Multimodel anomaly detection on spatio-temporal logistic datastream with open anomaly detection architecture


Oktay T., Yogurtcuoglu E., Sarikaya R. N. , KARACA A. C. , Komurcu M. F. , SAYAR A.

EXPERT SYSTEMS WITH APPLICATIONS, vol.186, 2021 (Journal Indexed in SCI) identifier identifier

  • Publication Type: Article / Article
  • Volume: 186
  • Publication Date: 2021
  • Doi Number: 10.1016/j.eswa.2021.115755
  • Title of Journal : EXPERT SYSTEMS WITH APPLICATIONS
  • Keywords: Logistics, Big data, Streaming data, Anomaly detection, Trajectory anomaly detection, Driver identification, Driver behavior anomaly detection

Abstract

Logistic companies rely on heavily Vehicle Tracking Systems (VTS) to provide instant and valuable data about the vehicle's condition, the cargoes' statuses, and the trips. The data that is collected from the vehicles and drivers are generally not fully utilized to its full potential other than day-to-day business activities because of the difficulties created by the big data aspects of the logistic data in storing, processing, analyzing, and reporting which require expertise and infrastructure. Real-time analysis of the data coming from VTS and mobile devices provides enormous value in terms of businesses. In this research, by using real-world data supplied by an international logistic company, we have established a multimodel streaming analytic framework that detects trajectory anomalies and identifies drivers in parallel and near-real-time. The framework can be enhanced with any number of trajectory or driver identification models, all of which can be run in parallel and contribute to the overall anomaly detection process by the "consensus"of the included models. In this study, we provide a sample implementation of the framework with one trajectory and one behavior anomaly detection model suitable for logistic spatio-temporal datastream. Provided models do not require any business-specific data and only depend on the basic VTS and location data. An isolation-based model is used for trajectory anomaly detection which does not need the source and the destination while identifying sub-trajectories. The outputs of both models are combined on a dashboard which shows the vehicles' abnormal sub-trajectories along with the notification of whether or not the vehicles are being driven by the assigned driver.