Bearing Faults Detection Using Statistical Feature Extraction and Probability Based Distance: A Comparative Study

Bearing fault detection is crucial for maintaining rotating machines. It promotes a great deal of research but still faces challenges, such as the diversity of fault types, slight fault symptoms, and non-stationary operating conditions. Recently, fault detection schemes using statistical feature extraction approaches and probability-based distance have attracted much attention. Therefore, this work investigates the effectiveness of feature extraction approaches using Local Mahalanobis Distance, Principal Component Analysis, Kernel Principal Component Analysis, and Independent Component Analysis techniques. For each, we evaluate and compare the capability of feature analysis using Kullback-Leibler divergence, Jensen-Shannon divergence, Wasserstein distance, and Kolmogorov Smirnov distance. This comparative study uses the Case Western Reserve University bearing data and shows that the fault detection solution combining Local Mahalanobis distance with Wasserstein distance can achieve outstanding performance for accurate bearing fault detection. Additionally, the fault detection approaches using the Local Mahalanobis distance features are robust to non-stationary operating conditions.