Generator-related faults account for 27% of total wind turbine failures, with approximately 40% of induction machine faults caused by the bearing components. Therefore, early fault warning of generator bearing is crucial for reducing wind turbine operation and maintenance costs. However, there is a lack of exploration on how to enhance the convergence, robustness, and generalization of the fault warning model. To address this problem, this paper proposed a GRPO (Group Relative Policy Optimization)-based early fault warning method for wind turbine generator bearings. Inspired by DeepSeek’s GRPO algorithm, this paper proposed a novel GRPO Loss to constrain multiple base learners during training, and the best-performance base learner is selected as the fault warning model. Validated on two wind turbines from an actual wind farm in northern China, the proposed GRPO Loss enhances model robustness and generalization. The method achieves early warning of non-drive-end bearing faults up to 93 h in advance, while reducing RMSE and MAPE by an average of 1.5158 and 3.47395% respectively.