A Comprehensive Assessment of Gearbox Tooth Faults Based on Dynamic Modelling and Machine Learning

Abstract

Abstract This paper presents a dynamic model-based gearbox fault diagnosis using machine learning. A single-stage spur gear using an eight-degrees-of-freedom (DOF) dynamic model is developed and investigated with four gear tooth conditions, i.e., healthy tooth, 20 % tooth crack, 40 % tooth crack, and 60 % tooth crack. In the developed model, an analytically improved time-varying mesh stiffness (IAM-TVMS) model, which considers the effects of structural coupling of loaded tooth, nonlinear Hertzian contact stiffness, precise transition curve, and misalignment between base and root circle, and an improved tooth crack model, are incorporated to get a reliable system dynamic response. To make the simulated response more realistic, different levels of negative white Gaussian noise (−2dB to −10dB SNR) are added to the simulated signal. The simulated noisy signals are then segmented, and a total of 12 statistical indicators are calculated on each segmented signal to develop the feature matrix. Four different machine learning algorithms are used to classify the faults from the extracted feature matrix, and their performances are compared and discussed. The results show that the KNN classifier outperformed them all, with a classification accuracy of 90.5 %. The finding shows that the proposed method works well in the presence of intense noise and may help in identifying the faults in the system in quick time without expending too much on experimental test setup.