An integrated Conv1D–ResNet–GRU architecture for robust vibration data recovery in structural health monitoring

SHM systems often struggle with missing and noisy data, which hinders their ability to detect structural damage. Standard deep learning (DL) approaches have a split limitation: Convolutional Neural Networks (CNNs) excel at capturing local signal details but struggle with long-term trends, while Recurrent Neural Networks (RNNs) model temporal dependencies effectively but often overlook fine-grained local features. This study proposes an integrated Conv1D–ResNet–GRU architecture that explicitly addresses this tradeoff for vibration data recovery. In the proposed hierarchical design, Conv1D layers first extract fine-grained local features from raw acceleration signals; a ResNet-style residual block with a projection shortcut then facilitates gradient propagation and preserves informative representations; finally, stacked GRU (Gated Recurrent Unit) layers capture long-range temporal dependencies and produce multi-step reconstructions. The framework is evaluated on a laboratory-scale cable-stayed bridge model under realistic SHM scenarios, including missing-data ratios (5–15%), and additive Gaussian noise levels from 0% to 40%. A comprehensive comparison is conducted against baseline Conv1D and GRU models and two ablation variants (Conv1D–GRU and Conv1D–ResNet) trained under identical training settings. Across all forecasting horizons, missing-data levels, and noise conditions, the proposed Conv1D–ResNet–GRU outperforms the compared models, achieving lower RMSE and MAE and higher R2. These results demonstrate that combining residual convolutional blocks with GRU-based temporal modeling offers a robust and effective solution for vibration data reconstruction in SHM applications.