Research on fault diagnosis of rolling bearings in roller-to-roller printing units based on siamese network

Abstract

To realize a fault diagnosis of rolling bearings in a real R2R printing unit, a method based on Siamese Network is proposed in this work. First, vibration signals in rolling bearing were changed into a series of time–frequency spectra with Continuous Wavelet Transform, and thus the frequency components with time in various scales can be reflected as images. Siamese Networks with sub-nets composed of both Convolutional Neural Network (CNN) and Depth-wise Separable Convolution Network (DSCN) were proposed and established for fault diagnosis; meanwhile, fault samples were divided into sample twins to solve the problem of small samples. As to a database of rolling bearings, different kinds of faults with various degree, rotary speed and added noise were distinguished with both SN-CNN and SN-DSCN models successfully. Then an experiment for a R2R unit in printing press is also taken, there are all 7 classes of samples to be identified, and each group contains few numbers of samples. From this work, it can be seen that SN-CNN and SN-DSCN both can realize a fault diagnosis of rolling bearings in printing units based on 20 samples, which can be seen as a limited sample learning mission. Besides, SN-DSCN is proved to have a less time in training process compare to SN-CNN.