Hybrid Probabilistic Deep Learning for Damage Identification

In structural health monitoring, various types of sensors collect a large amount of data for structural defect detection. These data provide critical support for the application of machine learning for structural damage identification. However, machine learning relies heavily on training data, whose quality and distribution can affect the effectiveness of detection models in real-world damage identification. In addition, machine learning contains a large number of parameters that are highly uncertain, which results in the output of machine learning models is not always as reliable. These deterministic deep networks usually make overconfident decisions in some data. The ability of deep learning to provide safe and reliable decisions is very important when applied in the field of engineering. In order to ensure the decision security of machine learning models, this paper proposes a hybrid probabilistic deep network for structural damage identification. The proposed method converts deterministic weights into a Gaussian distribution, which in turn quantifies the uncertainty in machine learning. Among them, variational inference is used for uncertainty modeling of probabilistic deep networks. These uncertainty metrics can be used to determine whether the output of the machine learning model is reliable. Nevertheless, the introduction of uncertainty weakens the learning ability of deep networks. Meanwhile, the number of parameters in the probabilistic layer is twice that of the deterministic layer for the same architecture. Therefore, probabilistic deep learning is more difficult to train compared to deterministic deep learning. To address these issues, deep learning with hybrid probabilistic and non-probabilistic layers needs to be investigated. This paper analyzed and discussed the effects of different numbers of probability layers on the effectiveness of structural damage identification. Finally, a series of experimental results showed that the proposed method is able to accurately identify structural damage while quantifying the decision uncertainty.