A two-step passenger flow anomaly detection scheme based on machine learning methods

Abstract Subway is an important transportation means for residents due to its large volume, punctuality and environmental friendliness. However, weather factors, sports events, concerts and some unexpected events can lead to a surge or abnormality in passenger flow, which brings enormous pressure to the management of stations and passenger flow guidance. Inspired by this, this paper formulates the abnormal passenger flows into different categories in terms of the characteristics and periodical trends, and proposes a two-step abnormal detection scheme to identify the anomalies and their type, and locate abnormal positions. First, two abnormal passenger flows recognition methods based on Jensen–Shannon divergence, dynamic time warping, and density-based spatial clustering of applications with noise are established to identify the station-level abnormal passenger flow. Then, a triple standard deviation algorithm based on sliding window is further proposed to identify the abnormal type and position. Real-world smart card data of the Beijing subway in China, and the manual mutation data of the real data are employed to evaluate effectiveness of our framework. The results show that our two-step scheme is superior to the state-of-the-art algorithms, which can detect out and locate abnormal passenger flows with various characteristics. On more mutation data, this paper discusses the performances on various anomalies of different types of stations in depth, which further indicates our framework is robust and effective in practice.