Lightweight Evidential Time Series Imputation Method for Bridge Structural Health Monitoring

Long-term data loss resulting from sensor malfunctions, communication interruptions, and other factors in Structural Health Monitoring (SHM) significantly undermines the reliability of damage identification and safety assessment. Existing methods—ranging from statistical approaches and low-rank matrix completion to traditional machine learning and deep learning imputation techniques—often suffer from either limited accuracy or excessive model size and slow inference, making deployment in resource-constrained scenarios difficult. To address these challenges, this paper proposes TEFN–Imputation, a lightweight and efficient time-series imputation model. This model utilizes observation-driven non-stationary normalization to mitigate the impact of time-varying characteristics and dimensional discrepancies. It employs linear projection for temporal length alignment and constructs BPA-style mass representations from dual perspectives of time and channel. Furthermore, it replaces strict Dempster–Shafer belief combination with an expectation-based evidential aggregation (readout), thereby significantly reducing computational overhead while enabling uncertainty-aware evidential indicators for interpretation rather than claiming a direct accuracy gain from uncertainty modeling. The observed accuracy and robustness improvements are primarily attributed to the normalization and dual temporal–channel modeling design under the same lightweight readout. Systematic experiments on two real-world bridge monitoring datasets, Z24 and Hell Bridge, demonstrate that TEFN consistently maintains low Mean Absolute Error (MAE) and minimal volatility across various combinations of training and testing missing rates, exhibiting high robustness against variations in missing rates and train–test mismatches. Concurrently, compared to RNN and large-scale Transformer baselines, TEFN reduces parameter count and CPU inference time by one to two orders of magnitude. Thus, it achieves a superior trade-off among accuracy, efficiency, and model scale, making it highly suitable for online SHM and imputation tasks in practical engineering applications. Across the settings on Z24, TEFN achieves a mean MAE of 0.218 with a standard deviation of 0.002, while using only 0.02 MB parameters and 2.73 ms per batch CPU inference.