开关电路中的潜在问题进化Petri网分析方法

Evolutionary Petri nets for Sneak Circuit Analysis of Switch Circuits

利用神经网络进行潜在通路分析(SCA)由于丢失了系统的结构信息,所以导致分析结果不可靠以及解释困难等问题;为了克服这个缺陷,将学习型Petri网(LPN)用于SCA;但是传统LPN有两个缺陷:(1)都是针对无回路PN模型,这不符合实际情况;(2)大部分都是利用BP算法进行学习,带来BP算法固有的缺陷;针对这两个缺陷,提出基于克隆选择算法(CSA)的LPN(CSALPN);首先对系统进行PN建模,然后利用CSA训练PN,使得PN既可以学习先验知识又可以利用系统的结构信息;为了提高LPN的泛化能力,引入了神经网络集成;具体方法就是将训练的所有抗体作为集成中的个体,然后通过简单加权集成输出;文章还提出了带回路的学习型PN不陷入死锁的充要条件;最后用CSALPN对一个典型的电路进行SCA;统计结果证实了该方法可以有效发现开关电路的潜在通路。

Sneak circuit analysis based on neural network may generate suspect result due to losing of architectural information of sys- tem. To overcome the shortcomings, the paper propose a novel approach for SCA, which introduce learning Petri Nets （LPNs） into SCA. But there are two shortcomings of conventional LPNs： 1. all of the LPNs was deigned for non--loop PNs, this was not suitable for reality. 2. Most of the LPNs use BP algorithm to adjust parameters of PNs, this can trap into local minimum and intolerant learning speed. The pa- per introduce clonal selection algorithm （CSA） into LPNs （CSALPNs）. CSA can train LPNs with loop and non--loop. The parameters of PNs were adjusted by CSA. Neural network ensemble was introduced to improve the generalization performance, the ensemble consist of the trained antibodies as individuals to predict possible functions. The sufficient and necessary condition of loop PNs for avoiding trapping in deadlock was proposed. A typical example was used to test CSALPNs. The statistical results demonstrate CSALPNs can discover effectively sneak circuits of switch circuits.