基于卷积神经网络的新奇检测技术在结构损伤识别中的应用

Application of Novelty Detection Technology Based on Convolutional Neural Network in Structural Damage Identification

针对新奇检测难以同时识别结构损伤时刻和损伤位置的问题,提出在新奇检测中引入卷积神经网络以实现损伤时刻和损伤位置的一次性确定。首先,采用小波包技术处理结构响应得到小波包能量,并将相邻测点对应频带的能量比作为新奇检测模型的特征向量;然后,以结构健康时的特征向量作为训练数据,建立健康模式下的基于卷积神经网络的新奇检测模型;接着,将结构实时输出的特征向量输入新奇检测模型,所得输出与健康状态的输出进行对比,并将输出和输入的欧氏距离作为新奇指标;最后,根据新奇指标的变化识别结构损伤时刻和损伤位置。数值模拟和实验室试验验证了该方法的有效性。

To solve the problem that novelty detection is difficult to identify the time and location of structural damage at the same time,convolutional neural network was introduced into novelty detection.Firstly,the wavelet packet technology was used to process the structural response to obtain the wavelet packet energy,and the energy ratio of the corresponding frequency bands of adjacent measurement points was applied as the feature vector of the novelty detection model.Then,taking the eigenvector of the healthy structure as the training data,the novelty detection model based on convolutional neural network under the health mode was established.Next,the real-time output feature vectors of the structure was input into the novelty detection model,the output with was compared the output of the health state,and the Euclidean distance between the output and input was used as the novelty index.Finally,the time and location of structural damage based on changes were identified in novelty index.The effectiveness of the method herein were verified through numerical simulation and laboratory experiments.