Main-bearing fault identification in full-facetunnel boring machines (TBMs) impacts excavation safety and efficiency. In practice, stringent reliability requirements and proactive maintenance yield extremely imbalanced monitoring data (fault samples $ \leqslant$1%), resulting in difficult-to-learn sparse fault features and a very high missed-detection risk. To address this issue, we propose a deep reinforcement learning driven TBM main bearing fault identification model (DRLimb) that reformulates the traditional static classification problem as a sequential decision optimization problem under a reinforcement learning framework. Specifically, the process is modeled as a Markov decision process; a dual-network architecture (online and target $ Q $-networks) with soft updates is used to ensure stable convergence of the policy learning process. Moreover, we design a decision-history-aware asymmetric reward mechanism that assigns higher rewards to correct fault identifications and stronger penalties to missed detections, forcing the agent to focus on sparse yet critical fault patterns and enhancing the identification sensitivity to minority fault modes. Theoretical analysis shows that setting the majority-class reward coefficient proportional to the imbalance ratio balances inter-class gradient contributions. Experiments on multiple TBM main-bearing datasets with extreme imbalance show that DRLimb achieves $ G $-mean and F1-score above 93.2%, outperforming state-of-the-art imbalance-aware diagnostic models and baselines.