融合注意力机制的改进DBN变工况齿轮箱故障诊断方法  

Improved DBN method with attention mechanism for the fault diagnosis of gearboxes under varying working condition

作　　者：张智禹 尹爱军[1] 谭建[2] ZHANG Zhiyu;YIN Aijun;TAN Jian(State Key Laboratory of Mechanical Transmissions,College of Mechanical Engineering,Chongqing University,Chongqing 400044,China;China Petroleum Xinan Oil&Gas Field Company Chongqing Gas Mine,Chongqing 400021,China)

机构地区：[1]重庆大学机械工程学院机械传动国家重点实验室,重庆400044 [2]中国石油西南油气田分公司重庆气矿,重庆400021

出　　处：《振动与冲击》2021年第14期47-52,共6页Journal of Vibration and Shock

基　　金：重庆市科技重大主题专项重点研发项目(cstc2018jszx-cyztzxX0032)。

摘　　要：针对齿轮箱在交变工况下运行时导致的故障模式难以识别、分类精度降低的问题,提出融合注意力机制的改进深度置信网络(DBN)变工况齿轮箱故障诊断方法。为解决齿轮箱单一时、频域特征反应故障信息不全面、异常不敏感问题,提取时域、频域、小波包时频域特征形成高维特征集。利用深度置信网络具有的贪心学习优势分别对其进行挖掘,同时结合注意力机制自适应对描述齿轮箱状态有效的特征给予更多“注意”,从而提高齿轮箱故障诊断精度。引进余弦损失函数降低深度置信网络对不同工况振动强度的敏感性,从而减轻网络拟合负担、提高泛化能力。齿轮箱变工况故障诊断试验结果表明,所提方法有效提高了变工况下齿轮箱故障诊断精度,同时具有很好的泛化能力。Aiming at the fact that the fault mode of gearboxes running under varying working condition is difficult to identify and the classification accuracy is reduced,an improved deep belief network(DBN)method with attention mechanism was proposed.Firstly,in order to solve the problem that single time domain or frequency domain characters are not comprehensive and insensitive for responding to gearbox fault information,the time domain,frequency domain and wavelet packet time-frequency domain features were extracted and synthesized to form a high-dimensional feature set.Then,making use of the greedy learning advantage of the DBN,the features were separately mined further.At the same time,the attention mechanism was used to adaptively give more attention to the features that effectively describe the gearbox state to improve the accuracy of gearbox fault diagnosis.Finally,the cosine loss function was introduced to reduce the sensitivity of the deep confidence network to the vibration intensity under different working conditions,thereby reduce the network fitting burden and improve the generalization ability.The fault diagnosis tests of a gearbox under varying working condition show that the proposed method effectively improves the fault diagnosis accuracy of the gearbox,and has good generalization ability.

关 键 词：注意力机制 余弦损失函数 深度置信网络(DBN) 齿轮箱 变工况故障诊断 

分 类 号：TH212[机械工程—机械制造及自动化]