Missing data imputation in tunnel monitoring with a spatio-temporal correlation fused machine learning model

Imputing missing values in structural health monitoring (SHM) data is essential for conducting data-driven analyses of tunnel structural stability. However, SHM data is dynamically changing with complex spatio-temporal correlations, making it particularly challenging to impute, especially for continuous or peak missing data. Therefore, this study aims to present a spatio-temporal correlations fused machine learning model (ST-ML) to accurately impute different forms of missing SHM data. Unlike existing methods, this approach considers various physical scenarios by incorporating Gaussian distribution and employing a bidirectional recursion structure to enhance the robustness of the model. Consequently, a series of ablation experiments and comparative analyses were conducted using SHM data from a case study to evaluate the rationale and necessity of the ST-ML model, as well as its imputation ability across datasets with varying levels of missing data. The experimental results demonstrate a significant improvement in imputation accuracy, with MAE and RMSE values reduced by at least 1.51 and 2.27, respectively, after considering spatio-temporal correlations and diverse physical cases. Moreover, the proposed model outperformed the commonly used models even under special cases, where the average imputation error was reduced by at least 1.15. These findings affirm the reliability of the proposed model.