Fully Distributed Multi-sensor Change-point Detection

Change-point detection has been a fundamental problem in quality control, hazards monitoring and cybersecurity. With the ever-growing complexity of system and enlarging number of sensors, multi-sensor detection algorithms are in real demand. However, most existing detection algorithms consider a centralized approach, meaning that one collects all data at one location, which sometimes may be infeasible for real-time detection due to delay and cost of communication. To address this issue, we propose a fully distributed CUSUM-based detection procedure inspired by the idea of average consensus. The algorithm is fully distributed, i.e., it only requires sensors to communicate with their neighbors and no fusion center is needed. We provide theoretical analysis showing that the performance of the fully distributed scheme can match the centralized algorithms under some mild assumptions. Numerical experiments demonstrate the good performance of the algorithm.