Reconstruction of long-term strain data for structural health monitoring with a hybrid deep-learning and autoregressive model considering thermal effects

Complete data are essential for implementing reliable structural health monitoring (SHM); however, data loss due to equipment malfunction or other potential factors is unavoidable. Therefore, it is necessary to develop reliable structural-response-reconstruction methods. However, previous strain reconstruction methods have three shortcomings that lead to relatively low accuracy: 1) they typically do not use modules with better ability to capture the long- and short-term variation patterns of time series, such as the bidirectional gated recurrent unit (BiGRU) and convolutional neural network (CNN); 2) they do not simultaneously consider the spatiotemporal correlation among sensors and the strong correlation between temperature and strain, each of which has been demonstrated to contribute to improving the reconstruction accuracy separately; and 3) they do not carefully consider the linear and nonlinear correlations among SHM data. To address these problems, we proposed a strain reconstruction method that combines a nonlinear deep-learning (DL) component with a linear autoregressive (AR) component. The method also utilizes BiGRU and CNN in the DL component to capture the long- and short-term patterns of the SHM data better, as well as the two above-mentioned data correlations. To fully validate the performance of the proposed method, the long-term SHM data of a long-span steel box girder suspension bridge were utilized for validation. The results show that the hybrid DL and AR model can reconstruct the long- and short-term missing data with higher accuracy under different scenarios than previous models, such as the CNN.