基于多尺度知识蒸馏与增量学习的滚动轴承故障诊断方法  

A rolling bearing fault diagnosis method based on multi-scale knowledge distillation and continual learning

作　　者：夏逸飞 皋军[1] 邵星[1] 王翠香[1] XIA Yifei;GAO Jun;SHAO Xing;WANG Cuixiang(School of Information Engineering,Yancheng Institute of Technology,Yancheng 224000,China;School of Mechanical Engineering,Yancheng Institute of Technology,Yancheng 224000,China)

机构地区：[1]盐城工学院信息工程学院,江苏盐城224000 [2]盐城工学院机械工程学院,江苏盐城224000

出　　处：《振动与冲击》2024年第12期276-285,共10页Journal of Vibration and Shock

基　　金：国家自然科学基金(62076215);教育部新一代信息技术创新项目(2020ITA02057);盐城工学院研究生科研与实践创新计划项目(SJCX23_XZ032)。

摘　　要：为了缓解单任务轴承故障诊断方法在不同工况诊断时产生的灾难性遗忘问题,提出一种基于多尺度知识蒸馏与增量学习(multi-scale knowledge distillation and continual learning,CL-MSKD)的滚动轴承故障诊断方法。以一维卷积神经网络作为CL-MSKD主要框架,余弦归一化层作为多任务共享的分类器,通过标签与特征两个尺度的知识蒸馏实现模型知识的保存与传递。CL-MSKD能够以一个统一结构的网络模型对在不同工况下的轴承故障进行诊断,通过知识压缩方法不断地学习和保存知识,最终缓解增量阶段产生的灾难性遗忘问题,提升跨工况场景下轴承故障诊断性能。试验表明,CL-MSKD能够有效缓解灾难性遗忘并保持良好的诊断效果。在任务环境差异较大的情况下,准确率指标仍能达到97.09%,与其他增量方法相比稳定性更好,精度更高。To alleviate the catastrophic forgetting problem caused by the single-task bearing fault diagnosis method under different working conditions,a rolling bearing fault diagnosis method based on multi-scale knowledge distillation and continual learning(CL-MSKD) was proposed.The one-dimensional convolutional neural network was used as the main framework of CL-MSKD,and the cosine normalization layer was used as a multi-task shared classifier.The model knowledge was preserved and transmitted through the knowledge distillation of label and feature scales.CL-MSKD can diagnose bearing faults under different working conditions with a unified structure network model,continuously learn and save knowledge through a knowledge compression method,and finally alleviate the catastrophic forgetting problem in the incremental stage,and improve the bearing fault diagnosis performance under cross-working conditions.The experiment shows that CL-MSKD can effectively alleviate catastrophic amnesia and maintain good diagnostic effect.In the case of large differences in task environments,the accuracy index can still reach 97.09%,which has better stability and higher precision than other incremental methods.

关 键 词：增量学习 知识蒸馏 卷积神经网络 轴承故障诊断 共享分类器 

分 类 号：TH212[机械工程—机械制造及自动化] TH213.3