STFT结合2D CNN-SVM的齿轮箱故障诊断方法  

Fault Diagnosis Method of Gearbox Based on STFT and 2DCNN-SVM

作　　者：谢锋云 汪淦[1,2] 王玲岚 李刚 朱海燕 谢三毛 XIE Fengyun;WANG Gan;WANG Linglan;LI Gang;ZHU Haiyan;XIE Sanmao(School of Mechanical Electronics and Vehicle Engineering,East China Jiaotong University,Nanchang 330013,China;Life-cycle Technology Innovation Center of Intelligent Transportation Equipment,East China Jiaotong University,Nanchang 330013,China)

机构地区：[1]华东交通大学机电与车辆工程学院,南昌330013 [2]华东交通大学智能交通装备全寿命技术创新中心,南昌330013

出　　处：《噪声与振动控制》2024年第4期103-109,共7页Noise and Vibration Control

基　　金：国家自然科学基金资助项目(52265068);江西省自然科学基金资助项目(20224BAB204050);载运工具与装备教育部重点实验室资助项目(KLCEZ2022-02);江西省教育厅资助项目(GJJ2200627);江西省研究生创新专项资金资助项目(YC2022-s481)。

摘　　要：为提高齿轮箱故障诊断的有效性和故障识别的准确率,提出一种基于短时傅里叶变换(Short-term Fourier transform,STFT)、二维卷积神经网络(Two-dimensional Convolutional Neural Network,2D CNN)和支持向量机(Support Vector Machine,SVM)相结合的齿轮箱故障识别方法。搭建JZQ250型定轴齿轮箱实验平台,利用加速度传感器获得齿轮箱振动信号,并对振动信号进行短时傅里叶变换得到二维时频图,然后将时频图输入到2D CNN中进行特征信息提取,通过2D CNN前向传播和反向传播对不同类别故障时频图信息进行训练,建立不同类别特征之间更深层次的联系,通过训练集和验证集loss曲线、准确率曲线和t-SNE可视化(t-Distributed Stochastic Neighbor Embedding,t-SNE)多种方法来反映模型训练程度,最后由SVM对故障类型进行识别。通过将所提出的方法与FFT-2D CNN、1D CNN-SVM和2D CNN-SVM对齿轮箱故障识别结果进行对比,本方法故障识别准确率最高,达到97.94%,且提出的方法具有很好的鲁棒性。In order to improve the effectiveness and the accuracy of gearbox fault diagnosis,a gearbox fault identification method based on the combination of short-term Fourier transform(STFT),two dimensional convolutional neural network(2D CNN)and support vector machine(SVM)is proposed.The JZQ250 fixed axis gearbox experimental platform is built.The acceleration sensor is used to collect the gearbox vibration signal,and the STFT is performed on the vibration signal to obtain a two-dimensional time-frequency map.Then,the time-frequency map is input into the 2D CNN for feature information extraction.The time-frequency map information of different types of faults is trained through 2D CNN forward propagation and backward propagation to establish a deeper relationship between different types of features.The training set and verification set loss curves,accuracy curve and t-Distributed Stochastic Neighbor Embedding(t-SNE)visualization are used to reflect the training degree of the model.Finally,SVM is used to identify the fault type.By comparing the fault identification results of gearbox with those of FFT-2D CNN,1D CNN-SVM and 2D CNN-SVM,the proposed method has the highest accuracy of fault identification of 97.94%,as well as good robustness.

关 键 词：故障诊断 齿轮箱 短时傅里叶变换 二维卷积神经网络 支持向量机 

分 类 号：TH113.1[机械工程—机械设计及理论]