对失效传感器具备鲁棒性的故障诊断方法  

Robust fault diagnosis method for failure sensors

作　　者：马现伟 范朝辉 聂为之 李东[4] 朱逸群 MA Xianwei;FAN Chaohui;NIE Weizhi;LI Dong;ZHU Yiqun(School of Future Technology,Tianjin University,Tianjin 300072,China;Strategic Assessments and Consultation Institute,Academy of Military Science,Beijing 100091,China;School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China;College of Intelligence and Computing,Tianjin University,Tianjin 300050,China;Metrology Center,State Grid Tianjin Marketing Service Center,Tianjin 300160,China)

机构地区：[1]天津大学未来技术学院,天津300072 [2]军事科学院战略评估咨询中心,北京100091 [3]天津大学电气自动化与信息工程学院,天津300072 [4]天津大学智能与计算学部,天津300050 [5]国网天津市电力公司营销服务中心计量中心,天津300160

出　　处：《浙江大学学报（工学版）》2024年第7期1488-1497,共10页Journal of Zhejiang University：Engineering Science

基　　金：国家重点研发计划资助项目(2020YFB1711700)。

摘　　要：针对传感器失效对故障诊断效果的影响,提出带有退化特征惩罚机制的故障诊断方法 (FDCFP).利用卷积神经网络提取局部特征,使用Transformer编码器进一步融合全局信息,缓解局部传感器失效对特征学习的影响.引入自注意力掩码对传感器失效引入的特征伪关联性进行惩罚,使编码器更关注未受失效影响的传感器数据.掩码在自注意力层参与权重计算,以较低的复杂度代价提高含有失效数据模型的鲁棒性.在CWRU、PU、SEU和XJTU-SY数据集上进行FDCFP的性能验证,当数据失效比例为0.7时,FDCFP在CWRU、PU和XJTU-SY数据集上的故障诊断准确率相较于现有方法分别提升了5.9%、3.1%和3.7%.A fault diagnosis method with corrupted feature penalties(FDCFP)was proposed to address the impact of sensor failure on fault diagnosis.The convolutional neural network was used to extract local features,and the Transformer encoder was used to further fuse global information,alleviating the influence of local sensor failures on feature learning.The self-attention masks were employed to penalize the pseudo-correlation between features introduced by the sensor failure,making the encoder attentive to the parts of the sensor data unaffected by failures.The masks were weighted at the self-attention layer to improve the model’s robustness to sensor failures in the presence of faulty data at a low complexity cost.The performance of the proposed method was verified in datasets such as CWRU,PU,SEU,and XJTU-SY.Compared to existing methods,the FDCFP achieved accuracy improvements of 5.9%,3.1%,and 3.7%respectively on the CWRU,PU,and XJTU-SY datasets at a 0.7 data failure ratio.

关 键 词：传感器失效 故障诊断 深度学习 注意力机制 TRANSFORMER 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TH165.3[自动化与计算机技术—控制科学与工程]