基于MACNN的柴油机故障诊断方法研究  

Fault diagnosis method for diesel engines based on MACNN

作　　者：程建刚 毕凤荣[1] 张立鹏[2] 李鑫 杨晓[1] 汤代杰 CHENG Jiangang;BI Fengrong;ZHANG Lipeng;LI Xin;YANG Xiao;TANG Daijie(State Key Laboratory of Engines,Tianjin University,Tianjin 300072,China;Tianjin Internal Combustion Engine Research Institute,Tianjin 300072,China)

机构地区：[1]天津大学内燃机燃烧学国家重点实验室,天津300072 [2]天津内燃机研究所,天津300072

出　　处：《振动与冲击》2022年第10期8-15,共8页Journal of Vibration and Shock

基　　金：天津市自然科学基金(18JCYBJC20000)。

摘　　要：高效、准确的故障诊断可以提高柴油机的安全性和可靠性。传统机械故障诊断方法中人工参与程度过高,对识别结果带来诸多不确定性。针对这一问题,提出一种基于多重注意力卷积神经网络(multiple attention convolutional neural networks,MACNN)的端到端故障诊断方法。该方法采用多层卷积神经网络(convolutional neural networks,CNN)结合卷积注意力模块(convolutional block attention module,CBAM)对原始时域数据进行特征提取;然后,对多维卷积输出特征图进行重组以保留其序列信息;最后,直接采用序列注意力机制完成序列特征的学习。经采用实测柴油机缸盖振动信号数据进行验证后表明:面对8分类柴油机故障数据集,MACNN能够达到97.88%的识别准确率,测试100个样本用时仅为0.35 s。与现有多种传统故障诊断方法和端到端故障诊断方法相比,均具有更好的诊断效果。In the traditional mechanical fault diagnosis method for diesel engines,the degree of human participation is too high,which brings high uncertainty to results.To solve the problem,an end-to-end fault diagnosis method was proposed based on multiple attention convolutional neural networks(MACNN).In the method,multi-layer convolutional neural networks(CNN)and convolutional block attention module(CBAM)were combinedly used to extract features from the original time-domain data,and then the multi-dimensional feature map of convolutional output was recombined to retain its sequence information.Finally,the sequential attention mechanism was directly used to learn the sequence feature.The results show that MACNN can achieve a recognition accuracy of 97.88%for eight-class fault data sets of diesel engines,and the time taken to test 100 samples is only 0.35 s.Compared with other traditional fault diagnosis methods and end-to-end fault diagnosis methods,the proposed MACNN has better diagnosis effect.

关 键 词：多重注意力 卷积神经网络(CNN) 故障诊断 端到端 

分 类 号：TH17[机械工程—机械制造及自动化] TN911[电子电信—通信与信息系统]