| With the development of industrialization,mechanical equipment plays an important role in human living environment.Bearing is the core component of mechanical equipment,and its health status is related to the operation of the whole mechanical equipment and even the whole factory.Therefore,the research on bearing fault diagnosis has important theoretical significance and application value.The traditional fault diagnosis method needs high professional knowledge and complex feature extraction process,and its development in the field of fault diagnosis is limited.Deep learning has certain advantages in the field of vision,which opens up a new way for the field of fault diagnosis.Compared with one-dimensional vibration signal,the research of deep learning fault diagnosis method based on two-dimensional image is less.Therefore,this paper mainly studies the fault diagnosis method based on the combination of two-dimensional image and deep learning.The research contributions of this paper are as follows:(1)A bearing fault diagnosis method CWT-MDRSN based on the combination of continuous wavelet transform time-frequency image and multiscale depth residual shrinkage network is proposed.Aiming at the problem of low accuracy of bearing fault diagnosis under the background of noise,CWT-MDRSN takes Res Net18 as the infrastructure and introduces soft thresholding under the attention mechanism,which can enhance key features and eliminate redundant features.And multiscale dilated convolution is introduced,and its position in the network is studied to obtain multi-scale feature information.Experiments show that the proposed method can effectively suppress the interference of noise on vibration signal.(2)A bearing fault diagnosis method based on multi-sensor fusion time-frequency diagram and Shuffle Net2 MC is proposed.Aiming at the problem that most deep neural networks have many parameters and high computing power,and can not be used for real-time diagnosis in embedded devices with limited hardware conditions,the method of multi-sensor fusion can express comprehensive fault characteristics.Shufflenet2 MC is based on Shuffle Netv2 and introduces mixed convolution of different sizes,which can improve the feature extraction ability of the network.The knowledge distillation theory is also introduced,and the more complex depth residual shrinkage network(DRSN50)is used to help the training of the basic network,which improve the parameter utilization of the basic network.Experimental results show that the parameters,computational power and accuracy of this method are better than other comparison algorithms. |
