A feature-enhanced deep learning-based framework for accurate reconstruction of acceleration responses in super-tall buildings

Structural Health Monitoring plays a critical role in ensuring the safety and long-term performance of high-rise buildings; however, its effectiveness is often undermined by sensor malfunctions, data loss, and the strongly non-stationary nature of structural responses. Existing reconstruction methods typically rely on raw acceleration signals without a robust feature-selection process, making them sensitive to noise and incapable of maintaining accuracy when only a limited number of functional sensors remain. To overcome these limitations, this study introduces a feature-enhanced deep learning framework for reconstructing acceleration responses with major sensor loss. The method decomposes available measurements into Intrinsic Mode Functions using Empirical Mode Decomposition and applies Power Spectral Density – based correlation analysis to identify the most informative components. These features are then used to train a Long Short-Term Memory model whose hyperparameters are optimized through the Observer – Teacher – Learner-Based Optimization algorithm. The proposed approach is validated using acceleration data from Milad Tower in Tehran, equipped with 11 sensors. Results demonstrate that the framework can accurately reconstruct responses even when data from up to 10 sensors are missing. Performance metrics confirm its robustness, indicating that the method offers a reliable solution for enhancing SHM systems exposed to sensor failures and data loss.