基于预训练GoogleNet模型和迁移学习的齿轮箱故障检测方法  

Gearbox fault detection based on pre-trained GoogleNet modal and transfer learning

作　　者：杨魏华 阮爱国 黄国勇[1] YANG Weihua;RUAN Aiguo;HUANG Guoyong(Faculty of Civil Aviation and Aviation,Kunming University of Science and Technology,Kunming 650500,China;Wenshan Miao Zhuang Autonomous Prefecture Advanced Technical School,Wenshan 663000,China;China General Nuclear Power New Energy Holding Co.,Ltd.,Yunnan Branch,Kunming,650000,China)

机构地区：[1]昆明理工大学民航与航空学院,云南昆明650500 [2]文山苗族壮族自治州高级技工学校,云南文山663000 [3]中国广核新能源控股有限公司云南分公司,云南昆明650000

出　　处：《机电工程》2024年第2期262-270,共9页Journal of Mechanical & Electrical Engineering

基　　金：云南省科技厅重大专项(KKBD202265023)。

摘　　要：针对以往齿轮箱故障诊断中特征处理算法繁琐、人为因素影响较大等问题,提出了一种基于预训练GoogleNet模型和迁移学习(TL)的故障诊断方法。首先,利用连续小波变换(CWT)将离散时间序列转变为二维小波尺度图,构建了样本集;然后,对预训练模型进行了结构微调及参数微调使其符合任务需求,利用处理得到的训练样本对微调后的模型进行了微训练,使其达到理想精度,然后保存模型,再将其应用于故障分类任务;最后,为了对上述模型的可行性进行验证,利用昆明理工大学控制与优化重点实验室的平行齿轮箱数据以及东南大学的行星齿轮箱数据对微调模型进行了验证。研究结果表明:相比于传统卷积神经网络(CNN)以及未经预训练的GoogleNet模型,基于预训练GoogleNet模型和迁移学习的故障诊断方法在训练样本较少的情况下,其分类准确率均值仍然高达97.40%,且模型的收敛速度更快,对计算机算力的依赖程度更低。微调模型高层的方法能根据任务分类情况个性化设置模型输出,因此该模型能够适用于不同的场景。In order to address the issues of cumbersome feature processing algorithms and the significant human factors in the diagnosis of gearbox,a fault diagnosis method based on a pre-training model and transfer learning(TL)was proposed.Firstly,the continuous wavelet transform(CWT)was used to convert discrete time series into two-dimensional wavelet scale images to construct a sample set.Then,the pre-trained model was structurally fine-tuned and parameter-tuned to meet the task requirements.The fine-tuned model was further trained with processed training samples until the desired accuracy was achieved and the corresponding model was saved and applied to fault classification tasks.Finally,the fine-tuned model was validated by using parallel gearbox data from key laboratory of control and optimization of Kunming University of Science and Technology and the planetary gearbox data of Southeast University.The research results demonstrate that comparing to traditional convolutional neural networks(CNN)and non-pretrained GoogleNet models,the proposed model achieves an average classification accuracy of 97.40%with limited training samples.Additionally,the proposed model exhibits the characteristics of faster convergence speed,lower dependence on computational power since upper layers of model is modified.The method of fine-tuning the upper level of the proposed model can personalize output according to the task classification,so that the proposed model can be applied in different scenarios.

关 键 词：变速器 预训练网络 迁移学习 连续小波变换 尺度图 卷积神经网络 

分 类 号：TH132.41[机械工程—机械制造及自动化]