采用大核注意力机制的抗噪轴承故障诊断模型

Fault Diagnosis Model of Noise-resistant Bearings Using Large Kernel Attention Mechanism

针对滚动轴承故障诊断模型易受轴承工作环境噪声以及运行数据样本数量影响的特点,提出一种并行大核注意力机制卷积神经网络(PLKACNN)。首先,将一维时间序列通过短时傅里叶变换转化成二维图像作为模型的输入,使用并行大核注意力机制实现对不同维度的故障特征的提取;其次,将并行支路所得的特征图进行堆叠以获得整体信息,最终通过整合卷积层以及全连接层对整体信息进行学习以及分类。结果表明PLKACNN能够有效识别滚动轴承故障程度和故障位置,在带噪数据集上获得平均98.5%的准确率,并且在带噪小样本实验中获得92.81%平均准确率,证明所提PLKACNN具有较好的噪声鲁棒性以及泛化能力。

Aiming at the problem that the rolling bearing fault diagnosis models are susceptible to the noise of the bearing working environment and the number of vibration signal data samples,a parallel large kernel attention convolutional neural network(PLKACNN)is proposed.Firstly,the one-dimensional time series is transformed into a two-dimensional image by short-time Fourier transform,which is used as the input of the model.And the parallel large kernel attention module is used to extract the fault features of different dimensions.Then,the feature maps obtained from the parallel branches are stacked to obtain the overall information.Finally,the overall information is learned and classified by collected convolutional layer and fully connected layer.The results show that PLKACNN can effectively identify the degree and location of rolling bearing faults with average accuracies of 98.57%on the noisy dataset and 92.81%in the noisy small sample experiments.It proves that the proposed PLKACNN has good noise robustness and generalization ability.