基于深度卷积神经网络模型和XGBoost算法的齿轮箱故障诊断研究  

RESEARCH ON GEAR BOX FAULT DIAGNOSIS BASED ON DCNN AND XGBOOST ALGORITHM

作　　者：张荣涛 陈志高 李彬彬[1] 焦斌[1] ZHANG RongTao;CHEN ZhiGao;LI BinBin;JIAO Bin(Shanghai DianJi University,Shanghai 201306,China;China Nuclear Industry Maintenance Co.,Ltd.,Haiyan Branch,Haiyan 314300,China)

机构地区：[1]上海电机学院电气学院,上海201306 [2]中核检修有限公司海盐分公司,海盐314300

出　　处：《机械强度》2020年第5期1059-1066,共8页Journal of Mechanical Strength

基　　金：上海市科学技术委员会科研项目(17DZ1201200)资助。

摘　　要：针对齿轮箱的复合故障诊断问题,将深度卷积神经网络(Deep Convolutional Neural Network,DCNN)与XGBoost(e Xtreme Gradient Boosting)算法相结合,建立故障诊断模型。首先,利用深度卷积神经网络自适应提取原始振动加速度信号的特征矩阵。其次,将所得到的特征矩阵作为输入数据,运用网格调参法对XGBoost算法进行参数调整,得到XGBoost模型。最后,作为训练数据训练XGBoost模型,得到DCNN-XGBoost齿轮箱故障诊断模型。为了验证该模型的有效性和XGBoost算法的优越性,与DCNN-BP神经网络、DCNN-随机森林和DCNN-支持向量机三种模型作对比分析,并且对DCNN所得特征矩阵和人工提取的特征矩阵进行t-SNE可视化降维分析。结果表明,DCNN获得的特征矩阵可视化的效果优于人工提取的特征矩阵,并且随机森林的稳定性不如XGBoost算法,和BP神经网络相比,XGBoost算法在防止过拟合方面有一定的优势,SVM与DCNN的结合有其局限性,最后DCNN-XGBoost模型的诊断正确率和时间优于其他模型。In order to solve the problem of complex fault diagnosis of gearbox,the DCNN(Deep Convolution Neural Network)was combined with the XGBoost(e Xtreme Gradient Boosting)algorithm to establish the fault diagnosis model.Firstly,the DCNN Model was used to adaptively extract the feature matrix of the original vibration acceleration signal.Secondly,the feature matrix was used as input data,and the parameters of XGBoost algorithm were adjusted by lattice parameter method,then the XGBoost model was obtained.Most after that,the XGBoost model was trained by the feature matrix,so the gear box fault diagnosis model of DCNN-XGBoost was obtained.In order to verify the validity of the model and the superiority of XGBoost algorithm,the model was compared with three models:DNN-BP(Back Propagation neural network)model,DCNN-RF(Random Forest)model and DCNN-SVM(Support Vector Machine)model.The DCNN feature matrix and the artificial feature matrix were analyzed by t-SNE visualization algorithm,the results show that the visualization effect of DCNN feature matrix obtained is better than that of artificial feature matrix;Compared with XGBoost,the stability of Random Forest is not as good as that of XGBoost algorithm;Compared with BP neural network,XGBoost algorithm has some advantages in preventing over-fitting;The combination of SVM and DCNN has some limitations.Finally,the diagnostic accuracy and time of DCNN-XGBoost model is better than that of other models.

关 键 词：齿轮箱 故障诊断 卷积神经网络 XGBoost 

分 类 号：TH212[机械工程—机械制造及自动化] TH213.3