基于自注意力的时频特征融合轴承故障诊断方法  

Bearing Fault Diagnosis Method Based on Self-attentional Time-frequency Feature Fusion

作　　者：刘晶 张明辉 吴健 陈二龙 季海鹏 LIU Jing;ZHANG Minghui;WU Jian;CHEN Erlong;JI Haipeng(College of Artificial Intelligence,Hebei University of Technology,Tianjin 300400,China;Hebei Data Industrial Intelligent Engineering Research Center,Tianjin 300400,China;Tianjin Development Zone Jingnuo Data Technology Co.Ltd.,Tianjin 300400,China;Tianjin Electric Science Research Institute Co.Ltd.,Tianjin 300400,China;College of Materials Science and Engineering,Hebei University of Technology,Tianjin 300400,China)

机构地区：[1]河北工业大学人工智能与数据科学学院,天津300400 [2]河北省数据驱动工业智能工程研究中心,天津300400 [3]天津开发区精诺瀚海数据科技有限公司,天津300457 [4]天津电气科学研究院有限公司,天津300180 [5]河北工业大学材料科学与工程学院,天津300400

出　　处：《哈尔滨理工大学学报》2024年第6期1-12,共12页Journal of Harbin University of Science and Technology

基　　金：2021年度京津冀基础研究合作专项项目(21JCZXJC00050);2022年河北省自然科学基金(F2022202021);电气传动国家工程研究中心创新基金(GE2022ZL001).

摘　　要：针对故障诊断中单一时、频域信号蕴含特征不全面,导致传统故障诊断模型泛化能力不强的问题,提出一种基于自注意力的时频特征融合轴承故障诊断方法。该方法首先将原始信号和相应频谱作为双通道输入,时频交叉注意力能够通过时域发现频域的隐性特征,通过频域发现时域的隐性特征,经过多层时频交叉注意力进行全局特征提取的同时产生深度融合的时频融合特征;其次将时频融合特征输入改进的多尺度残差网络进一步提取局部特征,经全连接层降维并输出诊断结果;最后采用凯斯西储大学轴承数据集、帕德博恩大学轴承数据集和某工厂减速机轴承故障数据集进行实验验证,结果表明该方法诊断准确率及泛化能力显著提升,同工况诊断准确率可达99%以上、跨工况诊断准确率可达95%以上。Aiming at the problem that the single time and frequency domain signal contains incomplete features in bearing fault diagnosis,which leads to the poor generalization ability of traditional fault diagnosis models,a bearing fault diagnosis method based on self-attention time-frequency feature fusion was proposed.Firstly,the original signal and the corresponding spectrum were taken as dual channel inputs,and the time-frequency cross-attention was able to discover the implicit features in the frequency domain through the time domain,and the implicit features in the time domain were found through the frequency domain.After multi-layer time-frequency cross-attention for global feature extraction,deep fusion time-frequency fusion features were generated.Secondly,the time-frequency fusion features were input into the improved multi-scale residual network to further extract local features,and the dimension was reduced through the fully connected layer and the diagnosis results were output.Finally,the bearing fault data set of Case Western Reserve University and the reducer bearing fault data set of a factory were used for experimental verification.The results show that the diagnostic accuracy and generalization ability of the proposed method are significantly improved,and the diagnostic accuracy of the same working condition can reach more than 99%,and the cross-condition diagnosis accuracy can reach more than 95%.

关 键 词：轴承故障诊断 自注意力 特征融合 深度学习 多尺度残差网络 

分 类 号：TP306.1[自动化与计算机技术—计算机系统结构]