利用单传感器数据基于GAF-CNN的结构损伤识别

Structural Damage Identification Using Single Sensor Data Based on GAF‑CNN

为了减少损伤识别所需传感器数量,降低监测系统造价及海量数据的处理成本,提出了基于单传感器数据结合格拉姆角场(Gramian angular field,简称GAF)和卷积神经网络(convolutional neural networks,简称CNN)的结构损伤识别方法。采用GAF理论将原始振动信号分别转换为格拉姆角和场(Gramian angular summationfield,简称GASF)和格拉姆角差场(Gramian angular difference field,简称GADF)二维图像,以转换后的GASF和GADF两类图像数据集为输入,基于LeNet⁃5结构下的浅层卷积神经网络模型,训练最优二维CNN模型用于结构损伤识别。以国际桥梁维护和安全协会提出的结构健康监测基准模型结构及一榀钢框架结构为例,研究振动信号转化为二维图像算法、卷积神经网络模型参数、传感器布置位置及测量噪声对识别结果的影响。结果表明:所提算法仅需单个传感器数据即可实现损伤识别的目的,数值模拟及模型试验的损伤识别准确率均为100%,单条样本测试时间为8.5 ms左右,满足结构健康监测在线损伤识别的需求,且受传感器布置位置和噪声程度影响较小;GADF图较GASF图收敛效率更高,震荡幅度更小,受局部最优值影响较小,在样本数量规模一致的状态下,更易训练生成最优二维CNN模型。

A structural damage identification method using single sensor data based on Gramian angular field(GAF)and convolutional neural networks(CNN)is proposed,which can reduce the number of sensors needed for damage identification,and the cost of the monitoring system and the processing cost of massive data.In this method,the original vibration signal is converted into 2-dimensional images of the Gramian angular summation field(GASF)and the Gramian angular difference field(GADF)based on GAF theory respectively.The image features are extracted and input to the optimized 2-dimensional CNN mode based on LeNet-5 Structure,and the ideal damage identification result is finally obtained.The international association for bridge maintenance and safety bridge health monitoring benchmark structure and a steel frame structure are used to verify the effective⁃ness of the proposed method.Also the influences of the algorithm for vibration signal transformation into two-di⁃mensional image,convolutional neural network model parameters,sensor location and measurement noise on the damage identification results are analyzed.The results show that the proposed algorithm only needs single sensor data to achieve the purpose of damage identification,and the accuracy of damage identification is 100%in both numerical simulation and model test.The test time of a single sample is about 8.5 ms,which meets the re⁃quirements of online damage identification for structural health monitoring,and it is less affected by sensor loca⁃tion and noise level.Secondly,compared with GASF graph,GADF graph has higher convergence efficiency,smaller oscillation amplitude and less influence by local optimal value.Under the condition of the same sample size,it is easier to train and generate the optimal 2-dimensional CNN model.