To provide timely alerts for the unexpected surges of passenger flow in urban rail transit, an anomaly detection method was proposed based on the smart card data collected by the automated fare collection system. First, the length of the sliding time window was determined in light of the time-varying characteristics of passenger flow to adapt to the dynamic data environment. Then, a deep belief network model was developed to extract the features hidden in the target sample, and further recognize the pattern of the target sample within the time window. Finally, by mapping the target sample and the historical samples of the same pattern to a multi-dimensional feature space, the anomalies in metro passenger flow could thus be identified in terms of local outlier factor. According to the case study of Guangzhou Metro, results show that the pattern recognition accuracy of the proposed method was 92.5% and the anomaly detection false rate and accuracy were 3.98% and 91.9%, respectively. The detection performance of the method was related to the forms and degrees of the anomalies, and affected by the acceptance criteria of the response time. Overall, the proposed method could accurately detect a broad range of anomalies in the cost of a comparatively low false rate.