摘要
针对滚动轴承信号易受噪声干扰和智能诊断模型在不同工况下自适应性差的问题,提出了一种多尺度注意力卷积神经网络(MSACNN)模型.首先,将一维时间序列转化为二维图像作为模型的输入,在特征提取过程中,利用多尺度卷积结构拓宽网络的宽度并实现不同维度敏感特征的提取;然后,通过注意力机制对数据不同维度的特征赋予不同的权重,使模型更关注于最具类别区分度的区域,从而提高模型的特征学习能力;最后,通过全连接层的多分类函数实现滚动轴承的故障诊断.实验结果表明:与其他方法相比,该模型不仅能在同负载各测试集上达到很高的准确率,而且在变负载工况下具有较强的迁移泛化能力和鲁棒性;该模型在强噪声环境下也具有良好的诊断性能,较其他方法抗噪性优势明显.此外,通过可视化方法分析了该模型的特征学习过程和故障分类机理.
Aimed at the problem that rolling bearing signals are easily disturbed by noise and the intelligent diagnosis model has poor adaptability under different working conditions,a multi-scale attention convolutional neural network(MSACNN)model is proposed.First,the one-dimensional time series is converted into a two-dimensional image as the input of the model.In the feature extraction process,the multi-scale convolution structure is used to broaden the width of the network and realize the extraction of sensitive features of different dimensions.Then,the attention mechanism is used to assign different weights to the features of different dimensions of the data,so that the model pays more attention to the areas with the most classification,thereby improving the feature learning ability of the model.Finally,the failure diagnosis of the rolling bearing was realized by the multi-classification function of the fully connected layer.Experimental results show that compared with other methods,the proposed model can not only achieve high accuracy on each test set of the same load,but also has strong migration generalization ability and robustness in a variable load environment.Meanwhile,the model also has good diagnostic performance in a strong noise environment,and obvious advantages in noise immunity compared with other methods.In addition,the feature learning process and fault classification mechanism of the model are analyzed by visualization methods.
作者
丁雪
邓艾东
李晶
邓敏强
徐硕
史曜炜
Ding Xue;Deng Aidong;Li Jing;Deng Minqiang;Xu Shuo;Shi Yaowei(School of Energy and Environment,Southeast University,Nanjing 210096,China;National Engineering Research Center of Turbo-Generator Vibration,Southeast University,Nanjing 210096,China;School of Information Engineering,Nanjing Audit University,Nanjing 211815,China)
出处
《东南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2022年第1期172-178,共7页
Journal of Southeast University：Natural Science Edition
基金
国家自然科学基金资助项目(51875100,52005267)
中央高校基本科研业务费专项资金资助项目(3203002101C3).
关键词
滚动轴承
故障诊断
多尺度特征提取
注意力机制
卷积神经网络
rolling bearing
fault diagnosis
multi-scale feature extraction
attention mechanism
convolutional neural network
