Vehicle-assisted bridge damage identification has great application potential， but it is still difficult to extract damage-sen‐ sitive features from multi-source monitoring data and accurately evaluate the bridge damage status. To solve this problem， an At‐ tention-LSTM-based Feature Fusion Model （ALFF-Net） is proposed. The model improves the perception ability of Bi-LSTM cells for multi-scale feature information in time series data through a preset data reconstruction layer. Furthermore， by employing attention mechanism and feature fusion strategy， the model reduces the prediction difficulty of downstream branches of deep neural networks and further improves the modeling ability for the important dependency relationships in the sequence data. A monitoring dataset under different road roughness and vehicle speeds is generated through a vehicle-bridge interaction system simulation， and the bridge damage identification performance of the ALFF-Net model is comprehensively tested. The results show that the ALFF-Net model improves the damage identification accuracy by up to 19.30% compared to the classical LSTM network while signifi‐ cantly reducing computational costs， and the identification errors under different road roughness levels are less than 3%. More‐ over， by comparing the identification accuracy of the ALFF-Net model under different data-driven schemes， the robustness of the bridge damage detection results with synergistic multi-source monitoring data is verified.