Deep learning-based unsupervised methods for real-time condition monitoring of structures: a state-of-the-art survey

Real-time unsupervised condition monitoring of civil infrastructures has gained a great deal of attention during the past decade. This practice has been challenged by several factors such as the lack of a robust feature extraction strategy, scarcity of baseline data collected from the intact structure, the lack of information from missing data, and the hardship of specifying a dynamic threshold strategy. Thanks to the advances in deep learning techniques, the condition monitoring practice of civil infrastructures benefits largely from the strength of deep learning for feature extraction, amending missing information, and developing dynamic threshold settings. This survey studies some of the recent advances in real-time unsupervised condition monitoring of civil infrastructures. As such, it has been noted that the variational auto-encoder and generative adversarial networks are two main deep learning models that can address the aforementioned challenges. Therefore, a possible future path for research in this field can be towards mixing these deep learning models to address all the challenges of real-time unsupervised condition monitoring of civil infrastructures at once.