The single index based on modal strain energy and modal curvature is proved to be effective for damage localization, while its recognition accuracy is not satisfied and can not directly reflect damage degree. Therefore, multilevel data fusion methods are proposed here with three new indexes of modal strain energy dissipation rate (DR), change rate of cross-model modal strain energy (CR), and difference of modal curvature ratio (RD). Firstly, first-level, second-level, and third-level data fusion methods are deduced based on Bayes theory, weighted average criterion, and BP neural network, respectively. The first three order modes data of each index are fused in the first level, and their results are further fused between indexes in the second level, which will be applied for damage position judgement; moreover, results of the second level as input are last fused in the third level in order to predict structural damage degree. Secondly, a simply supported beam and a three-span continuous girder models are simulated to verify effectiveness of multilevel data fusion indexes. It can be concluded that the recognition results agree well with man-made damage cases whatever the positions or degrees are. Finally, a test study on a simply supported steel beam with different damage forms is carried out. The results show that the proposed multilevel data fusion methods have good abilities of sensitivity and anti-interference, are fault-tolerant, have robustness, and would provide certain effective experience for actual damage identification.