融合图像方法的半实时轴承故障诊断方法  

Semi-Real-Time Bearing Fault Diagnosis Method Combined Image Method

作　　者：王鹏志 张满囤 韩亚洪[1,2] 赵旭 王征 Wang Pengzhi;Zhang Mandun;Han Yahong;Zhao Xu;Wang Zheng(College of Intelligence and Computing,Tianjin University,Tianjin 300350;Tianjin Key Lab of Machine Learning,Tianjin 300050;School of Artificial Intelligence,Hebei University of Technology,Tianjin 300401;Tianjin International Joint Center for Virtual Reality and Visual Computing,Tianjin 300401)

机构地区：[1]天津大学智能与计算学部,天津300350 [2]天津市机器学习重点实验室,天津300050 [3]河北工业大学人工智能与数据科学学院,天津300401 [4]天津市虚拟现实与可视计算国际联合中心,天津300401

出　　处：《计算机辅助设计与图形学学报》2021年第6期895-903,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家重点研发计划(2019YFC1520600);国家自然科学基金(61573356);河北省自然科学基金(F2019202054).

摘　　要：轴承在高负荷的环境下长时间运转经常会发生故障,造成巨大损失.若能在故障发生前期进行准确感知,则可以在很大程度上减少损失.通过分析轴承故障问题的特点,提出一种半实时的高准确率诊断方法,使用双路宽卷积核深度卷积网络(deep convolutional neural networks with double paths and wider kernels,DWDCNN)作为实时诊断算法,在结果出现异常或轴承处于高噪声环境下的时候对轴承的振动数据使用短时傅里叶变换(short time Fourier transform,STFT),将其转换为图像,再使用轻型多尺度胶囊网络(smaller inception capsule net,SICN)进行二次诊断.使用该算法与现有其他算法在凯斯西储大学(Case Western Reserve University,CWRU)数据集和添加不同强度噪声后的CWRU数据集上进行实验,对准确性和计算效率进行对比.结果显示DWDCNN模型使用0.12 ms即可完成一次预测,且在SNR=−4 dB的条件下达到80.07%的准确率,而SICN模型虽然计算时间较长,但是可以在SNR=−4 dB的条件下达到95.00%的准确率.Bearings running in the high load environment for a long time often malfunction,resulting in huge losses.It can be reduced to a large extent if the fault can be detected accurately in the early stage.According to the analysis on characteristics of the bearing fault problem,a semi-real-time and high-accuracy diagnosis method is proposed.First,deep convolutional neural networks with double paths and wider kernels(DWDCNN)are used as a real-time diagnosis method.When the result looks abnormal or the bearing is in a high-noise environment,short-time Fourier transform(STFT)is used on the vibration data of the bearing to convert it to image,and smaller inception capsule net(SICN)is used for secondary diagnosis.Then a comparison experiment between the proposed models and other existing models on Case Western Reserve University(CWRU)dataset and CWRU dataset with different intensities of noise is made on the aspects of accuracy and time performance.The result shows that DWDCNN can accomplish one prediction within 0.12 ms,and the accuracy can be achieved of 80.07%under the condition of SNR=−4 dB.Although using more time,the accuracy of SICN can be achieved of 95.00%under the condition of SNR=−4 dB.

关 键 词：轴承 故障诊断 振动信号 卷积神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]