基于故障关联的多故障概率诊断方法

Probabilistic Diagnosis Approach to Diagnosing Multiple-fault Programs with Fault Correlation

故障软件诊断的必要性在于真实世界中的软件几乎都会包含一个以上的故障。与单故障不同,多个故障的传播及其关联导致软件诊断更复杂,不确定性更高,概率推理因而被用于适应多故障程序的特殊性。提出了一种新的基于变形概率图FCG及其推理的软件诊断方法。相比于BARINEL方法和经典的贝叶斯网,FCG的特别之处在于采用了无向图上候选故障及其关联关系的贝叶斯推理和Noisy-or推理,而候选故障及其关联可以从程序语句间的控制依赖关系和数据依赖关系中创建。从西门子套件到更大的space,grep程序的实验,无论是在处理单故障还是处理多故障的情况下,实验结果都证明了FCG的有效性,其诊断效果比LOUPE,Ochiai,Tarantula甚至BARINEL方法都准确。

Diagnosing multiple-fault software is necessary because almost all real-world software contains more than one fault.Unlike single-fault,the propagation and correlation of multiple faults in software lead to more complexity and great uncertainty,and probabilistic reasoning is thus applied to accommodate such uniqueness.This paper proposed a new probabilistic reasoning method to diagnose multiple-fault software by using variant probabilistic graphs FCG and their inference.Distinguished from the BARINEL method and the classical Bayesian network,FCG features Bayesian and Noisy-or inference from undirected graph consisting of candidate faults and their correlation which can be set up from statement-level control and data dependencies.Experiments were conducted on programs ranging from Siemens suite to larger ones like space and grep.The experimental results validate the effectiveness of the present approach in handling programs no matter with single fault and with multiple faults,and especially it is more accurate than competitors such as LOUPE,Tarantula,Ochiai and even BARINEL.