A time–frequency heterogeneous dual-domain joint sampling–reconstruction and uncertainty quantification network for structural monitoring data reconstruction

Structural health monitoring (SHM) systems generate massive volumes of data during long-term operation, imposing substantial challenges on storage and transmission. Compressed sensing, which leverages sparse sampling combined with deep-learning-based reconstruction, has emerged as an effective means to alleviate these bottlenecks. However, existing methods often suffer degraded reconstruction fidelity at high compression ratios or when limited to a single domain (time or frequency), and they typically lack sampling strategies tailored to inherent signal characteristics as well as principled quantification of reconstruction uncertainty—limitations that undermine their reliability in practical engineering applications. To address these shortcomings, this study proposes a time–frequency heterogeneous, dual-domain joint sampling–reconstruction and uncertainty-quantification network specifically designed for SHM data. The model performs joint sampling and reconstruction in separate time-domain and frequency-domain branches and fuses their complementary outputs through a designed multi-scale weighted fusion module to produce the final reconstruction. Uncertainty quantification of the reconstructed signals is obtained naturally via repeated Monte Carlo forward passes. The proposed method is validated on laboratory pedestrian-bridge tests and on an in-service steel truss bridge. Results show that, across a range of compression ratios, the method achieves excellent reconstruction performance in both time and frequency domains, with modal parameters of the reconstructed signals essentially matching those of the originals. Compared with conventional random undersampling and single-domain reconstruction, the network-learned dual-domain undersampling strategy yields higher reconstruction accuracy, reduced variability, and improved robustness. Furthermore, a parallel dual-channel sampling scheme designed for sensor-level deployment enables in situ compressed acquisition and high-fidelity downstream reconstruction in real SHM environments.