基于时频变换和卷积神经网络的结构损伤识别∗

Structural Damage Identification Based on Time-frequency Transform and Convolutional Neural Network

摘要:

为了解决将单传感器时域数据直接作为卷积神经网络（CNN）的输入所引起的损伤识别精度不高的问题，提出基于小波包变换（DWPT）和快速傅里叶变换（FFT）的卷积神经网络识别方法。以短钢梁桥现场试验测得的数据集为例，将单传感器数据样本分别进行 DWPT 和 FFT 变换，使用变换后的特征训练 1D‐CNN 网络，训练好的网络测试精度有明显的提升，其识别精度均高于多个传感器数据直接作为输入的识别精度。同时分析了对噪声样本和异源（结构上未曾参与网络训练的传感器）数据的识别情况，结果表明对含噪声样本先进行时频变换再训练网络能显著提升对噪声样本的识别精度，而且能改善训练好的网络难以对异源传感器数据进行识别的问题，最后通过卡塔尔大学看台现场试验数据进一步论证上述结论。

Abstract:

To solve the problem of low accuracy in damage identification caused by using single sensor time-domain data directly as input for a convolutional neural network， a convolutional neural network identification method based on wavelet packet transform and fast Fourier transform is proposed. Taking the dataset measured in the field experiment on a short steel bridge as an example， single sensor data samples are transformed using DWPT and FFT respectively， and the transformed data are used to train the 1D-CNN network. The test accuracy of the trained network is obviously improved， and its recognition accuracy is higher than that of multiple sensor data directly as input. At the same time， the recognition of noise samples and data from sensors that have not participated in network training in structure is analyzed. The results show that the time-frequency transformation of noise samples before training the network can significantly improve the recognition accuracy of noise samples， and can improve the problem that the trained network is difficult to identify data from the sensor that have not participated in network training. Finally， the above conclusions are further demonstrated by test data from the grandstand of Qatar University.