ML-based bevel gearbox fault diagnosis: An extensive time domain feature extraction approach with limited data

Abstract

Data-driven based gear box fault diagnosis is one of the efficient approach as the gear drive train is highly nonlinear parametrically excited system. The data driven approach is solely dependent on the quality of the data acquired from the physical system. These data are direction-sensitive and carry different information depending on the state of the gearbox. In the literature, mostly one-directional data and their traditional time-domain feature matrix are utilized for gearbox fault diagnosis. The proposed work presents for the first time the effect of multi-directional data with a unique technique for data preparation based on various concatenation approaches. Along with this, the fifty extensive time domain features are extracted and fed to three different feature selection (FS) techniques and five popular machine learning (ML)-classifiers for fault diagnosis purposes. An experimental set-up with a bevel gearbox under different fault conditions was used to acquire the vibration signal data using a tri-axial accelerometer. The paper studies the three cases of data preparation: case 1: single-directional data at a time; case 2: horizontal concatenation of multi-directional data at a time; and case 3: vertical concatenation of multi-directional data at a time for gearbox fault diagnosis. Results show that overall, case 3 performed well and gave improved accuracy under different FS techniques and ML-classifiers as compared to cases 1 and 2.