基于粒子滤波和不确定图的动态系统诊断方法

在动态系统基于模型诊断中,状态空间大小与元件个数、时间是双指数关系.K-Best枚举方法每个时刻只考虑K个可能性最大的状态,有效减小了枚举空间,但当系统复杂庞大或诊断周期长时,状态更新仍是一项巨大工程.提出一种结合粒子滤波和不确定图的动态系统诊断方法PF_LUG,利用粒子在状态空间的分布近似其概率,并用不确定图标签的反向匹配代替传统的正向轨迹枚举.算法有效解决了由时间导致的计算量增长问题,使时间对复杂度的影响由指数运算降为乘积运算.仿真结果表明该算法的运行时间相对诊断周期线性增长,比K-Best枚举有明显优势.

Approach based on particle filter and uncertainty graph to diagnosis for dynamic systems

In model-based diagnosis of dynamic systems, the scale of state space is exponential to both the number of components and time-steps. K-Best enumeration considers only K states with maximum probabilities at each time-step, which reduces the enumeration space. But state updating is not feasible when the system is complex or the diagnostic duration is long. An approach named PF_LUG was presented, which is based on particle filter and labeled uncertainty graph. The probability of state was approximated by the number of particles sampling the state. And the traditional enumeration was replaced by label matching in back-tracking process. It reduced the computation cost of time-step by moving the term from exponent to multiplier in complexity function. The experimental results show that the running time increases linearly via time-step and outperforms K-Best enumeration apparently.