基于时频融合多级注意力机制的双通道CNN轴承故障诊断模型  

Bearing Fault Diagnosis Model Based on Dual-channel CNN with Time-Frequency Fusion Multi-Level Attention Mechanism

作　　者：冯新[1] 陈儒晖 杨雄[1] FENG Xin;CHEN Ruhui;YANG Xiong(Zhicheng College,Fuzhou University,Fuzhou 350002,China)

机构地区：[1]福州大学至诚学院,福建福州350002

出　　处：《贵州大学学报（自然科学版）》2024年第6期70-77,共8页Journal of Guizhou University:Natural Sciences

基　　金：福建省创新战略研究计划资助项目(2023R0034)。

摘　　要：为进一步提高轴承故障诊断准确率,提出了一种基于快速傅里叶变换(fast fourier transform,FFT)和变分模态分解(variational mode decomposition,VMD),并融合多级注意力机制的双通道卷积神经网络(convolutional neural networks,CNN)模型用于滚动轴承故障诊断。首先,将一维故障信号经过FFT和VMD处理后进行堆叠,作为双通道CNN的输入;其次,将预处理后的数据分别通过基于通道注意力和全局注意力的二维CNN提取重要特征;再次,利用交叉注意力机制将两个通道提取的特征进行融合;最后,经过全连接层和softmax分类器进行故障诊断。试验结果表明:采用该方法在美国凯斯西储大学10类轴承故障数据集的平均准确率达到100%,其诊断精度优于常见的故障预测模型和单通道模型,有利于促进轴承的智能故障诊断研究和实际应用。To further improve the accuracy of bearing fault diagnosis,a dual-channel convolutional neural network(CNN)model incorporatingfast fourier transform(FFT),variational mode decomposition(VMD),and multi-level attention mechanisms is proposed for diagnosing rolling bearing faults.First,the one-dimensional fault signal is processed using FFT and VMD,and the resulting data is stacked as the input for the dual-channel CNN.Second,the preprocessed data is passed through a two-dimensional CNN based on channel attention and global attention to extract important features.Third,a cross-attention mechanism is employed to fuse the features extracted from the two channels.Finally,fault diagnosis is performed through a fully connected layer and a softmax classifier.Experimental results show that this method achieves an average accuracy of 100% on the 10-class bearing fault dataset from Case Western Reserve University.Its diagnostic accuracy is better than that of common fault prediction models and single-channel models,which is conducive to promoting the research and practical application of intelligent fault diagnosis of bearings.

关 键 词：故障诊断 时频融合 注意力机制 双通道卷积神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]