A novel, machine learning-based feature extraction method for detecting and localizing bearing component defects

Vibration analysis for conditional preventive maintenance is an essential tool for the industry. The vibration signals sensored, collected and analyzed can provide information about the state of an induction motor. Appropriate processing of these vibratory signals leads to define a normal or abnormal state of the whole rotating machinery, or in particular, one of its components. The main objective of this paper is to propose a method for automatic monitoring of bearing components condition of an induction motor. The proposed method is based on two approaches with one based on signal processing using the Hilbert spectral envelope and the other approach uses machine learning based on random forests. The Hilbert spectral envelope allows the extraction of frequency characteristics that are considered as new features entering the classifier. The frequencies chosen as features are determined from a proportional variation of their amplitudes with the variation of the load torque and the fault diameter. Furthermore, a random forest-based classifier can validate the effectiveness of extracted frequency characteristics as novel features to deal with bearing fault detection while automatically locating the faulty component with a classification rate of 99.94%. The results obtained with the proposed method have been validated experimentally using a test rig.