An interpretable multi-scale deep learning framework for dam displacement prediction and automated early warning of structural health monitoring

Accurate prediction of displacement and early warning are fundamental to dam structural health monitoring (SHM). Hydrostatic, thermal, and aging-related environmental drivers affect dam displacements at different response rates, leading to heterogeneous hysteresis effects. Existing models often rely on fixed or single-scale temporal representations, or introduce historical displacement observations as inputs, thereby limiting their ability to distinguish short-term hydrostatic responses, thermal hysteresis effects, and long-term cumulative deformation, and weakening the independence of predicted displacements for early warning. This study develops an interpretable environmental factor-driven deep learning (DL) framework, termed past decomposable mixing with gated attention residual (PDM-GAR) network, for dam displacement prediction and early warning. The PDM module is adopted to decompose historical environmental sensing data into multiple temporal scales to represent environmental evolution at different temporal resolutions. The GAR module integrates gated recurrent units, multi-head attention, and residual connections to directly predict future dam displacement from environmental monitoring data. PDM-GAR was trained and evaluated using 11 years of monitoring data from a concrete gravity dam, reducing the prediction errors by up to 56.3% compared to statistical and DL baseline models. Prediction residuals were further modeled using kernel density estimation to establish multi-level warning thresholds. In addition, perturbation experiments were performed to quantify the impact of input anomalies and to identify sensitive variables and timesteps. These results demonstrate that PDM-GAR provides a reliable prediction-warning-interpretation framework for dam SHM.