基于深度特征选取的旋转机械跨域故障诊断

Cross-domain fault diagnosis of rotating machinery based on deep features selection

在实际的工业场景中,对旋转机械进行故障诊断时,存在着标签故障样本不足和数据分布差异的问题,为此,基于深度特征选取和迁移学习方法,提出了一种新的跨域故障诊断方法。首先,利用深度自编码器进行了深度特征提取,利用不同激活函数下的深度自编码器提取出的深度特征,构建了深度特征池;然后,采用提出的面向跨域诊断的特征选取方法,选取了可迁移特征用于后续的特征迁移学习,利用所提出的改进联合分布适应方法,降低了源域和目标域特征数据间分布差异;最后,基于经迁移学习后的有标签源域样本和无标签目标域样本,对故障识别分类器进行了训练,并利用机械故障模拟实验台的轴承和电机故障数据,开展了旋转机械跨域故障诊断的实验。研究结果表明:与对比模型相比,所提出的方法能够取得更优秀的跨域故障诊断性能;在选取合适的特征数时,其最大故障诊断准确率明显高于其他对比模型(其中,轴承为95.42%,电机为88.67%)。

In practical industrial scenarios,when diagnosing rotating machinery faults,there were problems such as insufficient labeled fault samples and differences in data distribution.For this reason,a new cross-domain fault diagnosis was proposed based on deep feature selection and transfer learning methods.Firstly,the deep feature extraction was carried out by using the deep autoencoder,and the deep feature pool was constructed by using the deep features extracted by the deep autoencoder under different activation functions.Then,the proposed features selection method for cross domain diagnosis was used to select transferable features for the subsequent feature transfer learning.The proposed improved joint distribution adaptation was used to reduce the distribution differences between source and target domains.Finally,based on the labeled source domain samples and unlabeled target domain samples after transfer learning,the fault recognition classifier was trained,and the cross-domain fault diagnosis experiment was carried out through the bearing and motor fault data of the mechanical fault simulation test-bed.The research results show that the proposed diagnosis method can achieve the better cross-domain fault diagnosis performance than the comparison models,and its maximum fault diagnosis accuracy(bearing:95.42%,motor:88.67%)is significantly higher than other comparative models when the suitable number of features were selected.