Aiming at the problems of large discrepancies between dynamic model response and the distribution of measured data, as well as the poor generalization in fault diagnosis methods driven by simulation data, a gearbox fault diagnosis method based on digital twin driven wavelet attention transfer network is proposed. Firstly, a virtual model of the gear transmission system was built using the lumped parameter method to map the physical system. Key parameters were optimized with measured normal data, and fault mechanism models were integrated to generate a rich twin fault dataset. Secondly, a discrete wavelet attention-based feature extraction network was designed, combining the multi-scale signal decomposition capability of discrete wavelet transform with the channel attention mechanism that dynamically focuses on strongly correlated fault features. This model effectively extracts domain-invariant fault features from both twin and measured data in the wavelet domain. Then, to address differences in marginal and conditional distributions between twin and measured data, a joint subdomain adaptation criterion was proposed by combining maximum mean discrepancy (MMD) and dynamic local MMD. This criterion measures the joint feature distribution discrepancy between the two domains, enabling the transfer of the gear twin model to real-world fault diagnosis. Finally, the proposed method was experimentally validated on a multi-stage parallel gearbox test bench. Results showed that it achieved superior diagnostic performance across all transfer tasks, with an average classification accuracy of 98.10%. The method effectively enables fault diagnosis transfer from twin data to measured data under conditions of limited labeled high-quality fault data.