A CUSUM-like approach for online change-point detection on bus door systems

摘要

Anomaly detection on sequential data is common in many domains such as fraud detection for credit cards, intrusion detection for cyber-security or military surveillance. This study is motivated by the predictive maintenance of pneumatic doors in transit buses. In this context, each available observation is a multidimensional trajectory (or curve) representing the couple (air pressure of actuators, door position) during an opening/closing cycle of a door. The data acquisition is processed through embedded sensors. Inspired by the CUSUM test, this paper deals with an on-line change-point detection procedure on sequential data where each observation consists in a multivariate curve. The system is considered out of control when a change occurs in the curves probability distribution. A generative model, inspired by a regression model, is used to characterize a curves sequence in the curves space. The unknown parameters of these distributions are estimated using the maximum likelihood principle. Experimental studies performed on realistic data demonstrate a certain practicability of the proposed method.