振动与冲击第38卷第15期JOURNAL OF VIBRATION AND SHOCKVol. 38 No. 15 2024SADBN及其在滚动轴承故障分类识别中的应用杨宇，罗鹏，甘磊，程军圣(湖南大学汽车车身先进设计制造国家重点实验室，长沙410082)摘要：传统的智能诊断方法一般都是基于“特征提取+分类器”模型，其核心在于特征值的提取以及分类器的
设计。针对不同的诊断对象,通常需要根据先验知识提取不同的故障特征值,这必将给最终的诊断结果带来诊断误差;与
此同时，传统的分类器一般使用浅层模型，这使得其难以表征信号与装备运行状况之间复杂的映射关系。作为深度学习 算法典型代表之一的深度信念网络(Deep Belief Network,DBN)，可以直接从原始信号中提取特征并具有深度学习能力,
因而已受到越来越多研究者的关注。但是DBN依然存在网络结构需要人为设定的缺陷，这也限制了 DBN在工程实际中 的应用。为解决DBN网络结构难以确定及如何提升其在工程实际应用中的诊断效率问题,提出了一种新的深度信念网
络，艮卩结构自适应深度信念网络(Structure Adaptive Deep Belief Network,SADBN) o与DBN相比,SADBN可以自适应地确
定网络结构,有效提高诊断效率。对滚动轴承故障振动信号的分析结果表明了改进网络的有效性。关键词：深度学习；DBN；网络结构；SADBN；滚动轴承故障诊断中图分类号：TH113.1 文献标志码：A DOI： 10.13465/j. cnki. jvs. 2024.15.002SADBN and its application in rolling bearing fault identification and classificationYANG Yu9 LUO Peng, GAN Lei. CHENG Junsheng(State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body, Hunan University, Changsha 410082, China)Abstract: Traditional fault diagnosis methods are based on a feature extraction plus classifier ” model, and its
core is feature values9 extraction and classifier design. Aiming at different diagnosis objects, different fault feature values should be extracted usually according to priori knowledge, these bring diagnosis errors to the final diagnosis results. At the
same time, the traditional classifier generally uses a shallow layer model, and this makes it difficult to characterize a complex mapping relation between signals and equipment operating conditions. As a typical representative of the deep learning algorithm, the deep belief network ( DBN) can be used to directly extract features from original signals and it has a deep learning ability, so it receives more and more attentions of researchers. But the DBN has a disadvantage of its
network structure needing artificial setting to limit 辻s application in engineering practice. Here, in order to solve problems
of its structure being difficult to set and how to improve its diagnosis efficiency in engineering practical application, a new
DBN named the structure adaptive deep belief network (SADBN) was proposed. It was shown that compared with DBN,
SADBN can be used to determine a network' s structure adaptively, and effectively improve its fault diagnosis efficiency. The analysis results of rolling bearing fault vibration signals verified the effectiveness of the network improved with
SADBN.Key words: deep learning ； DBN; network structure; SADBN; rolling bearing fault diagnosis传统的智能诊断方法过于依赖提取的特征值以及 专家诊断经验知识，在面对如何实现大型化、高速化、复杂化装备系统准确、快速、便捷的在线监测与实时故 障诊断这类问题时，传统的智能诊断方法就显得有点
力不从心3］。因此，迫切需要研究新的方法来满足工
基金项目：国家自然科学基金(51575168； 51375152)；国家重点研发计
程实际的需求。Hinton等⑶在《Science》上提出深度学习理论，由
划项g (2016YFF0203400)；智能型新能源汽车国家2011协同创新
中心、湖南省绿色汽车2011协同创新中心资助收稿日期：2024-01 -19修改稿收到日期：2024 - 04 - 28此开启了机器学习在学术界和工业界的浪潮。深度学 习的宗旨在于构建深层次的网络结构模型，学习数据
第一作者杨宇女，博士，教授,1971年生通信作者罗鹏男，硕士生,1989年生E-mail:yzluopeng@ 163. com
中隐含的特征，获取数据丰富的内在信息。相较于传 统的智能诊断方法，深度学习方法有以下三大优
SADBN及其在滚动轴承故障分类识别中的应用
