A spatiotemporal recurrent neural network for missing data imputation in tunnel monitoring

Given the swift proliferation of structural health monitoring (SHM) technology within tunnel engineering, there is a demand on proficiently and precisely imputing the missing monitoring data to uphold the precision of disaster prediction. In contrast to other SHM datasets, the monitoring data specific to tunnel engineering exhibits pronounced spatiotemporal correlations. Nevertheless, most methodologies fail to adequately combine these types of correlations. Hence, the objective of this study is to develop spatiotemporal recurrent neural network (ST-RNN) model, which exploits spatiotemporal information to effectively impute missing data within tunnel monitoring systems. ST-RNN consists of two moduli: a temporal module employing recurrent neural network (RNN) to capture temporal dependencies, and a spatial module employing multilayer perceptron (MLP) to capture spatial correlations. To confirm the efficacy of the model, several commonly utilized methods are chosen as baselines for conducting comparative analyses. Furthermore, parametric validity experiments are conducted to illustrate the efficacy of the parameter selection process. The experimentation is conducted using original raw datasets wherein various degrees of continuous missing data are deliberately introduced. The experimental findings indicate that the ST-RNN model, incorporating both spatiotemporal modules, exhibits superior interpolation performance compared to other baseline methods across varying degrees of missing data. This affirms the reliability of the proposed model.