Bearing Fault Severity Classification Based on EMD-KLD : A Comparative Study for Incipient Ball Faults

Rolling-element bearings are the most prevalent components in rotating machinery (RM). They need to be paid a particular attention for diagnosis and maintenance to avoid mechanical system failures which in turn leads to severe equipment damage, economic losses and human life annoyance. According to the literature, the ball bearings remains a sensitive and critical component. To efficiently diagnose the related defects, it requires the development of particularly accurate and sensitive methods contrarily to Inner Race, Outer Race or Cage faults which can be evaluated more easily. From this point of view, this work will be devoted to first the detection of Ball Faults (BF) with different severities. For this purpose, a database provided by the Case Reserve Western University (CWRU) is used. The proposed methodology is based on a particularly tuned EMD-KLD solution to create the features to be used in the fault classification. An Empirical Mode Decomposition (EMD) procedure decomposes the original vibration signals into several Intrinsic Mode Functions (IMFs). Using a statistical analysis evaluation, the most relevant ones are selected and then the Kullback-Libeler Divergence (KLD) is computed for creating the features. The most appropriate learning system is finally studied as a comparative classification study using three prevalent techniques namely: Directed Acyclic Graph SVM (DAG SVM), K-Nearest Neighbor (KNN) and Decision Tree (DT). An analysis of the influence on the results for different parameters related to these methods are be presented focusing on the evolution of the Training Accuracy Rate (T-rAR), the Testing Accuracy Rate (T-sAR), the Testing and Training time (T-st and T-rt respectively).