Bearing Fault Diagnosis Method Based on Vision Measurement and Convolutional Neural Network

Abstract

Convolutional neural network (CNN) can recognize different characteristics of mechanical vibration signals and classify them intelligently. However, CNN requires plentiful vibration signal samples to train. The vibration signals collected by accelerometer have the shortcoming of less sample data and single feature. In the bearing fault diagnosis work, our main contribution is to propose a method of vision measurement combining with CNN. First, the industrial high-speed camera collects the short videos of the bearing working under different conditions. Second, we select multiple pixels on the bearing edge in the image and apply the improved phase-based algorithm to extract the micro-vibration acceleration signal. After micro-vibration acceleration data is preprocessed, the bearing datasets of different types are formed. Finally, a one-dimensional convolutional neural network (1D-CNN) model is designed to recognize and classify characteristics of bearing fault signals. To verify the reliability and validity of our proposed method, we conducted a series of experiments on the bearing fault samples acquired from a rotor platform. The results fully demonstrate that our proposed method is feasible and accurate in bearing fault diagnosis.