基于迁移学习的滚动轴承在线故障诊断  

On-line Fault Diagnosis of Rolling Bearing Based on Transfer Learning

作　　者：毛冠通 洪流 王景霖 Mao Guantong;Hong Liu;Wang Jinglin(Mechanical and Electronic Engineering,Wuhan University of Technology,Wuhan 430070,China;Aviation Key Laboratory of Science and Technology on Fault Diagnosis and Health Management,Shanghai Aero Measurement&Control Technology Research Institute,Shanghai 201601,China)

机构地区：[1]武汉理工大学机电工程学院,湖北武汉430070 [2]航空工业上海航空测试技术研究所故障诊断与健康管理技术航空科技重点实验室,上海201601

出　　处：《航空科学技术》2020年第1期61-67,共7页Aeronautical Science & Technology

基　　金：航空科学基金(20173365001)~~

摘　　要：故障检测与诊断是工业设备稳定、可靠、安全运行的关键。近年来,由于深度学习模型具有自动特征识别的能力,在数据驱动故障检测方法中得到广泛的应用。一般来说,深度学习模型都是针对同一问题的历史数据进行训练,然后使用同源新数据在训练好的模型里进行检测。而迁移学习能够有效地解决目标域与源域中不同但相似的问题,因此本文提出一种基于迁移学习的卷积神经网络(TCNN)在线故障诊断方法。首先,通过短时傅里叶变换(STFT)将时域信号数据转换为包含丰富信息的频域图像,作为适合卷积神经网络(CNN)的输入。然后,构建在线CNN,在线CNN能够从频域图像中自动提取特征并进行故障分类。最后,为了提高在线CNN的实时性,还构建了多个离线CNN,并对相关数据集进行了预训练。通过将离线CNN的浅层结构迁移到在线CNN,在线CNN可以显著提高实时性,成功地解决了在有限训练时间内达到期望的诊断精度问题。提出的方法在CWRU大学轴承数据集上进行了验证,达到了预期的诊断精度。Fault detection and diagnosis is the key to the reliable and safe operation of industrial equipment.In recent years,deep learning algorithms have been widely used in data-driven fault detection because of its capability of automatic feature recognition.Generally,deep learning algorithms are trained by historical data for the same problem and then tested in trained models using homologous new data.As the transfer learning has the potential to address the problems that are different but still similar with each other in target domain and source domain,this paper proposes an online fault diagnosis method based on a deep Transfer Convolutional Neural Network(TCNN)framework.First,the time-domain signal data is transformed into a time-frequency domain containing rich information by Short Time Fourier Transform(STFT),which serves as the input that is suitable for Convolutional Neural Network(CNN).Then,an online CNN network is constructed,which can automatically extract features from frequency-domain images and classify faults.Finally,in order to improve the real-time performance of online CNN,several offline CNN are constructed and the relevant data sets are pre-trained.By transferring the shallow structure of offline CNN to online CNN,online CNN can significantly improve the real-time performance and successfully solve the problem of achieving the expected diagnostic accuracy within the limited training time.The proposed method is verified on the bearing fault data set of CWRU Bearing Data Center,and achieve the expected diagnostic accuracy.

关 键 词：卷积神经网络 在线故障诊断 短时傅里叶变换 迁移学习 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]