Deep Learning has been increasingly utilized in structural damage detection. Existing relevant studies often highlight the benefits of supervised deep learning in the intelligent identification of bridge damage. Notably, however, supervised deep learning methods encounter specific challenges in processing real-world monitoring data to reflect damage. Typical challenges include: (i) the need for a large dataset derived from a vast number of labeled damaged cases, which are extremely difficult to obtain from real-world monitoring data, experiments, and numerical simulations; (ii) the inability of available damage sample spectra to fully capture the damage information underlying the dynamic responses of bridges; (iii) the likelihood of requiring manual intervention to discriminate damage from the outputs of deep learning models, which is inefficient when dealing with massive amounts of monitoring data. To address these challenges, this study proposes an intelligent unsupervised deep learning framework for damage identification in bridges. The framework is characterized by three innovative technical elements: (1) a Deep Convolutional Autoencoder (CAE) model with a hybrid loss function is developed to provide an intelligent system that can identify bridge damage without requiring a large dataset of damaged cases; (2) Wavelet transmissibility pattern spectra are established to characterize damage information embedded in dynamic responses in a more efficient manner; and (3) an Ordering Points To Identify the Clustering Structure (OPTICS)-based damage picker is proposed to achieve automatic discrimination of damage cases. The feasibility of this framework is numerically demonstrated through the detection of damage in a curved bridge, and its effectiveness is experimentally validated on a laboratory-scale suspension bridge with induced damage. The results indicate that the proposed framework can automatically and accurately identify damage in bridges. This framework provides a solution for intelligent and data-driven bridge damage detection.

中文翻译：

---

使用带有小波传输率模式谱的深度卷积自动编码器进行桥梁无监督损伤检测的智能框架


深度学习越来越多地应用于结构损伤检测。现有的相关研究经常强调监督深度学习在桥梁损伤智能识别方面的好处。然而值得注意的是，有监督的深度学习方法在处理现实世界的监测数据以反映损害方面遇到了特定的挑战。典型的挑战包括：（i）需要从大量标记的损坏案例中得出的大型数据集，而从现实世界的监测数据、实验和数值模拟中很难获得这些数据集； (ii) 可用的损伤样本谱无法充分捕捉桥梁动态响应的损伤信息； (iii) 可能需要人工干预来区分深度学习模型输出的损害，这在处理大量监测数据时效率低下。为了应对这些挑战，本研究提出了一种用于桥梁损伤识别的智能无监督深度学习框架。该框架具有三个创新技术要素：（1）开发了具有混合损失函数的深度卷积自动编码器（CAE）模型，以提供无需大量损坏案例数据集即可识别桥梁损坏的智能系统； (2) 建立小波透射率模式谱，以更有效的方式表征动态响应中嵌入的损伤信息； (3)提出了一种基于排序点识别聚类结构(OPTICS)的损伤选择器，以实现损伤情况的自动判别。 该框架的可行性通过曲线桥损伤检测进行了数值论证，并在实验室规模的诱发损伤悬索桥上通过实验验证了其有效性。结果表明，所提出的框架可以自动、准确地识别桥梁的损坏。该框架为智能和数据驱动的桥梁损伤检测提供了解决方案。