A multimodal fusion network guided by environmental semantics for reconstructing missing bridge strain trends

Long term bridge strain trends describe the gradual evolution of structural responses under environmental actions and operational loads, providing important information for structural health monitoring and condition assessment. However, communication interruptions, sensor abnormalities, and unstable data transmission can lead to missing strain data, which affects the completeness of monitoring records and the reliability of subsequent analysis. This study proposes an environmental semantics-guided multimodal fusion network for reconstructing missing bridge strain trends. The network mainly consists of an environmental semantic modulation module and a temporal reconstruction module. The environmental semantic modulation module encodes environmental textual descriptions into semantic representations and introduces environmental contextual information through feature-wise semantic modulation. The temporal reconstruction module uses a one-dimensional convolutional neural network (1D CNN) to extract local trend variation features and a modified Transformer encoder to model long-range temporal dependencies, enabling environment-aware strain trend reconstruction. Based on strain data obtained from a real bridge field monitoring experiment, this study constructs several typical missing scenarios and validates the proposed method through comparison with multiple baseline methods. The results show that the proposed model can effectively recover the original strain trends under different missing conditions and outperforms the compared methods in terms of reconstruction accuracy and robustness. Ablation experiments further demonstrate the effectiveness of the environmental semantic modulation and temporal reconstruction modules.