Abstract: Aiming at the current situation that supervised deep learning is mainly used to extract fault features and detect coarse-grained types of faults in rolling bearing fault diagnosis, a fine-grained fault diagnosis method for rolling bearings integrated with Gaussian mixture models (GMM) and deep residual shrinkage networks (DRSN) is proposed. The GMM model integrates multiple Gaussian distribution functions to fit the distribution of fine-grained fault data and realize the clustering of bearing vibration signals without labels. The attention mechanism in DRSN model focused on the more critical information for the current task from a large number of fault feature information. Soft threshold is designed to set different thresholds for bearing samples in different health states. The method is validated by collecting 30 bearing health states from Case Western Reserve University (CWRU) data. The results show that integrate the unsupervised model with the deep learning model, which can process the bearing fault data without labels, achieve the purpose of fine-grained classification of bearing faults, provide a basis for subsequent equipment maintenance, and have good practical engineering significance and popularization.

Key words: fine-grained fault diagnosis, rolling bearing, Gaussian mixture model, deep residual shrinkage network, unsupervised learning, deep learning

摘要： 针对目前滚动轴承故障诊断主要采用监督式深度学习提取故障特征以及检测故障种类为粗粒度的现状，提出一种基于高斯混合模型（Gaussian mixed models，GMM）和深度残差收缩网络（deep residual shrinkage networks，DRSN）的滚动轴承细粒度故障诊断方法。GMM模型集成多个高斯分布函数，拟合细粒度故障数据的分布情况，实现对没有标签的轴承振动信号进行聚类，DRSN模型中注意力机制从大量故障特征信息中聚焦于对当前任务更为关键的信息，软阈值化旨在为处于不同健康状态的轴承样本设置不同的阈值。在凯斯西储大学（Case Western Reserve University，CWRU）滚动轴承故障数据中收集30种轴承健康状态对该方法进行了验证，结果表明，将非监督模型与深度学习模型融合，能够处理不含标签情况下的轴承故障数据，实现对轴承故障进行细粒度分类的目的，为后续的设备维护提供依据，具有较好的实际工程意义和推广性。

关键词: 细粒度故障诊断, 滚动轴承, 高斯混合模型, 深度残差收缩网络, 非监督学习, 深度学习