基于深度学习波形图像识别的轴承故障诊断方法  

Fault diagnosis method for bearing based on deep learning waveform image recognition

作　　者：余刃[1] 谢旭阳 王天舒 彭俏[1] 陈玉昇 YU Ren;XIE Xu-yang;WANG Tian-shu;PENG Qiao;CHEN Yu-sheng(College of Nuclear Science and Technology, Naval Univ. of Engineering, Wuhan 430033, China)

机构地区：[1]海军工程大学核科学技术学院,武汉430033

出　　处：《海军工程大学学报》2021年第4期76-82,共7页Journal of Naval University of Engineering

基　　金：海军工程大学科研发展基金资助项目(425317k304)。

摘　　要：为解决传统的轴承故障诊断方法需要大量专业知识的问题,提出了一种基于深度学习振动信号波形图像识别的轴承故障在线自动诊断方法。首先,采用InceptionV3模型作为预训练模型,设计了深度学习和迁移学习相结合的InceptionV3模型训练方法;然后,采用交叉熵作为损失函数,用于评价模型训练效果,给出了进行故障诊断的方法步骤,并用轴承在正常和不同故障状态时的振动数据,开展了方法有效性验证实验;最后,采用主成分分析法分析了InceptionV3模型提取的特征参数对不同故障模式的聚类效果,并通过对比分析运用与不运用迁移学习时InceptionV3模型的训练次数和训练时间,验证了迁移学习方法对模型训练速度的改善效果。结果表明:所提出的方法对不同故障状态有较高的识别精度,并且可以降低计算资源的要求,保证诊断过程的实时性。Aiming at the problem of traditional bearing fault diagnosis methods requiring a large amount of expertise,online automatic fault diagnosis method based on deep learning waveform image recognition of vibration signal was proposed.The InceptionV3 model was used as the pre-training model,and the training method for InceptionV3 model based on deep learning and transfer learning was designed.Cross-entropy was used as the loss function to evaluate the training effect of InceptionV3 model.The fault diagnosis steps were given,and the method was tested with vibration data of bearings in normal and different fault states.The clustering effect of CNN features for different fault states was analyzed through PCA.And the training steps and training time of InceptionV3 models with and without transfer learning was compared to verify the improvement on training speed.The results show that the proposed method has high recognition accuracies for different fault states.And the method requires less computing resource while ensuring real-time fault diagnosis of bearing.

关 键 词：轴承故障论断 InceptionV3 深度学习 波形图像识别 

分 类 号：TH17[机械工程—机械制造及自动化]