Objective Due to an unreasonable coupling degree among multi-disciplinary systems, urban rail transit UTO （unattended train operation） systems face issues such as cross-pass alarms, waterfall effects, and difficult fault location. Therefore, it is necessary to study a multi-agent collaborative alarm optimization method.

Method  A collaborative alarm optimization method based on multi-agent reinforcement learning is proposed. Specialized systems, including signaling, rolling stock, and power supply, are modeled as collaborative agents, and an alarm optimization reward function is designed. The QMIX algorithm is adopted to solve the collaborative decision-making problem through a "decentralized execution with centralized training" mechanism. A mixing value network is used to learn the optimal coupling degree threshold, achieving intelligent alarm aggregation and root cause identification. The effectiveness of the method is verified via a digital twin platform.

Result & Conclusion The method reduces the alarm volume by 60%, improves the fault location accuracy by 42%, reduces the processing time by 45%, and enables the professional coupling degree to be adjusted to an ideal range of 0.30 to 0.40. Multi-disciplinary system alarms are successfully optimized from highly coupled but low-collaboration to moderately coupled and highly efficient collaboration.