针对传统深度学习方法的网络隐含层和参数异常庞大且训练时间较长的特点，本文提出了一种基于双重注意力机制和改进Inception模块的卷积神经网络（convolutional neural network，CNN）模型来识别框架结构损伤。首先，通过局部最大值同步挤压变换将结构的振动响应信号转化为二维时频图并作为卷积神经网络的输入，然后基于改进Inception模块搭建二维卷积神经网络，最后通过双重注意力机制增强相关度高的损伤特征从而成功识别结构的损伤位置和损伤程度。通过IASC-ASCE SHM Benchmark结构I阶段数值模拟数据和卡塔尔大学看台模拟器数据集验证所提方法的有效性，研究结果表明：该方法不仅可以减少模型参数的个数和加快模型收敛速度，而且在面对框架结构多类别损伤识别问题时具有较高的准确率和较强的抗噪性能。

In order to address the issues of enormous numbers of parameters and long training time in traditional deep learning methods, a new structural damage identification method based on convolutional neural network (CNN) model with dual attention mechanism and improved Inception module is proposed. Firstly, the vibration response signals from engineering structures are firstly transformed into two-dimensional time-frequency spectrograms using local maximum synchrosqueezing transform, and then the obtained time-frequency spectrograms are used as inputs of CNN. Secondly, a two-dimensional CNN is established based on an improved Inception module. Finally, a dual attention mechanism is introduced to explore damage features with higher relevance, leading to a successful identification of structural damage location and severity. The effectiveness of the proposed method is validated by two numerical simulation cases of IASC-ASCE SHM Benchmark Phase I model and the Qatar University Grandstand Simulator dataset. The results demonstrate that the proposed method not only reduces model parameters and accelerates model convergence, but also has high accuracy and robust noise resistance in the context of multi-category damage identification on frame structures.