An Investigation of Ensemble Learning Algorithms for Fault Diagnosis of Roller Bearing

Abstract

Roller Bearing (RB) is one of the critical mechanical components in rotating machineries. Failure of a bearing may cause the fatal breakdown of an entire machine and inestimable financial losses due to its continuous rotation. Hence, it is significant to diagnose the fault accurately at an early stage so that it helps in predictive maintenance of the machine from malfunctioning. In the recent developments, Machine Learning (ML) has shown a drastic change in the way we predict, analyze and interpret the results. In this paper, a diagnostic technique is being proposed to identify the bearing faults that employs ensemble learning algorithms such as Bagging, Extra Tree and Gradient Boosting classifiers. The proposed method includes 1) Pre-processing of vibration data 2) Extracting statistical features such as Mean, Standard Deviation, Kurtosis, Crest Factor and Mel-Frequency Cepstral Co-efficient (MFCC) features and 3) Training the Ensemble Learning algorithms for classifying the various faults based on extracted features. For experimentation, vibration data is collected from the Case Western Reserve University (CWRU) Laboratory to diagnose 12 different fault types associated with Inner Race (IR), Outer Race (OR), Ball fault and normal bearing of varying diameters. Results shows that Ensemble learning algorithms performs better based on MFCC features as compared to statistical features.