An automated procedure for detection and identification of ball bearing damage using multivariate statistics and pattern recognition

This paper suggests an automated approach for fault detection and classification in roller bearings, which is based on pattern recognition and principal components analysis of the measured vibration signals. The signals recorded are pre-processed applying a wavelet transform in order to extract the appropriate high frequency (detailed) area needed for ball bearing fault detection. This is followed by a pattern recognition (PR) procedure used to recognise between signals coming from healthy bearings and those generated from different bearing faults. Four categories of signals are considered, namely no fault signals (from a healthy bearing) inner race fault, outer race fault and rolling element fault signals. The PR procedure uses the first six principal components extracted from the signals after a proper principal component analysis (PCA). In this work a modified PCA is suggested which is much more appropriate for categorical data. The combination of the modified PCA and the PR method ensures that the fault is automatically detected and classified to one of the considered fault categories. The method suggested does not require the knowledge/ determination of the specific fault frequencies and/or any expert analysis: once the signal filtering is done and the PC's are found the PR method automatically gives the answer if there is a fault present and its type.