基于卷积神经网络和递归图的桥梁损伤智能识别

Intelligent Damage Detection for Bridge Based on Convolution Neural Network and Recurrence Plot

为改进目前传统损伤识别方法对桥梁局部小损伤识别能力较弱的不足,提出利用深度学习方法中的卷积神经网络对桥梁损伤进行统计模式识别.根据卷积神经网络对损伤特征向量的需求,将车桥耦合振动下的原始结构响应信号进行小波包滤波和重构,之后通过递归分析获取不同损伤工况的递归图,将其作为新型的损伤特征图像作为卷积神经网络的输入.在此基础上提出基于卷积神经网络和递归图的桥梁结构损伤识别计算流程和方法.对一座连续梁桥进行不同位置和程度的损伤模拟,提取小波包频带能量及递归图等损伤特征向量,并进行基于多种统计模式识别算法的损伤识别.结果表明:与其他特征向量相比,递归图蕴含更丰富的损伤信息;与支持向量机和BP神经网络等传统统计模式识别方法相比,卷积神经网络能够通过逐层智能学习实现更准确的特征自动提取和区分,从而实现损伤位置和损伤程度的更精准识别.

To improve the deficiencies of the current traditional damage detection methods,such as the inadequate detection ability for local damage and minor damage,the convolution neural network based on deep learning method is proposed to improve statistical damage pattern recognition for bridge.According to the demand of the damage characteristic vectors of the convolutional neural network,the original structural responses of vehicle-bridge coupling vibration are regenerated by wavelet packet filtering and reconstruction,and the recurrence plots of different damage cases are obtained as the damage characteristics images and the input of convolution neural network according to recurrence analysis.On this basis,the corresponding calculation process and method for damage detection of bridge structure based on convolution neural network and recurrence plot are established.The damage simulation for a continuous beam bridge is carried out,including damage locations and degrees.The damage feature vectors such as frequency band energy of wavelet packet and recurrence plots are extracted,and damage detection based on multiple statistical pattern recognition algorithm is carried out.The results showed that the damage information in recurrence plots is more abundant compared with other feature vector,and the convolutional neural network can fulfill automatic extraction and distinguishing feature more accurately according to layer-by-layer intelligent learning,compared with the traditional damage detection methods,such as support vector machine and BP neural network.Hence,the damage detection method based on convolution neural network and recurrence plot can achieve more identification accuracy for damage location and degree.