Comparison of conventional method of fault determination with data-driven approach for ball bearings in a wind turbine gearbox

Abstract

The presented investigation on fault diagnosis of ball bearings compares the conventional method using FFT spectra with a data-driven approach using Support Vector Machines (SVMs). Three different cases of bearings (one healthy and two faulty bearings with different crack thickness) were used as experimental cases. The experimentally obtained time-domain acceleration signals were converted to the frequency-domain and de-noised using optimal wavelets selected based on relative magnitudes of Shannon entropy and energy values. The dominant peak was identified for each case and was subsequently compared with the characteristic bearing frequencies evaluated theoretically. The wavelet transformed time-domain experimental data was also used to train the SVM classifier. Also, the effect of statistical tools such as Principal Component Analysis (PCA) and Zero-phase Component Analysis (ZCA) on the classification accuracy of normal SVM and wavelet feature extraction-based SVM have been investigated.