Model-informed deep learning strategy with vision measurement for damage identification of truss structures

Structural damage identification approaches can be divided into two categories, i.e. data-driven approaches via statistical pattern recognition and model-based approaches via finite element (FE) model updating. These two approaches have their own merits, and their main shortcomings can be remedied by each other’s merits. Therefore, this study proposed a deep learning-based damage identification strategy involving both data-driven and model-based approaches, termed as model-informed deep learning (MIDL)-based strategy. This strategy first proposes a vision-based displacement estimation approach to extract structural displacement responses from video data. This approach reduces the displacement drift induced by conventional optical flow approaches and improves the tracking accuracy of feature points. Then, a calibrated FE model is built to construct data sets with different damage levels via FE model updating and time-history analysis. Following this, a one-dimensional convolutional neural network (1D CNN) is established to detect and localize structural damage by using direct displacement responses. Finally, FE model updating is performed again to quantify structural damage level with constrained targets. A truss structure is further used to evaluate the accuracy of the proposed strategy experimentally. Results illustrate that the proposed MIDL strategy, which uses time series to localize the structural damage, achieves a global location accuracy of 86.09% and avoids the feature extraction process. Meanwhile, with the known damage location, the efficiency and accuracy of structural damage quantification via rerunning model updating can also be significantly improved. In addition, the displacements estimated by the proposed approach have a good match with ground truth values with error standard deviations of less than 0.3 mm.