采用马氏决策过程和后验克拉美罗下界的多被动式移动传感器长期调度方法

Non-Myopic Scheduling Method of Multiple Passive Mobile Sensors Based on Markov Decision Process and Posterior Cramér-Rao Lower Bound

针对多被动式移动传感器协同工作时跟踪精度不稳定等问题,提出了一种基于多步预测的移动传感器长期调度方法。该方法结合部分可观马尔科夫决策过程(POMDP)构建多传感器调度模型,并基于后验克拉美罗下界(PCRLB)建立了传感器调度过程中的单步与长期代价函数;为有效减少计算复杂度,利用大量无迹采样粒子来近似估算长期代价值;通过将多约束非线性调度问题转化为决策树优化问题,可快速获取传感器长期调度方法,并给出了一种基于分支定界技术的改进决策树搜索算法。实验结果表明,所提方法能够实现移动式传感器的合理调度,在决策步长为2时,其目标跟踪精度相较于短期调度可平均提升6.08%;改进搜索算法的求解速度也更加迅速,能够有效满足在线调度的实时性要求。

A non-myopic scheduling method is proposed on the basis of multi-step prediction to solve the problem that tracking accuracies are not stable when multiple passive mobile sensors work together. A multi-sensor scheduling model is built based on the partially observable Markov decision process (POMDP), and a single-step cost function and a multi-step cost function in the scheduling process are given according to posterior Cramér-Rao lower bound (PCRLB). A large number of unscented sampling particles are used to approximate the multi-step prediction costs and to reduce the computation complexity. The sensor scheduling scheme is quickly obtained by transforming the multi-constraint nonlinear scheduling problem into a decision tree optimization problem, and solving the problem using an improved decision tree search algorithm based on the branch-and-bound technique. Simulation results and a comparison with the myopic scheduling method show that the proposed method can effectively make mobile sensors move reasonably, and the tracking accuracy improves by 6.08% on average when the decision step size is two. It is concluded that the improved search algorithm solves the problem faster and meets the real-time requirement of online scheduling.