<p>Gears play a fundamental role in mechanical power transmission, making the early detection of gear-tooth cracks vital for ensuring system reliability and avoiding costly failures. This paper introduces an integrated diagnostic framework that combines digital twin (DT) technology, addressing the scientific bottleneck of limited real-fault data availability, with supervised machine learning to detect and classify gear crack severity. The approach begins with the development of a dynamic model for a single-stage gearbox, capturing the essential physics of the system. A calibration procedure is then applied to align the model’s behavior with experimental measurements, forming a high-fidelity digital twin. Once calibrated, the DT is used to simulate various fault scenarios, including different crack severity levels. These simulations generate synthetic vibration signals that serve as training data for a Multi-class Support Vector Machine (MSVM) classifier. In parallel, experimental data collected from a controlled test bench under matching conditions is used for evaluation. By aligning model-based simulation with real-world measurements, the proposed framework enables reliable fault classification without relying on extensive historical failure data. This methodology highlights the potential of digital twins to enhance predictive maintenance strategies and supports the development of scalable data-efficient diagnostic systems for rotating machinery.</p>