为了提高损伤识别中单一指标对损伤的灵敏度和抗噪能力，基于模态应变能理论，提出联合主成分分析（principal component analysis, PCA）和熵权融合（entropy weight method, EWM）的两级特征融合方法，并使用北方苍鹰优化算法（northern goshawk optimization, NGO）结合门控循环单元（gated recurrent unit, GRU）进行桥梁损伤程度预测。首先，基于传统的模态应变能理论，构造出对角模态应变能比，由此衍生出对角模态应变能比变化率，对角模态应变能比耗散率，标准化对角模态应变能比差指标。其次，使用主成分分析实现指标内特征提取，熵权法融合指标间的特征，从而构造出加权决策指标（weighted decision index, WDI）。将单个模态应变能衍生指标输入到NGO-GRU混合神经网络中，损伤程度为输出，从而建立指标值与损伤程度之间的关系，进而实现损伤量化。通过三跨连续梁桥数值模型对所提出的方法进行验证，结果表明：加权决策指标具有良好的损伤定位能力和抗噪性，混合神经网络具有较高的损伤预测精度，预测准确率为91.14%。

In order to improve sensitivity and noise immunity of a single index to damage in damage identification, based on modal strain energy theory, a two-level feature fusion method combining principal component analysis (PCA) and entropy weight method (EWM) were proposed. The northern goshawk optimization (NGO) algorithm combined with gated recurrent unit (GRU) were used for bridge damage degree prediction. Firstly, based on traditional modal strain energy theory, the diagonal modal strain energy ratio was constructed, and then change rate of the diagonal modal strain energy ratio, dissipation rate of the diagonal modal strain energy ratio, and normalized difference index of the diagonal modal strain energy ratio were derived. Secondly, principal component analysis was used to extract features within the index, and entropy weight method was used to fuse features between indexes. Finally, weighted decision index (WDI) was constructed. The single modal strain energy derivative index was input into the NGO-GRU hybrid neural network, as well as damage degree was output, so as to established the relationship between index value and damage degree, and then realized damage quantification. The proposed method was verified by a three-span continuous beam bridge numerical model. The results show that weighted decision index has good damage location ability and noise immunity. The hybrid neural network has high damage prediction accuracy, with a prediction accuracy rate of 91.14%.