Damage detection in spur gears using vibration kurtosis analysis: An experimental data-driven mathematical model and machine learning approach

This study investigates the application of kurtosis analysis for condition monitoring and damage detection in gears, specifically focusing on pitting and tooth breakage caused by repeated operational cycles and manufacturing defects. Predictive models were developed using machine learning (ML) and an experimental data-driven mathematical model (MM), which correlates geometric, material, and operational parameters to predict the vibration kurtosis response. The average percentage error between predicted and experimental kurtosis values is 10.44% for tooth breakage using ML and 3.59% for the MM. For pitting, the errors are 7.01% for ML and 1.05% for the MM. In both cases, the MM outperforms ML, confirming its higher accuracy and applicability in gear fault detection. The study emphasizes the potential of kurtosis in predictive maintenance, enabling early damage detection, timely interventions, and improved system performance, thereby extending gear lifespan and minimizing the risk of unexpected failures. A hybrid approach that combines the stability of MM with the adaptability of ML could provide the most effective solution for gear fault diagnostics.