基于迁移堆栈自编码器的轴承故障诊断方法

Bearing Fault Diagnosis Method Based on Transfer Learning and Stacked Auto-encoders

近年来,基于数据驱动的设备智能故障诊断方法是监测设备健康状况的重要手段,然而在实际中不同工况下含标注的监测数据严重缺乏,导致智能故障诊断的模型难以有效构建。提出一种基于迁移堆栈自编码器的轴承故障诊断方法,能成功解决不同工况下轴承故障智能诊断的问题。首先,将不同工况下轴承原始振动信号数据进行快速傅里叶变换转化成频域信号,得到带标签的源域和不带标签的目标域数据集。其次,使用基于堆栈自编码器的多分类网络结构对源域数据进行特征提取,为防止过拟合加入Dropout层和批标准化层,从而有效提高特征的提取。最后,利用多核最大均值差异作为评价源域和目标域的距离指标,实现域不变特征提取并进行迁移学习。将该方法用于不同工况下滚动轴承的数据集进行验证,结果表明目标域样本充足时轴承故障诊断分类准确率能够达到99.4%,目标域样本为源域样本5%时其分类准确率能达到95.2%,具有较好的应用前景。

In recent years,the intelligent fault diagnosis method of equipment based on data-driven is an important means to monitor the health status of equipment.However,in engineering practice,the serious lack of monitoring data with labels under different working conditions makes it difficult to construct the intelligent fault diagnosis model effectively.In this paper,a bearing fault diagnosis method based on transfer learning and stacked auto-encoders is proposed,which can successfully solve the problem of bearing fault intelligent diagnosis under different working conditions.Firstly,the original vibration signal data of bearings under different working conditions are converted into frequency domain signals by fast Fourier transform,and the data sets of source domain and target domain with and without labels are obtained.Then,the multi-classification network structure based on the stacked auto-encoders is used to extract the features of source domain data.To prevent overfitting,dropout layers and batch standardization layers are added to effectively improve the feature extraction.Finally,multi-kernel maximum mean discrepancies is used as the index to evaluate the distance between the source domain and the target domain so as to achieve domain invariant feature extraction and transfer learning.The proposed method is used to verify the data sets of rolling bearings under different working conditions.The results show that the classification accuracy of bearing fault diagnosis can reach 99.4%when the target domain samples are sufficient,and the classification accuracy rate can reach 95.2%when the target domain samples are only 5%of the source domain samples.This method is proved to have a good application prospect.