摘要
推力轴承是核主泵关键零部件,其碰磨故障会导致轴瓦严重磨损甚至轴瓦烧焦,严重危害核主泵服役安全。考虑到核主泵运行工况复杂多变且实际核主泵推力轴承碰磨样本十分稀缺,导致传统基于机器学习的监测方法难以实施。为解决该问题,提出一种有效利用少量样本数据实现领域自适应的域对抗迁移学习模型,并应用于推力轴承碰磨监测。首先,根据核主泵轴瓦材质和轴承常见运行工况,开展乌金层材料摩擦基础试验并收集足量故障模拟数据作为特征迁移的源域数据;其次,开展推力轴承润滑失效碰磨试验,收集少量真实故障数据作为目标域数据;再次,对上述信号进行预加重及滤波处理,通过CNN-GRU网络将源域与目标域特征投射到同一特征空间,实现多领域特征提取与适配;最后,通过智能诊断网络完成变工况下未知标签推力轴承碰磨故障识别。实验结果表明,所提方法在可用样本较少时能准确有效实现推力轴承碰磨故障诊断,识别准确率可达100%,多次重复实验平均准确率相较TCA、DAN等方法提高了17.46%以上。
Thrust bearing is one of the key components of the primary pump,and its rub impact failure will lead to serious wear of the axial tile or even axial tile burning,which seriously jeopardizes the safety of the primary pump in service.Considering the complex and variable operating conditions of primary pumps,and the scarcity of actual samples of primary pump thrust bearing rub impact,traditional machine learning-based monitoring methods are difficult to implement.To solve this problem,a domain adversarial transfer learning model that effectively utilizes a small amount of sample data to achieve domain adaption is proposed and applied to thrust bearing rub impact monitoring.Firstly,according to the primary pump shaft tile material and common operating conditions of bearings,basic friction experiments on babbitt metal were carried out to collect sufficient failure simulation data as the source domain data for feature transfer.Secondly,rub impact experiments on lubrication failure of thrust bearings were conducted to collect a small amount of real failure data as target domain data.Then the above signal is pre-emphasized and filtered,and the CNN-GRU(Convolutional Neural Networks-Gate Recurrent Unit)network is used to project the source and target domain features into the same feature space to realize the multi-domain feature extraction and adaptation.Finally,the identification of thrust bearing rub impact faults under variable operating conditions with unknown labels is accomplished by the intelligent diagnosis network.The experimental results show that the proposed method can accurately and effectively realize the diagnosis of thrust bearing rub impact faults when there are fewer available samples,and the highest accuracy rate of identification reaches 100%,and the average accuracy rate of ten repetitive experiments is improved by more than 17.46%compared with TCA and DAN methods.
作者
赵青松
陈兴江
王琇峰
叶泉流
区瑞坚
王学灵
ZHAO Qingsong;CHEN Xingjiang;WANG Xiufeng;YE Quanliu;QU Ruijian;WANG Xueling(College of Mechanical Engineering,Xi'an Jiaotong University,Xi'an,Shaanxi Prov.710049,China;Suzhou Veizu Equipment Diagnosis Technology Co.,Ltd.,Suzhou,Jiangsu Prov.215211,China;State Key Laboratory of Nuclear Power Safety Monitoring Technology and Equipment of China Nuclear Power Engineering Co.,Ltd.Shenzhen,Guangdong Prov.518172,China)
出处
《中国核电》
2025年第2期181-188,共8页
China Nuclear Power
基金
自然基金委联合项目资助(U2267206)。
作者简介
赵青松(2000-),男,山西运城人,工程师,硕士,研究方向:旋转机械故障诊断(E-mail:zqs14735014711@163.com)。
