Fault Diagnosis Of Printing Press Bearings Based On Deformable Convolution Residual Neural Network

The indispensable rolling bear,a key element in printing presses,is essential for minimizing coefficient of motion friction and guaranteeing precision;their functioning state has a direct bearing on the printing press’s overall performance.Due to the corrosion of chemical substances on the rolling bear,or due to insufficient lubrication,it is easy to cause bearing failure.The bearing’s vibration signal can be a direct indicator of the machine’s functioning,thus necessitating the observance of the bearing’s operational state and the detection of any faults in a timely manner.The bearing vibration signal’s fault features are frequently obscured by the intense background noise in actual working conditions.The intricate working atmosphere,with its powerful background noise,is likely to cause issues in recognizing bearing features,and the network has issues such as gradient vanishing and over-fitting.Therefore,it is necessary to study the bearing fault diagnosis method under strong background noise.Therefore,this paper fuses the advantages of frequency slicing wavelet transform,deformable convolutional layer and residual neural network to propose an intelligent diagnosis method for printing machine bearing faults based on strong background noise,and the specific research contents are as follows:(1)In order to solve the problem of traditional preprocessing methods which need choosing the appropriate window function or wavelet basis function,modal confusion and endpoint effect,frequency slicing wavelet transform has been proposed to perform time-frequency analysis of bearing vibration signals,reduce the dependence of wavelets and wavelet packets on basis functions in the reconstruction process,and realize the reconstruction of signals in any frequency band and the accurate description of local features of signals.Through frequency slicing wavelet transform,the corresponding time-frequency map is generated,and the bearing fault characteristic information is effectively extracted,which lays the foundation for the subsequent intelligent bearing fault diagnosis model to accurately extract bearing faults.(2)Based on the fact that the traditional network model is prone to gradient disappearance and overfitting,the residual neural network as the bearing fault diagnosis model has been proposed,which solve the problems existing in the traditional network model.The performance of the residual neural network bearing fault diagnosis model is verified through experiments,and the results show that the average accuracy of the fault diagnosis model reaches 99.77%,and the average computing time is quicker,which proves the effectiveness of this proposed method.In order to improve the traditional residual neural network with complex and small fault features of strong background noise vibration signals,a deformable convolution structure to reconstruct the convolutional layers of the residual neural network has been proposed,and uses the adaptive shape of the deformable convolutional layers to improve the feature extraction ability of the network model,so that the model can learn more detailed features.The results show that the average accuracy of the deformable convolutional residual neural network model under strong background noise reaches 81.30%,and the performance is better than other fault diagnosis models.(3)In this paper,an experimental platform for printing press bearing fault diagnosis has been built,collect experimental data and conduct experimental validation for the characteristics of strong noise of printing press bearing fault vibration signal under actual working conditions.The results show that the bearing fault diagnosis model integrating frequency slicing wavelet transform,deformable convolution and residual neural network can achieve the classification of different faults of rolling bearings under strong background noise with an accuracy of 93.90%and good stability.