Early Fault Diagnosis and Classification of Ball Bearing Using Enhanced Kurtogram and Gaussian Mixture Model

Ball bearing failure is one of the major obstacles to the effective operation of large mechanical systems. During maintenance, the initial diagnosis of a fault within the bearing is key to reducing repair costs and improving the efficiency of the system. However, such faults are difficult to accurately diagnose due to noise and the unusual and unpredictable phenomena that they cause in the peaks of the measured signal. In this paper, we present an effective analytical technique for the early diagnosis of ball bearing faults based on vibration data derived from the bearings. We apply a feature extraction technique based on spectral kurtosis (SK) and then filter the results using statistical approaches. The actual defects in the bearings are evaluated in terms of a Gaussian mixture model; principal component analysis is then used to reduce the misclassifications caused by noise and weak fault symptoms. We verified the proposed algorithm experimentally and compared the results of our diagnostic technique to those obtained using the root mean square (rms) of the vibration data to evaluate the performance of the SK-based technique. The rms-based method can only be used to identify defects, whereas the SK-based method can evaluate the level and severity of the fault. In addition, the hybrid statistical SK-based process can diagnose faults without data on the rotational speed, actual load, or design specifications of the bearings or equipment that they are used in.