Long-term structural health monitoring strain modeling for bridges via data correlation and deep learning

Abstract Many of previous long-term strain modeling methods for bridges mechanically various data correlations, limiting their applicability to specific types of bridges. This study develops a more generalizable approach by quantitatively analyzing correlations between data using Pearson correlation coefficient (PCC) analysis to rationally select input data. A hybrid DL and autoregressive model with attention mechanism (DL-AR-ATT) model is employed as modeling tool, which has been demonstrated to outperform most of the state-of-the-art deep learning models. This study utilizes measured strain data from both concrete and steel box-girder bridge, along with air temperature data, to verify the modeling and generalizability performance of the proposed method. The results indicate that the combination of air temperature (AT) and atmospheric pressure is the optimal input combination for long-term strain modeling in bridge structures. However, when strain data is also included as input, the combination of strain data (with stronger correlations to the target modeling strain) and AT is the optimal input combination. Furthermore, the proposed method outperforms existing long-term strain response modeling methods for bridges in terms of modeling and generalizability performance.