基于频域特征变分自编码器的轴承故障诊断研究

Research on Bearing Fault Diagnosis Based on Frequency Domain Characteristic Variational Autoencoder

以往基于深度学习的机械故障诊断模型的泛化能力差,且网络复杂;训练一个通用的特征提取方法,对提取的特征进行分类,这一方法能够很好地解决此问题;针对这一问题,从两个方面展开研究:1)改进特征提取算法:将频域特征提取自编码器与变分自编码器结合,提出频域特征变分自编码器,使得提取到的特征更为鲁棒;2)离群点剔除:在特征提取之后,加入局部异常因子算法对离群点进行剔除,防止分类器过拟合,使得分类器泛化性能更好;故障诊断整体流程是,首先将样本数据输入变分频域特征提取自编码器进行特征提取,其次使用局部异常因子对离群点进行剔除,最后将特征输入分类器进行故障诊断。通过实验验证在不同损伤程度下特征提取的界限清晰,分类效果较好,同时表现出可迁移性,为后续故障诊断和迁移学习方法有一定的应用价值。

The generalization ability of previous deep learning-based mechanical fault diagnosis models is poor,and the network is complex.Training a general feature extraction method to classify extracted features is a good solution to this problem.Aiming at this problem,the research is carried out from two aspects:(1)Improved feature extraction algorithm:frequency domain feature extraction autoencoder is combined with variational autoencoder,and a frequency domain feature variational autoencoder is proposed to make the extracted features more accurate.(2)Outlier elimination:after feature extraction,a local anomaly factor algorithm is added to eliminate outliers to prevent the overfitting of classifier and make the generalization performance of classifier better.The overall process of fault diagnosis is to first input the sample data into the variational frequency domain feature extraction from the encoder for feature extraction,then use the local abnormal factor to eliminate outliers,and finally input the features to the classifier for fault diagnosis.It is verified by experiments that the boundary of feature extraction is clear under different damage degrees,the classification effect is good,and it shows transferability,which has certain application value for subsequent fault diagnosis and transfer learning method.