Intelligent fault diagnosis method plays an important role in ensuring the safe and reliable operation of gearbox. However, in actual industrial scenarios, high-quality labeled fault samples are scarce, which limits the performance of fault diagnosis. To tackle this problem, this research introduces a novel mechanism-data fusion approach for the imbalanced fault diagnosis of tooth root cracks. Firstly, the dynamic twin model of single-stage parallel shaft tooth root transmission system is established to obtain high-fidelity fault simulation data, aiming to acquire the dynamic response signals of different crack faults. Subsequently, the maximum mean square difference is used as the measurement criterion of the joint distribution alignment strategy to fully integrate mean and variance information, narrow the feature distribution difference between simulation domain and measured domain. Finally, the proposed mechanism-data fusion method has been validated on two types of unbalanced sample fault diagnosis tasks, demonstrating superior fault diagnosis performance compared to alternative methods.