Improvement in classification accuracy and computational speed in bearing fault diagnosis using multiscale fuzzy entropy

Abstract

The operation of ball bearings under varying faulty conditions comprises complex time-varying modulations in the acquired vibration signals. In such circumstances, the extraction of nonlinear dynamic parameters based on multiscale fuzzy entropy (MFE) and refined composite multiscale fuzzy entropy (RCMFE) have proved to be more efficient in fault recognition than the conventional feature extraction methods. However, the accuracy of the methods in classifying several fault classes should not arrive at the expense of higher computational cost. The two major factors responsible for affecting the computational cost are the sampling length and number of features. This paper investigates the capabilities of MFE and RCMFE methods to estimate several health states of bearing at a different range of sampling lengths and scale factors. The bearing condition comprises normal and defective states, where the defective state considers incipient and severe faulty conditions of bearing. The diagnosis capability of both methods is verified by employing the support vector machine classifier. Although the results demonstrate higher fault classification ability of RCMFE for both incipient and severe bearing faults, the results are more impressive, especially at a higher range of scale factors as well as at lower sampling lengths.