Data reconstruction leverages one-dimensional Convolutional Neural Networks (1DCNN) combined with Long Short-Term Memory (LSTM) networks for Structural Health Monitoring (SHM)

• Present methods and advantages and disadvantages of each method to overcome errors or data loss. • Propose a method to take advantage of the strengths of 1DCNN and LSTM to reconstruct lost sensor data in SHM. • Present the steps when using the proposed method to reconstruct sensor data. • The proposed method is performed on real data sets in the laboratory, and real bridge works are also used. The results show that the proposed method has good performance. SHM data collected in systems often face data loss due to transmission errors, sensor damage, or environmental impacts. Incomplete data can lead to erroneous assessments in evaluating structural safety in complex structures. Although data reconstruction has been studied, challenges are present in data reconstruction: (i) SHM data contains a large amount of noise; (ii) data structure is complex and doesn’t allow for simple linear or nonlinear formulation; (iii) reconstructed data needs to be accurate and reliable. This study proposes a hybrid deep learning approach combining the 1DCNN and LSTM network to reconstruct data within an SHM environment. The proposed model uniquely leverages 1DCNN for efficient spatial feature extraction and LSTM for capturing long-term temporal dependencies. Input data is strategically preprocessed through correlation-based sensor clustering and time-shift enhancement techniques. A hybrid model used the SHM data measurements before data loss to train models. The trained hybrid network can then reconstruct missing or erroneous data. The proposed method is validated on real datasets from different structures in various scenarios and can be applied in practice, achieving better performance and accuracy compared to other neural network-based methods. Quantitative results show that the hybrid model reduces the Mean Absolute Error (MAE) by 10–15% and achieves Modal Assurance Criterion (MAC) values exceeding 0.95, outperforming other baseline neural network models. These results highlight the model’s practical applicability for accurate SHM data reconstruction under both single- and multi-channel sensor failures.