The problem of fully distributed sequential change detection (FDSCD) is studied, where a bunch of distributed sensors collaboratively detect the occurrence of an abrupt event without the help of a fusion center. Due to the distributed nature, each sensor determines whether the event has occurred according to their own observations and messages received from their neighbors through noiseless links with finite capacity. In this paper, a novel algorithm is proposed for FDSCD that views every sensor as a fusion center and mimics the behavior of the cumulative sum (CUSUM) algorithm that is optimal in the presence of fusion center. One crucial factor in the proposed algorithm is the weighting factors that reduce the influence of stale messages circulated in the network. Design of these weighting factors based on genetic algorithms is provided. Quantizer design is also studied to meet the finite capacity constraint. Two designs are proposed where the first one numerically maximizes the total output Kullback-Leibler divergence and the second one again exploits the genetic algorithms. Extensive simulations are conducted whose results show that the proposed algorithm outperforms the current state-of-art in terms of the tradeoff between expected delay and average run length to a false alarm.

A fully distributed version of sequential change detection is studied, where a bunch of distributed sensors tries to collaboratively detect the occurrence of an abrupt event as quickly as possible in the absence of a fusion center. Every link connecting two nodes is subject to a finite bandwidth constraint. A novel stopping rule is proposed where each sensor computes a modified version of the well-known cumulative sum (CUSUM) statistic, exchanges a quantized version of this modified CUSUM statistic with its neighbors via links with a finite bandwidth, and then decides its view of whether the event has occurred. Using the trade-off between the worst-case expected detection delay (EDD) and average run length (ARL) of a false alarm as the performance metric, the proposed algorithm is shown, via extensive simulations, to outperform the current state-of-the-art based on average consensus. Moreover, the proposed stopping rule can adjust the quantization levels to comply with the bandwidth constraint and only exchanges information occasionally, which is more energy- efficient and much more spectrum-efficient than the one based on average consensus requiring infinite bandwidth and constantly exchanging information.