Deep learning methods for damage detection heavily depend on large labeled datasets, which are often lacking in actual structures. While numerical models can effectively simulate damage and generate labeled datasets, environmental uncertainties and modeling errors create a significant gap between these between the simulated data and the actual observations. Furthermore, the damage scenarios in actual structures often differ from numerical models. This study introduces the Partial Conditional Adversarial Domain Adaptation (PCADA) network, which enables effective cross-domain damage detection. First, the network uses a DenseNet architecture to extract damage-sensitive features from the acceleration response, and a classifier is trained on the source domain to predict damage labels. Then, features and label classifications are combined through multidimensional linear mapping. The resulting combined vectors serve as input to a domain discriminator, ensuring simultaneous alignment of features and classes from different domains. In addition, a transfer weight vector is designed to mitigate the negative impact of outlier label space data within the source domain. The conditional entropy loss is also applied to fine-tune the classifier, which enhances the robustness of the network. Finally, by training the feature generator and domain discriminator adversarially, the PCADA network acquires features that are sensitive to damage and invariant across domains, enabling it to detect damage across different domains effectively. The impact of different modeling errors on the proposed network performance is tested through numerical case studies, and cross-domain damage detection from numerical model to real scale model is also conducted. The results show that the proposed method can significantly compensate the difference between the numerical model and the actual structure, and has a high damage detection accuracy in the case of inconsistent damage types. This paper significantly enhances the practicality of domain adaptation techniques in actual structural damage detection.

中文翻译：

---

通过部分条件对抗域自适应进行跨域损伤检测


用于损伤检测的深度学习方法在很大程度上依赖于大型标记数据集，而实际结构中通常缺乏这些数据集。虽然数值模型可以有效地模拟损伤并生成标记的数据集，但环境不确定性和建模误差会在模拟数据和实际观测数据之间造成巨大差距。此外，实际结构中的损伤场景通常与数值模型不同。本研究介绍了部分条件对抗域适应 （PCADA） 网络，该网络可实现有效的跨域损伤检测。首先，该网络使用 DenseNet 架构从加速响应中提取损伤敏感特征，并在源域上训练分类器以预测损伤标签。然后，通过多维线性映射将特征和标签分类进行组合。生成的组合向量用作域鉴别器的输入，确保来自不同域的特征和类同时对齐。此外，还设计了转移权重向量来减轻源域中异常值标签空间数据的负面影响。条件熵损失还用于微调分类器，从而增强网络的稳健性。最后，通过对抗性地训练特征生成器和域判别器，PCADA 网络获得了对跨域损伤和不变敏感的特征，使其能够有效地检测不同域的损伤。通过数值算例分析，验证了不同建模误差对所提网络性能的影响，并进行了从数值模型到实尺度模型的跨域损伤检测。 结果表明，所提方法能够显著补偿数值模型与实际结构的差异，在损伤类型不一致的情况下具有较高的损伤检测精度。该文显著提高了域自适应技术在实际结构损伤检测中的实用性。