基于时频域改进型胶囊网络的轴承故障诊断  

Bearing Fault Diagnosis Based on Time-frequency Domain Improved Capsule Network

作　　者：孙岩 彭高亮[1] SUN Yan;PENG Gao-liang(School of Mechanical and Electrical Engineering, Harbin Institute of Technology, Harbin 150000, China)

机构地区：[1]哈尔滨工业大学机电工程学院,哈尔滨150000

出　　处：《组合机床与自动化加工技术》2021年第2期6-8,13,共4页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：国家自然科学基金(面上项目)(51875138)。

摘　　要：为提高噪声工况下轴承故障诊断的准确性,结合振动信号时频域特征规律,提出了一种时频域改进型胶囊网络的诊断方法,采用振动信号低频段短时傅立叶变换实现时频域特征构建,搭建具有直线性感受野的胶囊网络LR-Capsulenet模型,模型前端采用直线性卷积核实现时频域特征识别,再通过三次动态路由方法构建数字胶囊层,动态路由过程采用向量乘法体现线与线空间位置关系,实现故障类型区分。为验证方法的可行性和有效性,选用西储大学轴承数据集验证,与常见卷积神经网络进行对比,结果表明文中提出的方法,具有更高的准确性和稳定性,诊断准确率97%左右。In order to improve the accuracy of the noise condition of bearing fault diagnosis,combined with the feature of vibration signal time and frequency domain,this paper proposes a time-frequency domain modified capsule network diagnostic method,with low frequency domain vibration signal feature built the implementation,short time Fourier transform(STFT)structures with linear feel wild capsule network(LR-Capsulenet)model,the front end of model use the linear convolution to verify current frequency domain identification,then through three times of dynamic routing method to build digital capsule layer,dynamic routing process use vector multiplication manifests the line and the line space position relations,implement different types of faults.In order to verify the feasibility and effectiveness of the method,the bearing data set of Western Reserve University was selected for verification,and the results were compared with the common convolutional neural network.The results show that the method proposed in this paper has higher accuracy and stability,and the diagnostic accuracy is about 97%.

关 键 词：故障诊断 时频域 直线性感受野 胶囊网络 动态路由 准确率 

分 类 号：TH133[机械工程—机械制造及自动化] TG506[金属学及工艺—金属切削加工及机床]