基于注意力机制的多尺度卷积神经网络电机轴承故障诊断

Multi Scale Convolutional Neural Network Fault Diagnosis of Motor Bearing Based on Attention Mechanism

针对现有的卷积神经网络(Convolutional Neural Network,CNN)故障诊断方法只能提取单尺度特征,丢失了故障敏感信息,无法正确表达电机轴承的健康状态的问题,提出了注意力机制的多尺度卷积神经网络(Multi-scale Convolutional Neural Network,MSCNN)故障诊断方法,将多尺度特征提取整合到传统的CNN结构中。通过不同尺寸的卷积核捕获信号的多尺度特征,使模型获得多样性的特征表达;引入注意力机制(Attention Mechanism,ATT),对提取的特征自适应的评分和赋值,将注意力集中在敏感特征上,让模型学习到高级特征;最后利用公开数据集进行实验验证,结果表明,所提方法诊断精度高,具有较好的泛化性能。

In response to the problem that existing Convolutional Neural Network(CNN)fault diagnosis methods can only extract single scale features,lose fault sensitive information,and cannot accurately express the health status of motor bearings,a Multi-scale Convolutional Neural Network(MSCNN)fault diagnosis method with attention mechanism is proposed,which integrates multi-scale feature extraction into traditional CNN structures.By capturing multi-scale features of signals through convolutional kernels of different sizes,the model achieves diverse feature representations.Introducing Attention Mechanism(ATT)to adaptively score and assign extracted features,focusing attention on sensitive features and allowing the model to learn advanced features.Finally,experimental verification was conducted using publicly available datasets,and the results showed that the proposed method has high diagnostic accuracy and good generalization performance.