基于GCN的多源变工况滚动轴承故障诊断

Multi-source rolling bearing fault diagnosis under variable working conditions based on GCN

滚动轴承是旋转机械的关键部件,其健康状况的识别非常重要。迁移学习作为一种有效工具被广泛应用于故障诊断领域,但单源迁移学习方法可能存在泛化性能较差甚至引起负迁移,造成识别效果不佳的问题。提出一种基于多感受野图卷积网络(GCN)的多源迁移学习方法(MS-GCN),通过在多个源域数据上学习迁移知识,实现变工况下滚动轴承的故障诊断。该方法首先利用小波变换将振动数据样本转换为二维时频图样本,将获得的N组源域样本和目标域样本进行构建得到N组源域-目标域样本数据对;其次,先利用深度卷积网络学习每组数据对的高维特征,再由多感受野图卷积网络学习所提特征的数据结构,使得自适应方法能充分学习域不变特征,更有效地将源域与目标域特征进行对齐,训练得到N组分类器;最后,取N组分类器分类结果的平均值为目标域样本的状态识别结果。基于江南大学轴承数据集对所提方法展开实验验证,在3组不同的变工况轴承故障诊断任务中,所提方法对4种不同状态(正常、内圈故障、外圈故障及滚动体故障)的分类准确率均在99%以上,与其他方法相比诊断准确率提升了0.22~8.27个百分点。对比结果表明:所提方法对变工况下滚动轴承的故障进行识别,可以有效地诊断出轴承的故障类型,具有一定的工程实用价值。

Rolling bearing is a key component of rotating machinery,and its health status identification is very important.Transfer learning is widely used as an effective tool in the field of fault diagnosis,but the single-source migration learning method may lead to poor generalization performance or even cause negative migration,resulting in poor recognition results.In this paper,a multi-source transfer learning method(MS-GCN)based on multireceptive field graph convolutional network(GCN)was proposed.By learning transfer knowledge on multiple source domain data,the fault diagnosis of rolling bearings under variable working conditions was realized.Firstly,the vibration data samples were converted into two-dimensional time-frequency diagram samples by wavelet transform.N sets of source domain samples and target domain samples were constructed to obtain N sets of source domain-target domain sample data pairs.Secondly,the deep convolution network was used to learn the high-dimensional features of each set of data pairs.Then,the data structure of the proposed features was learned by the multireceptive field graph convolution network,so that the adaptive method can fully learn the domain invariant features,align the source domain and the target domain features more effectively,and train N groups of classifiers.Finally,the average value of the classification results of N groups of classifiers was taken as the state recognition result of the target domain samples.Based on the bearing data set of Jiangnan University,the proposed method is experimental validation,and the classification accuracy of the proposed method for the four different states(normal,inner failure,outer failure and rolling element failure)are above 99%in three different sets of variable operating conditions bearing fault diagnosis tasks,which improves the diagnostic accuracy by 0.22~8.27 percentage points compared with other methods.The comparison results show that the proposed method for the identification of faults in rolling bearings under variable operating conditions can effectively diagnose the type of bearing faults,which has a certain engineering practical value.