Novel Mamba-based framework with enhanced generalization ability for multi-task dam response reconstruction

• The latest Mamba Model was applied to reconstruct dam data. • Mamba model performs excellently in reconstructing complex dam data with significant noise. • A reasonable time window size is recommended based on efficiency and accuracy analysis. The integrity and reliability of dam health monitoring data are fundamental to structural safety assessments. However, long-term operational dams often suffer from data loss, exacerbated by environmental noise in real-world conditions. Accurate data reconstruction is crucial for disaster early warning and engineering management. Recent advancements in deep learning have shown significant potential in this domain. The selection of deep learning models has increasingly become a case-specific engineering problem due to the lack of effective comparisons between different data reconstruction methods. There is no consensus on the most suitable model architecture, and the performance in feature extraction, reconstruction accuracy, and generalization capability remains inconsistently evaluated. This study proposes an innovative multi-task reconstruction framework based on the state-space architecture Mamba model. A comprehensive dam response dataset was constructed. This dataset incorporated factors including temperature, water level, and noise coupling. Subsequently, the study systematically compared the performance of Mamba and six other mainstream deep learning models in dam data reconstruction tasks to provide a basis for model selection in multi-task response reconstruction. Leveraging its unique state-space model characteristics and linear computational complexity, Mamba outperformed the six other models in reconstruction accuracy and generalization ability. Notably, in reconstructing data with complex nonlinear responses, the Mamba model demonstrated superior robustness, efficiently capturing long-term dependencies, with a coefficient of determination ( R ²) of 98% and a mean absolute error ( MAE ) below 0.18 mm. The Mamba model not only achieves high-precision structural response reconstruction but also exhibits excellent noise resistance, offering a novel solution for nonlinear data reconstruction.