Abstract: The real-time monitoring of rotating machinery health is very important,and the rolling bearing fault is the focus of research.It is difficult to diagnose the machinery with complex structure efficiently and accurately by traditional fault diagnosis method.Deep learning develops rapidly in the field of mechanical fault diagnosis due to its strong ability of data analysis and learning.To improve the diagnosis accuracy of traditional Convolutional Neural Network (CNN),a CNN-LSTM model with Batch Normalization (BN) was proposed by considering the defects of Long Short Term Memory Network (LSTM) in diagnosis.Through testing with the CWRU bearing dataset,the results demonstrated that the diagnosis accuracy and efficiency of the hybrid model by batch normalization improved.The proposed method obtained a diagnosis result superior to the traditional deep learning fault diagnosis methods and could efficiently and accurately diagnose faults at various positions and degrees under various loads.

Key words: rolling bearing, fault diagnosis, convolutional neural networks, long short term memory neural networks, batch normalization

摘要： 旋转机械健康状况的实时监测十分重要,其中滚动轴承故障更是研究的重点。传统的故障诊断方法难以高效准确地诊断出结构复杂的机械故障,而深度学习强大的数据分析和学习能力,使其在机械故障诊断领域发展迅速。为了提高传统卷积神经网络(CNN)在诊断中的准确度,并考虑到长短记忆网络(LSTM)诊断时间较长的缺陷,提出一种批标准化(BN)的CNN-LSTM模型。在凯斯西储大学轴承数据集上的实验表明,批标准化提高了该混合模型的诊断精度和效率。该方法获得了优于传统深度学习故障诊断方法的诊断结果,能够高效准确地进行多种负荷下、多种故障位置以及多种故障程度的诊断。

关键词: 滚动轴承, 故障诊断, 卷积神经网络, 长短时记忆网络, 批标准化