摘要
针对卷积神经网络在处理滚动轴承时域信号时难以充分提取特征、故障样本稀少及模型泛化性能不足的问题,提出一种基于注意力机制的增强卷积神经网络小样本故障诊断方法。首先,使用连续小波变换将轴承振动信号转化为二维时频图像,以便可视化其特征。然后,通过数据增强扩充样本数据,提升模型在小样本情况下的泛化性。为提高特征提取和模型泛化能力,使用MixConv将ConvNeXt V2模型的7×7卷积层重构为不同大小的并行卷积核,增强多尺度特征提取效果;引入卷积注意力机制模块(CBAM)提升关键特征识别能力。该模型在凯斯西储大学、东南大学和渥太华大学的故障数据集上进行实验验证。实验结果表明,所提模型对不同故障的识别率均为100%,与目前常用的7个模型相比,在相同条件下故障识别准确率最高,具有较强的泛化性能。
Aiming at the problems that convolutional neural network is difficult to fully extract features,the fault samples are scarce and the model generalization performance is insufficient when dealing with rolling bearing time domain signals,an enhanced convolutional neural network small-sample fault diagnosis method based on the attention mechanism is proposed.First,continuous wavelet transform is applied to convert the bearing vibration signal into a two-dimensional time-frequency image to visualize the features.Then,the sample data is expanded by data enhancement to improve the generalization of the model in the case of small samples.To improve the feature extraction and model generalization ability,the 7×7 convolutional layers of the ConvNeXt V2 model are reconstructed into parallel convolutional kernels of different sizes using MixConv to enhance the multi-scale feature extraction effect;and the CBAM attention mechanism is introduced to improve the key feature recognition ability.The model is experimentally validated on fault datasets from CWRU,SEU and University of Ottawa.The experimental results show that the proposed model has a recognition rate of 100%for different faults,which is the highest recognition rate for faults under the same conditions compared with the 7 models commonly used at present,and has strong generalization performance.
作者
宋存利
王子卓
时维国
SONG Cunli;WANG Zizhuo;SHI Weiguo(Dalian Jiaotong University,Dalian 116028,China)
出处
《中国惯性技术学报》
北大核心
2025年第1期96-106,共11页
Journal of Chinese Inertial Technology
基金
国家自然基金(62276042)
辽宁省教育厅项目资助课题(LJKMZ20220828)。
关键词
滚动轴承
故障诊断
注意力机制
连续小波变换
卷积神经网络
rolling bearings
fault diagnosis
attention mechanism
continuous wavelet transform
convolutional neural network
作者简介
宋存利(1975-),女,教授,硕士生导师。从事智能制造,故障诊断,小目标物体检测研究;通讯作者:时维国(1973-),男,教授,从事多智能体系统、故障诊断、网络控制研究。
