基于BPSO-M1DCNN的行星齿轮箱故障诊断方法研究  

Fault diagnosis method of planetary gearbox based on BPSO-M1DCNN

作　　者：郭永伦 吴国新[1] 刘秀丽[1] 徐小力[1] GUO Yong-lun;WU Guo-xin;LIU Xiu-li;XU Xiao-li(The Ministry of Education Key Laboratory of Modern Measurement and Control Technology,Beijing Information Science&Technology University,Beijing 100192,China)

机构地区：[1]北京信息科技大学现代测控技术教育部重点实验室,北京100192

出　　处：《机电工程》2021年第10期1277-1283,共7页Journal of Mechanical & Electrical Engineering

基　　金：国家重点研发计划资助项目(2020YFB1713203);北京学者计划资助项目(2015-025);北京市教委科研计划资助项目(KM202011232001)。

摘　　要：采用卷积神经网络对旋转部件进行故障诊断时,其对多尺度的故障特征利用有限,且网络层结构和超参数调试费时费力,针对上述问题,提出了一种基于离散二进制粒子群优化多尺度一维卷积神经网络的BPSO-M1DCNN算法。首先,对M1DCNN网络进行了初始化设计,采用了BPSO算法自适应调整超参数和网络结构构建BPSO-M1DCNN网络;然后,将原始振动数据输入BPSO-M1DCNN网络,进行了特征学习和提取,将学习到的故障特征进行了分类输出;最后,将该算法应用于行星齿轮箱的故障诊断试验,并将其结果与用BPSO-BP神经网络、一维卷积神经网络、M1DCNN网络的结果进行了对比分析,利用变化曲线表示M1DCNN网络、BPSO-M1DCNN网络的正确率和损失率,采用混淆矩阵显示各类故障诊断精度,并利用T-SNE算法对其特征学习过程进行了可视化。研究结果表明:相比BPSO-BP神经网络、1DCNN网络、M1DCNN网络,基于BPSO-M1DCNN网络的行星齿轮箱测试集的平均准确率均有一定提升,应用于行星齿轮箱故障的诊断效果较好。When convolution neural network(CNN)was used in fault diagnosis of rotating parts,the utilization of multi-scale fault features was limited,and debugging of network layer structure and hyperparameter were time-consuming and laborious.Aiming at the above problems,a BPSO-M1DCNN algorithm was proposed,which was based on discrete binary particle swarm optimization(BPSO)to optimize multi-scale one dimensional convolution neural network(M1DCNN).Firstly,M1DCNN network was initialized to design,BPSO algorithm was used to adaptively adjust the hyperparameters and network structure to build the BPSO-M1DCNN network.The original vibration data were input to the BPSO-M1DCNN network for feature learning and extraction.Finally,the learned fault features were classified and output.The algorithm was applied to the fault diagnosis of planetary gearbox,and it was compared with network models such as BPSO-BP neural network,one-dimensional convolutional neural network(1DCNN),and M1 DCNN network.The change curve was used to express the accuracy and loss rate of M1DCNN network and BPSO-M1DCNN network.Confusion matrix was used to show the accuracy of all kinds of fault diagnosis.T-SNE algorithm was used to visualize the feature learning process.The results indicate that,the average accuracy of planetary gearbox test set based on BPSO-M1DCNN network is higher than that of BPSO-BP neural network,1DCNN network and M1DCNN network,and a better fault diagnosis effect is achieved.

关 键 词：行星齿轮箱 故障诊断 多尺度一维卷积神经网络 二进制粒子群优化 

分 类 号：TH132.41[机械工程—机械制造及自动化] TH17