Abstract: Gearbox transmission systems typically operate at high speeds, rendering their key components susceptible to significant wear and impact damage. Given the inherent complexity of gearbox systems, simulation models often fail to accurately diagnose gearbox faults. To address this issue, this paper proposes a digital twin-based gearbox fault diagnosis method that combines physical modeling and data-driven approaches. First, by constructing two Boltzmann machines, features are extracted from the sensor data and the simulation data generated by the rigid-flexible coupled dynamic model. The information from these two data sources is mapped to a high-dimensional space to form a joint representation, which is then integrated with a multi-layer feed-forward neural network to create a multimodal information fusion model for real-time fault detection. However, during the practical application of digital twin-driven fault diagnosis methods, a discrepancy between the virtual space and the physical space often exists. Therefore, this paper develops an adaptive correction framework based on the multi-objective locust optimization algorithm to enhance the fidelity of the virtual space. Experimental results demonstrate that the proposed digital twin method reduces the information error between the physical space and the virtual space and significantly improves the accuracy of fault diagnosis of gearbox.