为解决结构的健康监测问题,找到合适的结构损伤识别特征,使用卷积神经网络提取结构特征来识别损伤,并通过IASC-ASCE SHM Benchmark第一阶段模拟数据验证其有效性,同时与小波包频带能量特征、前五阶本征模态函数能量特征做同分类器准确率对比,证明了卷积神经网络在自动提取特征方面的优势.在分析卷积神经网络自动提取特征的鲁棒性时,发现单一噪声数据训练的特征抗噪能力有一定局限性,为了获得更好的特征抗噪能力,提出混合噪声训练模式,验证了含噪声0％ ～ 50％的样本数据,均取得良好识别结果.同时在进行卷积核特征可视化工作中发现,混噪模式训练的卷积核能够识别更多阶次的频率信息.%Here,a convolution neural network was used to extract structural features,identify damage and solve problems of structural damage identification. The effectiveness of this method was verified with IASC-ASCE SHM Benchmark Phase 1 simulation data. Then,comparing the same classifier accuracies for energy characteristics of the convolution neural network,the wavelet packet and the first 5 IMFs obtained by EMD,advantages of the convolution neural network in automatically extracting features were proved. In analyzing the robustness of features'automatic extraction of the convolution neural network,it was found that the characteristic anti-noise ability of a single noise data training mode is limited. In order to acquire the better characteristic anti-noise ability,a mixed noise training mode was proposed. The validity of this training mode was verified using the sample data with noise of 0％-50％ to obtain good recognition results. At the same time,it was found in visualization of the convolution's kernel features that the convolution kernel of the mixed noise training mode can identify more orders of frequency information.
展开▼
