Early fault detection based on ball bearing vibration analysis using multinomial logistic regression

Abstract

Rotating machines play an important role in several applications such as transportation, industry and military. Maintaining continuous operations of these machines is vital to these applications. The reliability of these rotating machines depends on ball bearing health. Bearing can fail from many factors, e.g., impurity of lubricant, improper installation, etc. All of these defects can be detected by vibration analysis. Therefore, vibration signals are utilized to develop early fault detection of ball bearing to warn operators of any anomalies in the system. This paper presents a fault detection algorithm using a logistic regression method to assess the health state of the system. This regression utilizes the node energy of wavelet packet decomposition, kurtosis and crest factor as key features. The resulting signals are processed using the Savitzky-Golay smoothing algorithm before determining the time location where the defects occur with a confidence interval of 99%. The proposed method was tested with actual vibration data of bearing which were provided by the FEMTO-ST institute and showed that it is capable of accurately predicting defect time only 10 seconds after the actual defect occurred. The worst-case prediction of 970 seconds after defect occurs also observed during experiment. Although the system did not set off alarm for fault detection after 970 seconds, one can observe an increase in the fault probability trends in early state. Therefore, the proposed algorithm can be utilized for early fault detection in ball bearing applications.