在轨道交通客流实时监视业务中，车站实时客流告警的目的是帮助业务人员快速、准确的定位到路网中可能存在大客流风险的车站，达到提示预警的作用。如果告警阈值偏小，则告警频率会增加，使业务人员无法准确判断最需要关注的大客流风险车站，并对告警提示产生麻木心态；如果报警阈值偏大，则车站的突发大客流风险有可能未被监测到，从而无法及时采取应对措施。因此，需要提出一套科学、合理、适用性强的告警阈值。研究利用传统统计学方法和基于孤立森林的异常检测方法训练设定告警阈值，并通过应用效果测试验证对两种算法进行比选，确认利用孤立森林异常值判别思路训练的告警阈值更满足业务目标。研究成果按照“一站一方案”、“一个时段一方案”的原则，为北京市轨道交通指挥中心应急调度、诱导乘客和信息发布服务平台提供了车站实时进、出站量告警阈值，支撑客流实时监视与大客流风险预警。

In the rail transit real-time passenger flow monitoring business, a station real-time passenger flow alarm is to help the operator quickly and accurately locate the station that may have a significant passenger flow risk in the network to achieve the role of early warning. If the alarm threshold is too small, the alarm frequency will increase so that the operator cannot accurately judge the high passenger flow risk station that needs to be paid attention to most and have a numb mentality to the alarm prompt; if the alarm threshold is too large, the risk of sudden large passenger flow in the station may not be monitored, and responses cannot be taken in time. Therefore, the primary purpose of this paper is to propose a set of scientific, reasonable, and applicable alarm thresholds. This paper uses the traditional statistical method and the isolation forest method based on machine learning to train and set the alarm threshold. In addition, the two algorithms are compared through the application effect test. It is confirmed that training the outliers in the isolation forest ensures that the alarm threshold is more aligned with the business goals. According to the principle of "one station, one plan" and "one time period, one plan", the research results of this paper provide the alarm threshold of the station’s real-time inbound and outbound volume for the dispatching emergency command, passenger guidance and information service platform of the Beijing Metro Network Control Center(BMNCC), and support the real-time monitoring of passenger flow and the risk early warning of large passenger flow.