A Comparison of Machine Learning Methods for the Diagnosis of Motor Faults Using Automated Spectral Feature Extraction Technique

Centrifugal pumps (CPs) are mainly composed of impeller and bearings. The operation of the CPs is disturbed if any of its components is faulty. Bearings faults are reported to be the major reason for pump failures. Condition monitoring of the CPs helps in early diagnosis and assists to keep machines in working condition with minimum maintenance costs. Famous non-intrusive techniques known as motor current analysis (MCA) have been reported in the literature for the detection of bearing anomalies. However, limited literature is available to diagnose the minor scratches in the bearing surface. Recent research on the diagnosis of bearing scratches identification through MCA has shown some promising results. The comparison of machine learning and convolutional neural networks (CNNs) was performed in the classification of healthy bearings and faulty bearings (holes and scratches). The fault classification accuracy of 89.26% was reported which is very low. The low amplitudes of the bearing scratch in the MCA spectrum, environment noise and utilization of conventional feature extraction techniques were the key reasons for the low accuracy. This problem has been tackled in this paper by developing an automated frequency features extraction algorithm (ASFEA) to extract useful feature from the integrated current and voltage sensors data. ASFEA operates based on the feature location identification in the spectrum, feature extraction, measuring the amplitude of the fault component and comparing it with the statistical threshold. The experimental data has been collected and the performance of the ASFEA has been tested on several machine learning techniques and the better classification accuracy of ASFEA (SVM 96.87%, k-NN 100%, NBC 96%, Gbdt 96.87%, CNN 100%) has been achieved as compared to previously reported methods.