基于ECA-ResNet的轴承故障诊断方法  

Fault Diagnosis Method for Bearings Based on ECA-ResNet

作　　者：杨向兰 孙士保 王国强[2] 石念峰[2] 谢扬筱 YANG Xianglan;SUN Shibao;WANG Guoqiang;SHI Nianfeng;XIE Yangxiao(College of Information Engineering,Henan University of Science and Technology,Luoyang 471023,China;School of Computer and Information Engineering,Luoyang Institute of Science and Technology,Luoyang 471023,China;College of Computer Science and Technology,Shanghai University of Electric Power,Shanghai 200090,China)

机构地区：[1]河南科技大学信息工程学院,河南洛阳471023 [2]洛阳理工学院计算机与信息工程学院,河南洛阳471023 [3]上海电力大学计算机科学与技术学院,上海200090

出　　处：《轴承》2025年第1期102-110,共9页Bearing

摘　　要：针对传统的轴承故障诊断方法存在强噪声与变负载下关键特征提取不充分的问题,提出了一种基于有效通道注意力机制和残差网络(ECA-ResNet)的轴承故障诊断方法。首先,利用短时傅里叶变换与伪彩色处理方法将轴承振动信号转换为三通道彩色图像,以此作为输入样本集;其次,基于卷积和池化操作对数据进行浅层边缘特征提取与压缩,并在有效通道注意力网络(ECA-Net)中引入残差网络模块,解决网络退化的问题,建立通道之间的联系,自适应提取深层关键特征;最后,引入Dropout抑制模型过拟合,并以Softmax层作为分类器实现故障诊断。用凯斯西储大学与江南大学轴承数据集进行验证,结果表明该方法在强噪声和变负载下的平均准确率分别为97.5%,93.69%,抗噪性、泛化性和通用性较好。Aimed at the problem that traditional fault diagnosis methods for bearings cannot extract the key features sufficiently under strong noise and variable load,a fault diagnosis method for bearings is proposed based on effective channel attention-residual network(ECA-ResNet).Firstly,the bearing vibration signals are converted into threechannel color images using short-time Fourier transform and pseudo-color processing method,which is used as input sample set.Secondly,the shallow edge features of data are extracted and compressed based on convolution and pooling operation,and the residual network block is introduced into efficient channel attention network(ECA-Net)to solve the problem of network degradation,establish the connection between channels,and extract the deep key features adaptively.Finally,Dropout is introduced to suppress the model overfitting,and the Softmax layer is used as classifier for fault diagnosis.The verification is conducted using bearing data sets from Case Western Reserve University and Jiangnan University.The results show that the average accuracy of the proposed method can reach 97.5%and 93.69%under strong noise and variable load respectively,having better noise resistance,generalization and versatility.

关 键 词：滚动轴承 故障诊断 傅里叶变换 深度学习 残差 

分 类 号：TH133.3[机械工程—机械制造及自动化] TP183[自动化与计算机技术—控制理论与控制工程]