基于M2-MHA Block轻量化模型的小样本跨工况轴承故障诊断  

Small Sample Cross-Condition Bearing Fault Diagnosis Based on M2-MHA Block Lightweight Model

作　　者：邓兴超 朱冠华 张清华 张发振 赵乾惠 DENG Xingchao;ZHU Guanhua;ZHANG Qinghua;ZHANG Fazhen;ZHAO Qianhui(School of Information and Control Engineering,Jilin Institute of Chemical Technology,Jilin Jilin 132022,China;School of Automation,Guangdong University of Petrochemical Technology,Maoming Guangdong 525000,China)

机构地区：[1]吉林化工学院信息与控制工程学院,吉林吉林132022 [2]广东石油化工学院自动化学院,广东茂名525000

出　　处：《机床与液压》2025年第7期31-39,共9页Machine Tool & Hydraulics

基　　金：国家自然科学基金重点项目(61933013)。

摘　　要：针对轴承故障诊断中存在的训练数据不充分、模型泛化能力差以及计算复杂度大等问题,提出一种基于轻量化卷积神经网络的小样本跨工况轴承故障诊断方法。采用原始振动信号与量纲一量辅助数据集的并行输入方式,搭建基于深度可分离卷积的多输入多尺度(M2)特征提取架构,避免了仅使用原始振动信号可能导致的特征不充分问题。此外,提出一种多头注意力块(MHA Block),以提升训练效率和诊断性能。最后,通过迁移学习技术实现了基于小样本的跨工况诊断,并在凯斯西储大学数据集上进行实验验证。结果表明:所提方法在源域下的平均诊断精度达99.8%,且模型参数量和模型大小仅分别为28 789和540.5 kB;在小样本跨工况迁移诊断中,采用100个样本进行训练、500个样本进行测试,平均诊断精度高达99.3%;文中所提方法能够在低计算量条件下,实现高准确率与良好的泛化性能。Aiming at the problems such as insufficient training data,poor model generalization ability and large computational complexity in bearing fault diagnosis,a small-sample cross-condition bearing fault diagnosis method based on lightweight convolutional neural network was proposed.A multi-input multi-scale(M2)feature extraction architecture based on depth-wise separable convolution was constructed by using the parallel input of the original vibration signal and the dimensionless auxiliary data set,which avoided the problem of insufficient features caused by using only the original vibration signal.Additionally,a multi-head attention block(MHA Block)was proposed to improve training efficiency and diagnostic performance.Finally,transfer learning technique was employed toachievecross-condition diagnosis based on small samples,and the experimental verification was carried out on the Case Western Reserve University dataset.The results show that the average diagnostic accuracy of the proposed method in the source domain is 99.8%,with model parameters and size being only 28789 and 540.5 kB,respectively.Moreover,in small-sample cross-condition transfer diagnosis,100 samples were used for training and 500 for testing,and the average diagnostic accuracy reaches 99.3%.The proposed method can achieve high accuracy and good generalization performance under the condition of low computation.

关 键 词：故障诊断 轻量化 多输入多尺度 多头注意力块 小样本 跨工况 

分 类 号：TH133.33[机械工程—机械制造及自动化]