Missing Data Recovery for Dam Safety Monitoring via Hierarchical Clustering-Assisted Probabilistic Principal Component Analysis

Missing data is inevitable in dam monitoring systems and poses significant challenges to dam safety monitoring and early warning. Such missing values typically manifest in either discrete or continuous patterns. However, conventional imputation methods are often inadequate for scenarios in which both discrete and continuous missing patterns occur together within structural responses, particularly when external environmental covariates, such as reservoir water level and air temperature, are also unavailable. To address this limitation, a hybrid unsupervised framework called hierarchical clustering-assisted probabilistic principal component analysis is proposed. Within this framework, Ward’s hierarchical clustering is first applied to discern latent spatiotemporal patterns and facilitate adaptive clustering of monitoring variables. Subsequently, probabilistic principal component analysis models are constructed within each cluster to enable probabilistic reconstruction of both discrete and continuous missing entries. Benefiting from its probabilistic latent variable formulation, the method not only yields accurate point estimates but also quantifies residual noise uncertainty through the estimated isotropic variance, thereby providing a measure of estimation reliability. The proposed method was systematically validated using incomplete radial deformation monitoring data from an arch dam across four representative scenarios that combine varying discrete missing ratios with continuous block gaps. The results show that the proposed method consistently outperforms benchmark techniques in reconstruction accuracy and pattern preservation. Without requiring external environmental data, the framework offers a reliable solution for dam safety monitoring data reconstruction, supporting structural condition assessment and safety management.