变工况下轴承故障的残差对抗网络诊断方法  

Residual Adversarial Neural Network Diagnosis Method for Bearing Faults under Variable Working Conditions

作　　者：洪晓翠 段礼祥[1] 徐继威[2] 付强[2] Hong Xiaocui;Duan Lixiang;Xu Jiwei;Fu Qiang(College of Safety and Ocean Engineering,China University of Petroleum(Beijing);PetroChina Tarim Oilfield Company)

机构地区：[1]中国石油大学(北京)安全与海洋工程学院 [2]中国石油塔里木油田分公司

出　　处：《石油机械》2022年第5期32-42,共11页China Petroleum Machinery

基　　金：中石油战略合作科技专项“海外长输油气管道灾害监测预警及动力设施诊断技术研究”(ZLZX2020-05-02)。

摘　　要：实际工程中传统的以恒定转速和平稳信号为前提的故障诊断方法难以有效提取故障特征,为此,将深度学习中的残差网络(ResNet)与迁移学习中的领域对抗网络(DANN)相结合,提出一种深度迁移方法——残差对抗网络(RANN)。RANN采用滑窗取样策略从原始振动信号中截取故障样本,构建了包含特征提取器、故障分类器和领域判别器的残差对抗网络,采用3个工况下的滚动轴承数据,共开展6组迁移诊断试验。研究结果表明:RANN相比于标准DANN,特征提取及故障诊断效果均有所改善,平均准确率提升了约2.5百分点;该残差对抗网络通过特征提取器与领域判别器的对抗训练,可以自适应逐层提取对工况信息敏感度低的域不变特征;相比于单通道输入,采用双通道输入平均故障诊断准确率提升了约1.3百分点。所得结论可以为变工况机械设备的故障诊断提供参考。In practical engineering,it is hard to extract fault features effectively with the traditional fault diagnosis method based on constant speed and stable signal.Therefore,this paper proposed the residual adversarial neural network(RANN),which adopts the sliding window sampling strategy to intercept fault samples from the original vibration signal,and constructs a residual adversarial neural network including feature extractor,fault classifier and domain discriminator,and a total of 6 groups of transfer diagnosis tests were carried out using rolling bearing data under three working conditions.The results show that:compared with the standard domain adversarial neural network(DANN),RANN is better in feature extraction and fault diagnosis performance,with the average accuracy increased by about 2.5%.Through the adversarial training of the feature extractor and domain discriminator,RANN can adaptively extract domain-invariant features with low sensitivity to working condition information layer by layer.It uses dual-channel input,allowing the average fault diagnosis accuracy about 1.3%higher than single-channel input.The conclusions can provide a new method for fault diagnosis of mechanical equipment under variable working conditions.

关 键 词：残差对抗网络 故障诊断 迁移学习 信号融合 滚动轴承 变工况 

分 类 号：TE93[石油与天然气工程—石油机械设备]