On fault diagnosis using image-based deep learning networks based on vibration signals

With the development of complex and intelligent mechanical products, the reliability of rotating machinery equipment increasingly gains much attention. As essential parts of rotating machinery equipment, bearing failures are becoming one of the leading root causes of mechanical failures. The critical problem of the bearing fault diagnosis is the method based on mechanical vibration signals. However, an accurate fault diagnosis is constantly challenged due to the nonlinearity and non-stationarity of vibration signals. This work investigates the fault diagnosis of bearing based on vibration signals, and an image-based deep learning method is proposed, in which vibration signals in the time domain are transformed into intrinsic mode functions (IMFs), and uncorrelated IMFs are removed. Rest IMFs reconstruct signals, which are transferred into symmetrized dot pattern (SDP) images. The fault diagnosis is then formulated as an image classification problem and solved with deep learning networks. In this work, the parameters of SDP are selected by optimizing an image similarity to improve the classification accuracy. The presented method is demonstrated with an experimental fault diagnosis of rolling bearings, and results show that the accuracy of the proposed method can be up to 99%.