This work, proposes a methodology for gear fault diagnosis, based on digital twin driven supervised classification. The challenge is to build a good classification model using a dataset combining both datasets from physical system and digital twin. Numerical data are based on a 24 D.O.F single stage gearbox model, updating by varying some parameters. The approach developed in this work is tested on two types of hybrid datasets; homogeneous and heterogeneous for training MLA’s. It allows to obtain diagnosis accuracy of gear defects about 0.86 and 0.95, with respectively, support vector machines (SVM) and k-nearest neighbors (KNN) algorithms. Training phase is ensured by dataset containing fist 25% (one historical state), and 50% experimental data (Two historical states) , completed respectively by 75%(Tree novel states to be detected) and 50% of numerical data (Two novel states to be detected) of training dataset.