基于残差胶囊网络的滚动轴承故障诊断研究

Fault diagnosis of rolling bearing based on residual capsule network

采用传统的滚动轴承故障诊断方法对时域信号进行特征提取时,过分依赖于专家知识,而且提取到的特征对故障信息表达不充分,针对这一问题,提出了一种基于残差网络和胶囊网络的滚动轴承智能故障诊断方法。首先,以原始振动信号作为输入,使用一维卷积神经网络对其时域信号进行了全局特征提取;然后,利用残差网络提取了数据的低层特征,并将其输入到胶囊网络中,进行了低层特征矢量化处理;随后,采用模糊聚类改进的动态路由方法完成了低层特征到高层特征的聚合,并进行了特征分类;最后,为了验证该方法的有效性,采用滚动轴承数据集对所提出的方法进行了试验验证,并将该方法诊断结果与其他深度学习方法诊断结果进行了比较。研究结果表明:残差胶囊网络在分类精度上达到了99.95%,并且在收敛速度方面得到了提高,通过t-sne可视化分析进一步证明了该网络模型具有自适应挖掘高层特征的能力;残差胶囊网络在滚动轴承故障诊断中具有良好的精确性和泛化性。

There was over-reliance on experts’knowledge when extracting time domain signals by the traditional fault diagnosis method of rolling bearings,and the fault information was expressed inadequately by features.Aiming at the problems,an intelligent fault diagnosis model based on residual network and capsule network was proposed.Firstly,raw vibration signal was used as input,and the one-dimensional convolution neural network was used to extract global features from the time domain signal,and then the residual network was used to extract the low-level features of the data,and they were sent to the capsule network to vectorize the low-level features,after that the low-level features were combined into advanced features and classified through dynamic routing process which was improved by fuzzy clustering.Finally,in order to verify the effectiveness of this method,the proposed method was tested through the rolling bearing data sets,and the diagnosis result of this method was compared with the diagnosis result of other deep learning methods.The research results indicate that the residual capsule network reaches 99.95% in classification accuracy,and the convergence speed has been improved.The t-distributed stochastic neighbor embedding(t-sne)visible analysis further verifies that the network model has the ability to self-adaptively mine high-level features.The residual capsule network possesses good accuracy and generalization in the fault diagnosis of rolling bearings.