Fault Diagnosis in Gearbox System with MODWT and MODWPT Using SVM and RF Algorithms

ABSTRACT For the reliability and safety of rotating machines, automatic fault diagnosis has become very important because of the multiplicity of vibration sources. In industry, most undesirable vibrations in rotating machinery are caused by gear and bearing failures. Such defects generate nonstationary signals, with the information useful for their diagnosis frequently being obscured by background noise. Therefore, a monitoring system is required to identify these faults at early stages. In this article, a new approach has been proposed to optimize the diagnostic performance of combined and uncombined gear-bearing faults, based on vibrational signal denoising using the maximal overlap discrete wavelet packet transform (MODWPT) and the maximal overlap discrete wavelet transform (MODWT). Two feature extraction methods have been used: nonlinear features, which are sensitive to the signal’s nature, and linear features that are sensitive to the signal’s energy. Feature performance has been assessed using the support vector machines (SVMs) and random forest (RF) classifiers. Applying the RF algorithm yielded 100 % accuracy for both methods. However, using SVM classifier, MODWPT proved to have higher accuracy (99.90 %) than MODWT (<97.88 %), especially with nonlinear features. Feature extraction is used for vibration analysis in time domain with accelerometer sensors in radial direction. The proposed approach is validated by experimental data, and results have consistently confirmed that features processed by the two proposed methods are better than those obtained by features that have not undergone wavelet processing.