Abstract: The traditional railway crew management model relying on unified assessment and static manual scheduling suffers from multiple management pain points, including mismatched staffing, delayed emergency response, crude fatigue control, and untargeted training. Based on multi-source passenger service big data, an integrated management framework of "portrait – evaluation – reinforcement learning optimization" is constructed. passenger service staff is divide into five groups by combining PCA with multi-index K-Means clustering; the CART decision tree is adopted to mine hierarchical assessment thresholds and rules affecting service quality, so as to establish differentiated evaluation criteria; A dynamic composite Q-Learning reinforcement learning model is designed for typical scenarios to formulate dynamic strategies for manpower allocation and emergency response. Field verification was carried out at three hub stations. The results show that service complaints decreased by 42.9% and emergency handling time was reduced by 64.5%, with passenger satisfaction and crew efficiency improved simultaneously. This research provides a practical technical scheme for digital and refined management of passenger transport services.