A Machine Learning Approach for Bearing Fault Detection Using Vibration-Based Advanced Statistical Time-Domain Features

In electromechanical systems, where their failure can significantly affect system lifetime and dependability, ball bearings are essential. Early identification and evaluation of bearing problems is crucial in high-reliability industries like energy and aerospace, where safety is paramount. Vibration measurements are commonly used to identify bearing fatigue faults, such as spalls, which result from a variety of excitation mechanisms influenced by fault size and operating conditions. This study presents a robust approach to bearing fault classification using some advanced statistical time-domain features, like, Mean Absolute Value (MAV), Simple Sign Integral (SSI), Slope Sign Changes (SSC), Wilson amplitude (Wamp), Skewness (SKW), Kurtosis (KURT), Crest factor (CRF), Root Mean Square (RMS) along with other conventional features extracted from vibration signals of a widely used bearing model, FAG QJ212TVP. We selected 7 most relevant features from an original set of 15 using a correlation threshold of 0.8 and validated accuracy by testing multiple classifiers with tenfold cross validation. The results have shown that the addition of advanced features improves the accuracy and Random Forest Classifier is having the maximum accuracy of 0.99, with area under Receiver Operator Characteristics curve as 1.