Early Inner Race Fault Detection On A Ball Bearing Setup Using Histogram of Oriented Gradients and Wavelet Subselection

Predictive maintenance is an industrial practice to detect component failure ahead in time and before major damage is done to the system. Bearings are susceptible to such damage phenomena and should be replaced before critical failure, therefore early detection proves important. For wide applicability, these detection methods should work on easily available sensor data and have limited computational complexity. This study presents a method for classifying the health of a bearing based on the machines vibrational sensor data and with additional focus on computational complexity. The process involves converting the signals to the frequency spectrum using continuous wavelet transforms, and identifying specific frequency ranges associated with damage phenomena in bearings. Two approaches were used: analyzing the wavelet responses and creating a scalogram image to locate relevant areas. The results obtained on a bearing monitoring data set, created within the Flanders AI Research program, were consistent for both approaches and identified a specific set of scales that resulted in reduced computational load whilst attaining high failure detection rates. Consolidation is achieved by repeating the procedure on two public data sets.