Comparative analysis of gearbox fault detection using ensemble learning techniques with vibration sensor data

Abstract

Gearbox fault detection plays a crucial role in ensuring the reliable operation of machinery and preventing costly downtime. This research thesis aims to develop and evaluate ensemble learning techniques for accurate detection of gearbox broken tooth conditions using vibration data from SpectraQuest's Gearbox Fault Diagnostics Simulator. The dataset comprises vibration readings from sensors under both healthy and broken tooth conditions. A thorough analysis of the Gearbox Fault Diagnosis Dataset was conducted, integrating time and frequency domain analyses to inform feature engineering. A comprehensive comparative analysis of bagging, boosting, stacking, and voting approaches was conducted. The standout performer is the AdaBoostClassifierET, achieving an accuracy of 87.56%, precision of 88.36%, recall of 86.38%, and an F1 score of 87.36%. Bagging methods also exhibit commendable performance, with the BaggingClassifierET achieving an accuracy of 87.38%, precision of 87.17%, recall of 87.50%, and an F1 score of 87.34%. The research also highlights the significance of base model choices in ensemble techniques, as different base model choices yielded different results in all four techniques. The study surpasses previous work by incorporating a comprehensive set of ensemble techniques, advanced feature engineering informed by time and frequency domain analyses, and a nuanced evaluation of overfitting concerns.