基于双重注意力机制-改进Inception模块的CNN模型识别框架结构损伤

Damage identification of frame structure based on CNN model with dual-attention mechanism and improved Inception module

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,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。

Here,aiming at problems of having network hidden layers,enormous numbers of parameters and long training time in traditional deep learning methods,a frame structure damage recognition method based on convolutional neural network(CNN)model with dual-attention mechanism and an improved Inception module was proposed.Firstly,vibration response signals of a frame structure were converted into 2-D time-frequency diagrams with the local maximum synchro-squeezing transform.Then,a 2-D CNN model was built based on the improved Inception module,and the obtained 2-D time-frequency diagrams were taken as its input.Finally,highly correlated damage features in the CNN model were enhanced with a dual-attention mechanism to successfully identify location and level of damages to frame structure.The effectiveness of the proposed method was verified using Phase I numerical simulation data of IASC-ASCE SHM Benchmark Structure and Qatar University stand simulator dataset.The study results showed that this method can not only reduce number of model parameters and accelerate model’s convergence speed,but also have higher accuracy and stronger anti-noise performance when facing multi-class damage recognition problems in frame structures.