首页|基于监督对比学习和混合注意力残差网络的隔膜泵单向阀故障诊断

基于监督对比学习和混合注意力残差网络的隔膜泵单向阀故障诊断

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
由于工业生产环境中的强噪声和其他环境激励,隔膜泵单向阀不同故障的特征呈现一定的相似性,导致传统深度学习方法对单向阀的故障状态难以准确识别.为解决这一问题,提出了一种结合监督对比学习和混合注意力残差神经网络(HA-ResNet)的隔膜泵单向阀故障诊断方法.首先,将注意力机制引入了残差神经网络以提升网络的学习能力,自适应调节了重要但微弱特征权重,并以恒等变换减少了有效信息被抑制现象;其次,提出了加权"监督对比损失(SCL)+交叉熵(CE)损失",调节单向阀不同故障状态数据之间的距离,明确了单向阀不同故障状态的分类边界与降低噪声或环境激励的干扰;最后,通过工程实测数据,对监督对比学习和HA-ResNet融合方法的有效性和稳定性进行了验证.研究结果表明:监督对比学习和HA-ResNet融合方法在隔膜泵单向阀验证集上的平均准确率达到了99.3%;与其他故障诊断方法相比,其在诊断精度和稳定性上都具有一定的优势,验证了该方法在噪声干扰条件下故障诊断的可靠性.
Diaphragm pump check valve fault diagnosis method based on supervised contrastive learning and hybrid attention ResNet
In the industrial production environment,strong noise and other environmental stimuli result in similarities in the characteristics of different faults in the diaphragm pump check valve,making it difficult for traditional deep learning methods to accurately identify the valve's fault status.To solve this problem,a method combining supervised contrastive learning and hybrid attention residual neural network(HA-ResNet)was proposed for fault diagnosis of diaphragm pump check valves.Firstly,the attention mechanism was introduced into the residual neural network to enhance the network's learning capabilities.The weights of important but weak features were adaptively adjusted by the attention mechanism,and the suppression of useful information through identity transformations was reduced.Secondly,a weighted"supervised contrastive loss(SCL)+ cross-entropy(CE)loss"was proposed to adjust the distances between different fault states of the check valve.The classification boundaries for different fault states were clarified by this method and the interference of noise or environmental stimuli was reduced.Finally,the effectiveness and stability of the method combining supervised contrastive learning and HA-ResNet were verified using real measurement data.The results show that the average accuracy of supervised contrastive learning and HA-ResNet fusion method reaches 99.3%on the validation set of the diaphragm pump check valve.Comparing with other fault diagnosis methods,supervised contrastive learning and HA-ResNet fusion method has advantages in diagnostic accuracy and stability.The reliability of this method for fault diagnosis under noise interference conditions has been verified.

diaphragm pumpcheck valvefault diagnosissupervised contrastive loss(SCL)hybrid attention residual neural networks(HA-ResNet)feature similaritydeep learning method

任洪兵、彭宇明、黄海波

展开 >

西南交通大学 汽车与能源动力研究所,四川 成都 610036

西南交通大学 先进驱动节能技术教育部工程研究中心,四川 成都 610036

隔膜泵 单向阀 故障诊断 监督对比损失 混合注意力残差神经网络 特征相似性 深度学习方法

四川省自然科学基金资助项目四川省科技成果转移转化示范项目

2023NSFSC03952022ZHCG0061

2024

10.3969/j.issn.1001-4551.2024.04.004

机电工程
浙江大学　浙江省机电集团有限公司

机电工程

CSTPCD北大核心
影响因子：0.785
ISSN：1001-4551
年,卷(期)：2024.41(4)
  • 25