基于卷积神经网络和迁移学习的结构损伤识别

针对传统基于机器学习损伤识别方法手工提取特征适应性差、识别能力弱等问题，提出一种基于卷积神经网络和迁移学习的新颖、快速结构损伤识别方法.首先根据损伤特征向量特点，提出原始信号的分帧处理流程；其次考虑多传感器数据融合要求，建立多通道一维卷积神经网络结构损伤识别模型，给出模型的整体流程和网络参数；然后采集不同通道和不同噪声水平下，模拟不同位置程度损伤的15层框架数值模型加速度数据，进行损伤识别；最后将网络模型进行迁移学习，对7层框架模型试验进行损伤识别，并验证所提方法的可行性、准确性和计算复杂性.结果表明，该方法实现了特征自适应提取、损伤位置和损伤程度的精准识别，具有突出的计算效率.

Structural damage detection based on convolutional neural networks and transfer learning

Conventional damage detection methods based on machine learning only use a small number of manually extracted features for recognition, and thus the damage recognition ability is weak. In order to solve such problems, a novel and fast structural damage identification method, using multi-channel one-dimensional (1D) convolution neural network (CNN), is presented in this study. Considering the characteristics of multi-acceleration sensor data fusion, a multi-channel 1D convolutional neural network structural damage identification model is established. According to the demand of damage feature vector, the frame processing flow of the original acceleration signal is proposed and the overall process and network parameters of the model are given. According to the requirements of the damage feature vector, the original signal frame processing process is proposed. A 15-layer framework numerical model is simulated with different positions and degrees of damage, and the acceleration data under different channels and noise levels are used to identify the damage. On this basis, the method was transplanted to a 7-layer frame model through transfer learning (TL). Finally, the feasibility, accuracy and computational complexity of the proposed method is analyzed. The results show that the method directly uses multi-channel acceleration time-domain signals to achieve feature adaptive extraction, thereby achieving more accurate identification of damage location and damage degree, and has outstanding computational efficiency.