This study develops a predictive maintenance framework for industrial gearboxes that employs CWT and XGBoost algorithms to improve early fault detection. Compared with the traditional method for detecting gearbox faults, the proposed approach can extract more useful time-frequency domain features from the vibration signals using CWT to accurately diagnose slight faults which are usually not detected by universal methods. XGBoost, a machine learning classifier, then uses these extracted features to classify the data to determine if the operation is normal or fault states, including early-stage gear cracks. To alleviate common problems such as imbalanced data, the framework incorporates Bayesian optimization and SMOTE (Synthetic Minority Oversampling Technique), attaining a considerable classification accuracy of 94.49%. This methodology has practical benefits such as minimizing downtime of the equipment, reducing maintenance costs, and improving the reliability of industrial operations, thus making it appropriate for real-world industrial applications to better meet the goals of Industry 4.0. Further research will aim to generalize fault detection to more types of gear faults and to determine how this could be integrated into industrial IoT systems to increase autonomous maintenance capabilities.