Datum unit optimization for robustness of a journal bearing diagnosis system

This paper describes a robust diagnosis method for a rotor system with a journal bearing. To enhance the robustness of a journal bearing diagnosis system, it is of great importance to define an optimum datum unit for featuring anomaly states of the rotor system. To support the research goal, this study makes use of three measures for class separation, including Kullback-Leibler divergence (KLD), Fisher discriminant ratio (FDR), and a newly proposed measure: probability of separation (PoS). From the viewpoint of class separability, this work found that PoS is more attractive than other methods for quantification of class separation. PoS offers favorable properties like normalization, boundedness, and high sensitivity. A generic algorithm integrated with one of three measures consistently suggested the optimum datum units among the feasible datum units. Optimum datum units were found to be one-cycle for time-domain features and sixty-cycles for frequency-domain features. The support vector machine (SVM) classifier with the optimum datum units was used for diagnosing a normal and three anomaly states. The health classification results showed that the proposed optimum datum units can outperform other datum units.