The operational state of the Tunnel Boring Machine (TBM) main bearing is critical to tunneling safety and efficiency, making fault identification essential. However, proactive maintenance strategies result in extremely imbalanced monitoring data, with fault samples often comprising less than 1%, which complicates sparse fault feature extraction and increases the risk of missed detection. To address this, a Deep Reinforcement Learning-based fault identification model (DRLimb) is proposed. This approach formulates fault identification as a Markov Decision Process and employs a dual-network architecture (online and target Q-networks) with soft updates to ensure stable training. An asymmetric reward function is designed to prioritize fault samples by assigning higher penalties and rewards, directing the agent’s focus toward sparse yet critical fault patterns. Theoretical analysis shows that setting the reward coefficient for the majority class equal to the class imbalance ratio balances gradient contributions across classes. Experiments on multiple severely imbalanced TBM main bearing datasets demonstrate that DRLimb consistently achieves G-mean and F1-Score values above 93.2%, significantly outperforming mainstream imbalanced learning diagnostic models and baseline approaches.