Intelligent fault diagnosis of bevel gearbox using features derived from wavelet scattering transform

Abstract

Gearbox faults can significantly affect the performance and reliability of the gearbox, leading to costly downtime and maintenance. Therefore, effectiveness and timely use of fault diagnosis techniques are essential to ensure the smooth operation of machinery and prevent catastrophic failures. Wavelet scattering transform (WST) gives low variance features which has translation invariance property, whereas conventional approaches wavelet transforms (WT), and wavelet power spectrum (WPS) are more sensitive to this property. The extracted features from WST are more stable and suitable for classification and regression. Thus, WST features can be used directly to the machine learning algorithms, whereas conventional WT, WPS, etc. require significant additional signal processing. Hence, the development of a new approach for vibration-based fault diagnosis of gearbox systems using WST and support vector machine (SVM) is the main objective of this study. In which, the best features based on the scattering coefficients have been used as the input features to the SVM. For this study, vibration signals of bevel gearbox have been acquired for fault scenarios, such as healthy gear, chipped, and missing tooth. The classification accuracy at various shaft speeds such as at 10 Hz, 20 Hz and 30 Hz is obtained as 94%, 92% and 95%, respectively, using SVM model. The SVM model is also compared with Logistic Regression model and found the same classification accuracy. The suggested approach combines the benefits of WST and SVM approach without significant signal processing for preparing SVM input features compared to conventional WT and SVM approach. © 2025 Elsevier B.V., All rights reserved.