基于无监督特征对齐的滚动轴承故障诊断

FAULT DIAGNOSIS OF ROLLING BEARING BASED ON UNSUPERVISED FEATURE ALIGNMENT

针对不同转速环境下采集到的滚动轴承振动数据特征分布不一导致待诊断样本标签较难获取的问题,提出一种基于深度迁移网络的故障诊断方法。首先,搭建领域共享的特征提取网络,利用卷积神经网络(Convolutional Neural Network, CNN)提取振动信号中敏感故障特征,并结合双向长短时记忆(Bi-directional Long Short-Term Memory, BiLSTM)网络进一步提取敏感故障特征中的时间信息;然后,在深度迁移网络中分别嵌入CORAL损失和JMMD损失,通过最小化二阶统计量差异和联合分布最大均值差异值,缩小源域和目标域特征分布差异,进而提取到两域的共同特征;最后,添加Softmax分类层,实现对目标数据的故障状态识别。结果表明,该方法在目标域数据无标签的情况下,平均识别准确率为97.87%,明显高于目前流行的其它5种领域自适应故障诊断方法。

Aiming at the problem that the characteristic distribution of rolling bearing vibration data collected under different speed environment is inconsistent and it is difficult to obtain the label of samples to be diagnosed, a fault diagnosis method based on deep migration network is proposed. In this model, a domain-shared feature extraction network is constructed. The convolutional neural network(CNN) is used to extract vibration signal sensitive fault features, and bi-directional Long short-term Memory is used to extract vibration signal sensitive fault features. BiLSTM) network to extract the time information of sensitive fault features;Then, CORAL loss and JMMD loss were embedded in the deep migration network, respectively. By minimizing the second-order statistical difference and the maximum mean difference of the joint distribution, the differences in the feature distributions of the source domain and target domain were reduced, and the common features of the two domains were extracted. Finally, add Softmax classification layer to realize fault status recognition of target data. The results show that the average recognition accuracy of this method is 97.87% when the target domain data is unlabeled, which is significantly higher than the other five popular adaptive fault diagnosis methods.