domain classifier(see Fig. 1). The feature extractor, composed of neural network layers, processes input data from both domains and generates representations intended to be domain invariant. The label classifier is trained using labeled data from the source domain to predict class labels. Since target data are unlabeled, they are not directly used in this part of training. To align the source and target feature distributions, a domain classifier is trained to distinguish the domain of origin for each sample. A Gradient Reversal Layer (GRL) is placed between the feature extractor and the domain classifier. During backpropagation, the GRL reverses the gradient’s sign, encouraging the feature extractor to produce representations that confuse the domain classifier. When the domain classifier is unable to di ff erentiate between source and target samples, the extracted features are considered e ff ectively domain-invariant. The overall model starts with source domain input data, which are passed through the feature extractor. The resulting representations are then used to predict class labels via a dense layer with softmax activation, where each neuron corresponds to a class. The class with the highest predicted probability is selected as the final output. In summary, DANNs are powerful tools for unsupervised domain adaptation, enabling accurate classification in a target domain without requiring labeled target data. By minimizing the distribution gap between domains through adversarial training, they achieve robust and transferable feature representations. The aim of this work is to evaluate the ability of adversarial networks to transfer knowledge between two di ff er ent structures: a masonry panel and a fac¸ade that includes the same panel within it. Although they di ff er in overall geometry and loading conditions, the two structures share identical dimensions, Gothic masonry pattern, and sensor placement in the corresponding area, making them suitable for applying domain adaptation techniques. This approach allows testing the robustness and generalization capability of the classifier, reducing the need for labeled data in the target domain and making structural monitoring more e ffi cient and applicable to real-world scenarios. The DANN was trained using the numerical model of the masonry panel, where seismic damage is simulated through top displacement, representing the e ff ect the fac¸ade could have on the panel itself. Subsequently, the network was tested on the fac¸ade, where the seismic action was simulated by controlling horizontal inertial forces, imposed with a vertically varying distribution of lateral acceleration: 0.3g at the base and 1.2g at the top. This configuration allowed for evaluating the network’s ability to generalize in a more complex structural context while maintaining high classification reliability. 3. Case study