Bearing Fault Diagnosis based on Convolutional Neural Network learning of time-domain vibration signal imaging

Mechanical fault diagnosis and analysis is of paramount importance to ensure reliable and safe operation of various industrial systems. As the massive field data becomes more available, data-driven fault diagnosis becomes feasible and prevalent. But the traditional methods have its limitations in feature extraction and most related research focus on improving the classification method for higher precision. However the feature information of the time signals is still an important part of the diagnosis which has been neglected. This paper proposed a novel method which makes use of the message in the raw time signals. Firstly, a conversion method is used to convert time signals into two-dimensional images. Then the convolutional neural network (CNN) is proposed to extract the features of the 2-D images. Finally, the problem of signal processing is transformed into the problem of image processing. Five typical faults are examined in the experiment using the Case Western Reserve University bearing dataset. The numerical result clearly confirms the effectiveness of the proposed solution.