Quantifying Uncertainty in Structural Damage Identification Using Bayesian Convolutional Neural Networks

Damage identification is a crucial component of structural health monitoring, and convolutional neural networks (CNNs) have been extensively applied as an efficient tool in structural damage detection. However, the performance of traditional CNNs is not entirely satisfactory when addressing the inherent uncertainties, such as measurement noise and modeling inaccuracies, in structural damage identification. A damage identification method based on Bayesian convolutional neural networks (BCNNs) is proposed in this study. Network parameters in the proposed approach are probabilistic distributions rather than fixed values. After training, the network can generate different outputs for the same input based upon repeated inferences. This leads to a statistical output distribution, offering a more reliable reflection of structural damage identification by incorporating uncertainty considerations. Furthermore, the proposed approach utilizes phase space matrix as input, instead of simply employing modal information or dynamic responses, making the network more sensitive and requiring only a few sensor measurements. Numerical studies on a frame structure are conducted. The results demonstrate the effectiveness of the proposed approach. The statistical properties of output effectively quantify the uncertainties inherent in the damage identification results.