Self-supervised deep learning for gnss time series imputation: a comparative study of neural network architectures

Global Navigation Satellite System (GNSS) time series are widely used for structural health monitoring (SHM) and deformation analysis, but real-world recordings frequently contain short and long contiguous gaps that degrade downstream interpretation. This study addresses the challenge of accurate imputation for GNSS displacement series by proposing a self-supervised learning framework that trains models directly on real, unlabelled data using contiguous-span masking. We evaluate four neural architectures (ANN, CNN, GRU, LSTM) under a unified pipeline comprising signal denoising (moving-average, Kalman smoothing, Haar wavelet), sliding-window segmentation, Z-score normalization, and middle-region masking. Experiments use a year-long 10-minute-sampled dataset from the Can Tho cable-stayed bridge (sensor can519501, x/y/z components) and assess reconstruction quality via R², MAE, and MSE on withheld masked segments. Results indicate that recurrent architectures, particularly LSTM, produce the most faithful reconstructions: LSTM attains the highest validation R² (≈0.948) and the lowest MAE (≈0.137) and MSE (≈0.052) among tested models, while GRU offers competitive performance and CNN/ANN show substantially weaker recovery. These findings demonstrate that masking-based self-supervision is an effective strategy for GNSS gap recovery and that LSTM-like sequence models are well suited to capture the long-range temporal dependencies in bridge displacement data. The proposed approach enhances the reliability and continuity of GNSS-derived time series for structural monitoring and can inform future multi-sensor fusion and uncertainty quantification work