用于轴承故障诊断任务的轻量化卷积网络  

Lightweight Convolutional Network for Bearing Fault Diagnosis

作　　者：刘辉 李阳 侯一民[3] LIU Hui;LI Yang;HOU Yimin(China Nuclear Industry Maintenance Co.,Ltd,Shanghai 201103,China;School of Mechanical Engineering,Northeast Electric Power University,Jilin 132012,China;School of Automation Engineering,Northeast Electric Power University,Jilin 132012,China)

机构地区：[1]中核检修有限公司,上海201103 [2]东北电力大学机械工程学院,吉林吉林132012 [3]东北电力大学自动化工程学院,吉林吉林132012

出　　处：《哈尔滨理工大学学报》2024年第4期80-88,共9页Journal of Harbin University of Science and Technology

基　　金：国家自然科学基金青年基金(52205518).

摘　　要：针对机械故障诊断领域中,许多具有优异性能的卷积模型受制于部署成本的因素而难以应用于工业实践的问题,研发参数更低的轻量化诊断方法。采用深度可分离卷积压缩网络参数量的方法,研究了深度可分离卷积构建轻量化轴承故障诊断模型的可行性,从而给出了保障模型诊断精度的条件下压缩卷积网络参数量的策略。在开源和自制的轴承振动数据集上,对所提出的方法有效性进行了验证。结果表明,使用深度可分离卷积对卷积模型进行参数压缩,能够在满足轻量化需求的情况下,确保模型具有高诊断精度(96.20±2.81%)。In the field of bearing fault diagnosis,many convolutional models with excellent performance face challenges in industrial applications due to deployment cost constraints.This paper aims to develop a lightweight diagnostic method with reduced parameters.We investigate the feasibility of using depthwise separable convolution to construct a lightweight bearing fault diagnosis model,thereby proposing a strategy to compress the parameters of the convolutional network while ensuring diagnostic accuracy.The effectiveness of the proposed method is validated on both publicly available and custom vibration signal datasets.The results demonstrate that compressing convolutional models using depthwise separable convolution allows for lightweight requirements while maintaining a high diagnostic accuracy(96.20±2.81%).

关 键 词：故障诊断 卷积神经网络 模型结构设计 振动(机械) 

分 类 号：TH212[机械工程—机械制造及自动化] TH213.3