多故障程序的概率诊断方法研究

发布时间：2018-02-27 19:37

本文关键词： 软件故障诊断多故障程序概率图模型测试用例依赖性故障传播故障遮掩　出处：《大连海事大学》2016年博士论文　论文类型：学位论文

【摘要】：自动化的软件故障诊断技术对于保证软件质量起着至关重要的作用。然而现有的故障诊断技术大多假设程序只存在一个故障,这种假设在实际的程序中是不现实的。相比单故障而言,多故障程序固有的不确定性会产生更多更复杂的问题,使得现有的故障诊断方法的效果并不理想。本文通过分析程序切片、基于统计的故障定位、基于模型的软件调试以及概率图模型诊断等软件故障诊断技术的研究现状,在形式化多故障程序诊断问题模型的基础上,针对多故障程序本身固有的不确定性问题,例如测试用例依赖性、故障传播以及故障遮掩等,重点研究基于扩展概率图模型的概率诊断方法,取得了以下研究成果：(1)通过对多故障程序的测试用例依赖、故障传播和故障遮掩问题的分析,提出感染图及其概率诊断方法(IGADER)将BARINEL技术推进一步。感染图利用感染连接的概念从依赖关系角度描述语句之间的相互作用,在此基础上IGADER识别冲突、产生并鉴别候选诊断。为验证IGADER的有效性,采用不同规模的单故障和多故障程序进行实验,实验结果表明IGADER的诊断精度好于经典的Tarantula、Och iai以及BARINEL等方法。(2)基于程序语句之间的控制依赖和数据依赖关系,用马尔可夫覆盖建立基于因果模型的二层贝叶斯网络模型——概率因果图(PCEG)。通过基于Noisy-or的“自顶向下”推理以及基于标准贝叶斯的“自底向上”推理,能够有效捕捉(循环)程序在控制流和数据流上的故障传播。采用同样的程序,不同大小的测试用例集进行实验,证明PCEG相比Tarantula、Ochiai以及BARINEL方法更能刻画语句之间深层次的因果关系,对测试用例的敏感性较低,能够控制循环语句导致的相似执行信息对诊断精确性的负面影响。(3)针对软件开发中存在的“虚假依赖”问题,提出扩展隐马尔可夫模型及其概率诊断方法EHMM。EHMM把程序特征看作是“隐”变量,对每个失败测试用例建立一个隐马尔可夫模型,再通过在一组隐马尔可夫模型上的推理来对所有“隐”变量的状态进行分类,并对分类后状态为faulty的变量,计算其可疑度作为诊断结果。为了验证EHMM的有效性,特别对包含一个故障、两个故障以及三个故障的带有“虚假依赖”的真实程序进行实验设计,结果表明EHMM方法在处理带有“虚假依赖”的程序时,诊断结果要好于PCEG、IGADER、Tarantula以及Ochiai等方法。(4)诊断系统实现与应用方面,本文设计了一个集成IGADER、PCEG以及EHMM等概率诊断方法的诊断系统PGDS。该系统能够应用于实际的多故障程序诊断问题以及学生学习的认知能力诊断问题,并取得很好的效果。
[Abstract]:Automated software fault diagnosis technology plays an important role in ensuring software quality. However, most of the existing fault diagnosis techniques assume that there is only one fault in the program. This assumption is not realistic in a real program. The inherent uncertainty of a multi-fault program creates more complex problems than a single failure. Through analyzing program slice, fault location based on statistics, software debugging based on model and probability diagram model diagnosis, the present situation of software fault diagnosis technology is discussed. On the basis of formalizing the model of multi-fault program diagnosis, the inherent uncertainty problems of multi-fault program, such as test case dependency, fault propagation and fault masking, are discussed. The probabilistic diagnosis method based on extended probabilistic graph model is mainly studied. The following research results are obtained: 1) through the analysis of test case dependence, fault propagation and fault masking of multi-fault program, The infection diagram and its probabilistic diagnosis method (IGADERA) are proposed to push the BARINEL technology forward. The infection diagram uses the concept of infection connection to describe the interaction between sentences from the angle of dependency, and then IGADER identifies conflicts. To verify the effectiveness of IGADER, experiments are carried out using single-fault and multi-fault programs of different sizes. The experimental results show that the diagnostic accuracy of IGADER is better than that of classical methods such as Tarantula Och iai and BARINEL. A two-layer Bayesian network model based on causality model, probabilistic causality diagram (PCEG), is established by using Markov covering. Through "top-down" reasoning based on Noisy-or and "bottom-up" reasoning based on standard Bayes, Can effectively capture (cyclic) programs on the control flow and data flow of fault propagation. Using the same program, different sizes of test cases set for the experiment, It is proved that PCEG can depict deeper causality between statements and is less sensitive to test cases than the Tarantula Ochiai and BARINEL methods. The ability to control the negative impact of similar execution information on diagnostic accuracy caused by loop statements.) the problem of "false dependencies" in software development, An extended hidden Markov model and its probabilistic diagnosis method, EHMM.EHMM, are proposed. The program feature is regarded as a "hidden" variable, and a hidden Markov model is established for each failure test case. Then, by reasoning on a set of hidden Markov models, the states of all "hidden" variables are classified, and the suspicious degree of variables whose state is faulty after classification is calculated as the diagnostic result. In particular, a real program with "false dependency" including one fault, two faults and three faults is designed. The results show that the EHMM method is used to deal with a program with a "false dependency". The results of diagnosis are better than those of PCEG IGADERA Tarantula and Ochiai. 4) the realization and application of the diagnostic system. In this paper, a diagnosis system, PGDS, which integrates IGADERPCEG and EHMM probabilistic diagnosis methods, is designed. The system can be applied to practical multi-fault program diagnosis problems and students' learning cognitive ability diagnosis problems, and has achieved good results.
【学位授予单位】：大连海事大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP311.53

【相似文献】