基于深度置信网络的旋转部件半监督故障诊断  

Semi-supervised fault diagnosis on rotating parts based on the deep belief network

作　　者：马航宇 周笛 潘尔顺[1] MA Hang-yu;ZHOU Di;PAN Er-shun(School of Mechanical Engineering,Shanghai Jiao Tong University,Shanghai 200240)

机构地区：[1]上海交通大学机械与动力工程学院,上海200240

出　　处：《机械设计》2021年第12期1-6,共6页Journal of Machine Design

基　　金：国家重点研发计划项目(2020YFB1711100);国家自然科学基金项目(52005327,72001138,72071127)。

摘　　要：机械旋转部件故障诊断中,时变的噪声干扰和动态的运转状态直接影响系统故障诊断正确率。文中基于深度置信网络(DBN)提取数据深层信息的能力,结合固定学习步长的信号分解技术提取故障特征;进一步针对DBN作为无监督学习的不足,结合奇异值分解技术,将监督信息融入到特征训练强化过程中,建立半监督视角下深度置信网络(S-DBN)分析架构,降低噪声干扰,实现旋转机械系统的智能故障诊断;最后,结合变工况滚动轴承系统采集数据,基于S-DBN计算不同程度信号干扰下智能故障诊断平均正确率为97.14%。与传统DBN网络等4种方法作比较,不同程度噪声干扰下,S-DBN具有较好的准确性、泛化性和抗噪性。For the fault diagnosis on mechanical rotating parts, the time-varying noise and dynamic operating status affect the correct rate directly. Based on the ability of deep belief network to extract deep information and combined with the signal decomposition technology of fixed learning step length, data characteristics of fault can be represented. Besides, in combination with singular value decomposition, supervision information can be integrated for the feature training process to make up for the disadvantage of DBN as unsupervised learning. The semi-supervised perspective of S-DBN can be established to reduce noise interference and identify the intelligent fault diagnosis on the rotating parts. Finally, with the consideration on the rolling bearings under variable conditions, the average correct rate of intelligent fault diagnosis under different degrees of signal interference is calculated with the help of S-DBN as 97.14%. Also, the results are compared with those which are worked out by means of the four methods, for example, the traditional DBN network. With the disruption from different noise, S-DBN has better accuracy, generalization and anti-noise capability.

关 键 词：故障智能诊断 半监督 固定学习步长 深度置信网络 噪声干扰 

分 类 号：TH17[机械工程—机械制造及自动化]